Fix app tests that mutate global defaults

- Bootloader: GRUB fallback text colors → yellow/brown (amber tone)
- CLI charts: all GPU metric series use single amber color (xterm-256 #214)
- Wallpaper: logo width scaled to 400 px dynamically, shadow scales with font size
- Support bundle: renamed to YYYY-MM-DD (BEE-SP vX.X) SRV_MODEL SRV_SN ToD.tar.gz
  using dmidecode for server model (spaces→underscores) and serial number
- Remove display resolution feature (UI card, API routes, handlers, tests)

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

audit/Makefile

@@ -1,7 +1,10 @@
 LISTEN ?= :8080
 AUDIT_PATH ?=
+EXPORT_DIR ?= $(CURDIR)/.tmp/export
+VERSION ?= $(shell sh ./scripts/resolve-version.sh)
+GO_LDFLAGS := -X main.Version=$(VERSION)
 
-RUN_ARGS := web --listen $(LISTEN)
+RUN_ARGS := web --listen $(LISTEN) --export-dir $(EXPORT_DIR)
 ifneq ($(AUDIT_PATH),)
 RUN_ARGS += --audit-path $(AUDIT_PATH)
 endif
@@ -9,10 +12,11 @@ endif
 .PHONY: run build test
 
 run:
-	go run ./cmd/bee $(RUN_ARGS)
+	mkdir -p $(EXPORT_DIR)
+	go run -ldflags "$(GO_LDFLAGS)" ./cmd/bee $(RUN_ARGS)
 
 build:
-	go build -o bee ./cmd/bee
+	go build -ldflags "$(GO_LDFLAGS)" -o bee ./cmd/bee
 
 test:
 	go test ./...

audit/cmd/bee/main.go

@@ -8,6 +8,7 @@ import (
 	"log/slog"
 	"os"
 	"runtime/debug"
+	"strconv"
 	"strings"
 
 	"bee/audit/internal/app"
@@ -21,30 +22,7 @@ var Version = "dev"
 func buildLabel() string {
 	label := strings.TrimSpace(Version)
 	if label == "" {
-		label = "dev"
-	}
-	if info, ok := debug.ReadBuildInfo(); ok {
-		var revision string
-		var modified bool
-		for _, setting := range info.Settings {
-			switch setting.Key {
-			case "vcs.revision":
-				revision = setting.Value
-			case "vcs.modified":
-				modified = setting.Value == "true"
-			}
-		}
-		if revision != "" {
-			short := revision
-			if len(short) > 12 {
-				short = short[:12]
-			}
-			label += " (" + short
-			if modified {
-				label += "+"
-			}
-			label += ")"
-		}
+		return "dev"
 	}
 	return label
 }
@@ -53,10 +31,19 @@ func main() {
 	os.Exit(run(os.Args[1:], os.Stdout, os.Stderr))
 }
 
-func run(args []string, stdout, stderr io.Writer) int {
+func run(args []string, stdout, stderr io.Writer) (exitCode int) {
 	slog.SetDefault(slog.New(slog.NewTextHandler(os.Stderr, &slog.HandlerOptions{
 		Level: slog.LevelInfo,
 	})))
+	defer func() {
+		if rec := recover(); rec != nil {
+			slog.Error("fatal panic",
+				"panic", fmt.Sprint(rec),
+				"stack", string(debug.Stack()),
+			)
+			exitCode = 1
+		}
+	}()
 
 	if len(args) == 0 {
 		printRootUsage(stderr)
@@ -82,6 +69,8 @@ func run(args []string, stdout, stderr io.Writer) int {
 		return runWeb(args[1:], stdout, stderr)
 	case "sat":
 		return runSAT(args[1:], stdout, stderr)
+	case "benchmark":
+		return runBenchmark(args[1:], stdout, stderr)
 	case "version", "--version", "-version":
 		fmt.Fprintln(stdout, Version)
 		return 0
@@ -98,8 +87,9 @@ func printRootUsage(w io.Writer) {
   bee preflight --output stdout|file:<path>
   bee export  --target <device>
   bee support-bundle --output stdout|file:<path>
-  bee web     --listen :80 --audit-path `+app.DefaultAuditJSONPath+`
+  bee web     --listen :80 [--audit-path `+app.DefaultAuditJSONPath+`]
   bee sat nvidia|memory|storage|cpu [--duration <seconds>]
+  bee benchmark nvidia [--profile standard|stability|overnight]
   bee version
   bee help [command]`)
 }
@@ -118,6 +108,8 @@ func runHelp(args []string, stdout, stderr io.Writer) int {
 		return runWeb([]string{"--help"}, stdout, stdout)
 	case "sat":
 		return runSAT([]string{"--help"}, stdout, stderr)
+	case "benchmark":
+		return runBenchmark([]string{"--help"}, stdout, stderr)
 	case "version":
 		fmt.Fprintln(stdout, "usage: bee version")
 		return 0
@@ -304,7 +296,7 @@ func runWeb(args []string, stdout, stderr io.Writer) int {
 	fs := flag.NewFlagSet("web", flag.ContinueOnError)
 	fs.SetOutput(stderr)
 	listenAddr := fs.String("listen", ":8080", "listen address, e.g. :80")
-	auditPath := fs.String("audit-path", app.DefaultAuditJSONPath, "path to the latest audit JSON snapshot")
+	auditPath := fs.String("audit-path", "", "optional path to the latest audit JSON snapshot")
 	exportDir := fs.String("export-dir", app.DefaultExportDir, "directory with logs, SAT results, and support bundles")
 	title := fs.String("title", "Bee Hardware Audit", "page title")
 	fs.Usage = func() {
@@ -390,9 +382,9 @@ func runSAT(args []string, stdout, stderr io.Writer) int {
 			archive, err = application.RunNvidiaAcceptancePack("", logLine)
 		}
 	case "memory":
-		archive, err = application.RunMemoryAcceptancePackCtx(context.Background(), "", logLine)
+		archive, err = application.RunMemoryAcceptancePackCtx(context.Background(), "", 256, 1, logLine)
 	case "storage":
-		archive, err = application.RunStorageAcceptancePackCtx(context.Background(), "", logLine)
+		archive, err = application.RunStorageAcceptancePackCtx(context.Background(), "", false, logLine)
 	case "cpu":
 		dur := *duration
 		if dur <= 0 {
@@ -407,3 +399,85 @@ func runSAT(args []string, stdout, stderr io.Writer) int {
 	slog.Info("sat archive written", "target", target, "path", archive)
 	return 0
 }
+
+func runBenchmark(args []string, stdout, stderr io.Writer) int {
+	if len(args) == 0 {
+		fmt.Fprintln(stderr, "usage: bee benchmark nvidia [--profile standard|stability|overnight] [--devices 0,1] [--exclude 2,3] [--size-mb N] [--skip-nccl]")
+		return 2
+	}
+	if args[0] == "help" || args[0] == "--help" || args[0] == "-h" {
+		fmt.Fprintln(stdout, "usage: bee benchmark nvidia [--profile standard|stability|overnight] [--devices 0,1] [--exclude 2,3] [--size-mb N] [--skip-nccl]")
+		return 0
+	}
+	target := args[0]
+	if target != "nvidia" {
+		fmt.Fprintf(stderr, "bee benchmark: unknown target %q\n", target)
+		fmt.Fprintln(stderr, "usage: bee benchmark nvidia [--profile standard|stability|overnight] [--devices 0,1] [--exclude 2,3] [--size-mb N] [--skip-nccl]")
+		return 2
+	}
+
+	fs := flag.NewFlagSet("benchmark", flag.ContinueOnError)
+	fs.SetOutput(stderr)
+	profile := fs.String("profile", platform.NvidiaBenchmarkProfileStandard, "benchmark profile: standard, stability, overnight")
+	devices := fs.String("devices", "", "comma-separated GPU indices to include")
+	exclude := fs.String("exclude", "", "comma-separated GPU indices to exclude")
+	sizeMB := fs.Int("size-mb", 0, "per-GPU benchmark buffer size in MB (0 = auto)")
+	skipNCCL := fs.Bool("skip-nccl", false, "skip multi-GPU NCCL interconnect benchmark")
+	if err := fs.Parse(args[1:]); err != nil {
+		if err == flag.ErrHelp {
+			return 0
+		}
+		return 2
+	}
+	if fs.NArg() != 0 {
+		fmt.Fprintf(stderr, "bee benchmark: unexpected arguments\n")
+		return 2
+	}
+
+	includeIndices, err := parseBenchmarkIndexCSV(*devices)
+	if err != nil {
+		fmt.Fprintf(stderr, "bee benchmark: invalid --devices: %v\n", err)
+		return 2
+	}
+	excludeIndices, err := parseBenchmarkIndexCSV(*exclude)
+	if err != nil {
+		fmt.Fprintf(stderr, "bee benchmark: invalid --exclude: %v\n", err)
+		return 2
+	}
+
+	application := app.New(platform.New())
+	logLine := func(s string) { fmt.Fprintln(os.Stderr, s) }
+	archive, err := application.RunNvidiaBenchmark("", platform.NvidiaBenchmarkOptions{
+		Profile:           *profile,
+		SizeMB:            *sizeMB,
+		GPUIndices:        includeIndices,
+		ExcludeGPUIndices: excludeIndices,
+		RunNCCL:           !*skipNCCL,
+	}, logLine)
+	if err != nil {
+		slog.Error("run benchmark", "target", target, "err", err)
+		return 1
+	}
+	slog.Info("benchmark archive written", "target", target, "path", archive)
+	return 0
+}
+
+func parseBenchmarkIndexCSV(raw string) ([]int, error) {
+	raw = strings.TrimSpace(raw)
+	if raw == "" {
+		return nil, nil
+	}
+	var indices []int
+	for _, part := range strings.Split(raw, ",") {
+		part = strings.TrimSpace(part)
+		if part == "" {
+			continue
+		}
+		value, err := strconv.Atoi(part)
+		if err != nil || value < 0 {
+			return nil, fmt.Errorf("bad gpu index %q", part)
+		}
+		indices = append(indices, value)
+	}
+	return indices, nil
+}

audit/cmd/bee/main_test.go

@@ -46,8 +46,6 @@ func TestRunUnknownCommand(t *testing.T) {
 }
 
 func TestRunVersion(t *testing.T) {
-	t.Parallel()
-
 	old := Version
 	Version = "test-version"
 	t.Cleanup(func() { Version = old })
@@ -62,6 +60,16 @@ func TestRunVersion(t *testing.T) {
 	}
 }
 
+func TestBuildLabelUsesVersionAsIs(t *testing.T) {
+	old := Version
+	Version = "1.2.3"
+	t.Cleanup(func() { Version = old })
+
+	if got := buildLabel(); got != "1.2.3" {
+		t.Fatalf("buildLabel=%q want %q", got, "1.2.3")
+	}
+}
+
 func TestRunExportRequiresTarget(t *testing.T) {
 	t.Parallel()
 

audit/internal/app/app.go

@@ -19,17 +19,18 @@ import (
 )
 
 var (
-	DefaultExportDir       = "/appdata/bee/export"
-	DefaultAuditJSONPath   = DefaultExportDir + "/bee-audit.json"
-	DefaultAuditLogPath    = DefaultExportDir + "/bee-audit.log"
-	DefaultWebLogPath      = DefaultExportDir + "/bee-web.log"
-	DefaultNetworkLogPath  = DefaultExportDir + "/bee-network.log"
-	DefaultNvidiaLogPath   = DefaultExportDir + "/bee-nvidia.log"
-	DefaultSSHLogPath      = DefaultExportDir + "/bee-sshsetup.log"
-	DefaultRuntimeJSONPath = DefaultExportDir + "/runtime-health.json"
-	DefaultRuntimeLogPath  = DefaultExportDir + "/runtime-health.log"
-	DefaultTechDumpDir     = DefaultExportDir + "/techdump"
-	DefaultSATBaseDir      = DefaultExportDir + "/bee-sat"
+	DefaultExportDir        = "/appdata/bee/export"
+	DefaultAuditJSONPath    = DefaultExportDir + "/bee-audit.json"
+	DefaultAuditLogPath     = DefaultExportDir + "/bee-audit.log"
+	DefaultWebLogPath       = DefaultExportDir + "/bee-web.log"
+	DefaultNetworkLogPath   = DefaultExportDir + "/bee-network.log"
+	DefaultNvidiaLogPath    = DefaultExportDir + "/bee-nvidia.log"
+	DefaultSSHLogPath       = DefaultExportDir + "/bee-sshsetup.log"
+	DefaultRuntimeJSONPath  = DefaultExportDir + "/runtime-health.json"
+	DefaultRuntimeLogPath   = DefaultExportDir + "/runtime-health.log"
+	DefaultTechDumpDir      = DefaultExportDir + "/techdump"
+	DefaultSATBaseDir       = DefaultExportDir + "/bee-sat"
+	DefaultBenchmarkBaseDir = DefaultExportDir + "/bee-benchmark"
 )
 
 type App struct {
@@ -40,6 +41,8 @@ type App struct {
 	sat       satRunner
 	runtime   runtimeChecker
 	installer installer
+	// StatusDB is the unified component health store (nil if unavailable).
+	StatusDB *ComponentStatusDB
 }
 
 type ActionResult struct {
@@ -112,9 +115,17 @@ func (a *App) RunInstallToRAM(ctx context.Context, logFunc func(string)) error {
 type satRunner interface {
 	RunNvidiaAcceptancePack(baseDir string, logFunc func(string)) (string, error)
 	RunNvidiaAcceptancePackWithOptions(ctx context.Context, baseDir string, diagLevel int, gpuIndices []int, logFunc func(string)) (string, error)
+	RunNvidiaTargetedStressValidatePack(ctx context.Context, baseDir string, durationSec int, gpuIndices []int, logFunc func(string)) (string, error)
+	RunNvidiaBenchmark(ctx context.Context, baseDir string, opts platform.NvidiaBenchmarkOptions, logFunc func(string)) (string, error)
+	RunNvidiaOfficialComputePack(ctx context.Context, baseDir string, durationSec int, gpuIndices []int, staggerSec int, logFunc func(string)) (string, error)
+	RunNvidiaTargetedPowerPack(ctx context.Context, baseDir string, durationSec int, gpuIndices []int, logFunc func(string)) (string, error)
+	RunNvidiaPulseTestPack(ctx context.Context, baseDir string, durationSec int, gpuIndices []int, logFunc func(string)) (string, error)
+	RunNvidiaBandwidthPack(ctx context.Context, baseDir string, gpuIndices []int, logFunc func(string)) (string, error)
 	RunNvidiaStressPack(ctx context.Context, baseDir string, opts platform.NvidiaStressOptions, logFunc func(string)) (string, error)
-	RunMemoryAcceptancePack(ctx context.Context, baseDir string, logFunc func(string)) (string, error)
-	RunStorageAcceptancePack(ctx context.Context, baseDir string, logFunc func(string)) (string, error)
+	ListNvidiaGPUStatuses() ([]platform.NvidiaGPUStatus, error)
+	ResetNvidiaGPU(index int) (string, error)
+	RunMemoryAcceptancePack(ctx context.Context, baseDir string, sizeMB, passes int, logFunc func(string)) (string, error)
+	RunStorageAcceptancePack(ctx context.Context, baseDir string, extended bool, logFunc func(string)) (string, error)
 	RunCPUAcceptancePack(ctx context.Context, baseDir string, durationSec int, logFunc func(string)) (string, error)
 	ListNvidiaGPUs() ([]platform.NvidiaGPU, error)
 	DetectGPUVendor() string
@@ -136,7 +147,7 @@ type runtimeChecker interface {
 }
 
 func New(platform *platform.System) *App {
-	return &App{
+	a := &App{
 		network:   platform,
 		services:  platform,
 		exports:   platform,
@@ -145,6 +156,10 @@ func New(platform *platform.System) *App {
 		runtime:   platform,
 		installer: platform,
 	}
+	if db, err := OpenComponentStatusDB(DefaultExportDir + "/component-status.json"); err == nil {
+		a.StatusDB = db
+	}
+	return a
 }
 
 // ApplySATOverlay parses a raw audit JSON, overlays the latest SAT results,
@@ -154,7 +169,7 @@ func ApplySATOverlay(auditJSON []byte) ([]byte, error) {
 	if err != nil {
 		return nil, err
 	}
-	applyLatestSATStatuses(&snap.Hardware, DefaultSATBaseDir)
+	applyLatestSATStatuses(&snap.Hardware, DefaultSATBaseDir, nil)
 	return json.MarshalIndent(snap, "", "  ")
 }
 
@@ -174,7 +189,7 @@ func (a *App) RunAudit(runtimeMode runtimeenv.Mode, output string) (string, erro
 		}
 	}
 	result := collector.Run(runtimeMode)
-	applyLatestSATStatuses(&result.Hardware, DefaultSATBaseDir)
+	applyLatestSATStatuses(&result.Hardware, DefaultSATBaseDir, a.StatusDB)
 	if health, err := ReadRuntimeHealth(DefaultRuntimeJSONPath); err == nil {
 		result.Runtime = &health
 	}
@@ -189,10 +204,7 @@ func (a *App) RunAudit(runtimeMode runtimeenv.Mode, output string) (string, erro
 		return "stdout", err
 	case strings.HasPrefix(output, "file:"):
 		path := strings.TrimPrefix(output, "file:")
-		if err := os.MkdirAll(filepath.Dir(path), 0755); err != nil {
-			return "", err
-		}
-		if err := os.WriteFile(path, append(data, '\n'), 0644); err != nil {
+		if err := atomicWriteFile(path, append(data, '\n'), 0644); err != nil {
 			return "", err
 		}
 		return path, nil
@@ -217,10 +229,7 @@ func (a *App) RunRuntimePreflight(output string) (string, error) {
 		return "stdout", err
 	case strings.HasPrefix(output, "file:"):
 		path := strings.TrimPrefix(output, "file:")
-		if err := os.MkdirAll(filepath.Dir(path), 0755); err != nil {
-			return "", err
-		}
-		if err := os.WriteFile(path, append(data, '\n'), 0644); err != nil {
+		if err := atomicWriteFile(path, append(data, '\n'), 0644); err != nil {
 			return "", err
 		}
 		return path, nil
@@ -514,6 +523,15 @@ func (a *App) ListNvidiaGPUs() ([]platform.NvidiaGPU, error) {
 	return a.sat.ListNvidiaGPUs()
 }
 
+func (a *App) ListNvidiaGPUStatuses() ([]platform.NvidiaGPUStatus, error) {
+	return a.sat.ListNvidiaGPUStatuses()
+}
+
+func (a *App) ResetNvidiaGPU(index int) (ActionResult, error) {
+	out, err := a.sat.ResetNvidiaGPU(index)
+	return ActionResult{Title: fmt.Sprintf("Reset NVIDIA GPU %d", index), Body: strings.TrimSpace(out)}, err
+}
+
 func (a *App) RunNvidiaAcceptancePackWithOptions(ctx context.Context, baseDir string, diagLevel int, gpuIndices []int, logFunc func(string)) (ActionResult, error) {
 	if strings.TrimSpace(baseDir) == "" {
 		baseDir = DefaultSATBaseDir
@@ -526,10 +544,56 @@ func (a *App) RunNvidiaAcceptancePackWithOptions(ctx context.Context, baseDir st
 	return ActionResult{Title: "NVIDIA DCGM", Body: body}, err
 }
 
+func (a *App) RunNvidiaTargetedStressValidatePack(ctx context.Context, baseDir string, durationSec int, gpuIndices []int, logFunc func(string)) (string, error) {
+	if strings.TrimSpace(baseDir) == "" {
+		baseDir = DefaultSATBaseDir
+	}
+	return a.sat.RunNvidiaTargetedStressValidatePack(ctx, baseDir, durationSec, gpuIndices, logFunc)
+}
+
 func (a *App) RunNvidiaStressPack(baseDir string, opts platform.NvidiaStressOptions, logFunc func(string)) (string, error) {
 	return a.RunNvidiaStressPackCtx(context.Background(), baseDir, opts, logFunc)
 }
 
+func (a *App) RunNvidiaBenchmark(baseDir string, opts platform.NvidiaBenchmarkOptions, logFunc func(string)) (string, error) {
+	return a.RunNvidiaBenchmarkCtx(context.Background(), baseDir, opts, logFunc)
+}
+
+func (a *App) RunNvidiaBenchmarkCtx(ctx context.Context, baseDir string, opts platform.NvidiaBenchmarkOptions, logFunc func(string)) (string, error) {
+	if strings.TrimSpace(baseDir) == "" {
+		baseDir = DefaultBenchmarkBaseDir
+	}
+	return a.sat.RunNvidiaBenchmark(ctx, baseDir, opts, logFunc)
+}
+
+func (a *App) RunNvidiaOfficialComputePack(ctx context.Context, baseDir string, durationSec int, gpuIndices []int, staggerSec int, logFunc func(string)) (string, error) {
+	if strings.TrimSpace(baseDir) == "" {
+		baseDir = DefaultSATBaseDir
+	}
+	return a.sat.RunNvidiaOfficialComputePack(ctx, baseDir, durationSec, gpuIndices, staggerSec, logFunc)
+}
+
+func (a *App) RunNvidiaTargetedPowerPack(ctx context.Context, baseDir string, durationSec int, gpuIndices []int, logFunc func(string)) (string, error) {
+	if strings.TrimSpace(baseDir) == "" {
+		baseDir = DefaultSATBaseDir
+	}
+	return a.sat.RunNvidiaTargetedPowerPack(ctx, baseDir, durationSec, gpuIndices, logFunc)
+}
+
+func (a *App) RunNvidiaPulseTestPack(ctx context.Context, baseDir string, durationSec int, gpuIndices []int, logFunc func(string)) (string, error) {
+	if strings.TrimSpace(baseDir) == "" {
+		baseDir = DefaultSATBaseDir
+	}
+	return a.sat.RunNvidiaPulseTestPack(ctx, baseDir, durationSec, gpuIndices, logFunc)
+}
+
+func (a *App) RunNvidiaBandwidthPack(ctx context.Context, baseDir string, gpuIndices []int, logFunc func(string)) (string, error) {
+	if strings.TrimSpace(baseDir) == "" {
+		baseDir = DefaultSATBaseDir
+	}
+	return a.sat.RunNvidiaBandwidthPack(ctx, baseDir, gpuIndices, logFunc)
+}
+
 func (a *App) RunNvidiaStressPackCtx(ctx context.Context, baseDir string, opts platform.NvidiaStressOptions, logFunc func(string)) (string, error) {
 	if strings.TrimSpace(baseDir) == "" {
 		baseDir = DefaultSATBaseDir
@@ -538,14 +602,14 @@ func (a *App) RunNvidiaStressPackCtx(ctx context.Context, baseDir string, opts p
 }
 
 func (a *App) RunMemoryAcceptancePack(baseDir string, logFunc func(string)) (string, error) {
-	return a.RunMemoryAcceptancePackCtx(context.Background(), baseDir, logFunc)
+	return a.RunMemoryAcceptancePackCtx(context.Background(), baseDir, 256, 1, logFunc)
 }
 
-func (a *App) RunMemoryAcceptancePackCtx(ctx context.Context, baseDir string, logFunc func(string)) (string, error) {
+func (a *App) RunMemoryAcceptancePackCtx(ctx context.Context, baseDir string, sizeMB, passes int, logFunc func(string)) (string, error) {
 	if strings.TrimSpace(baseDir) == "" {
 		baseDir = DefaultSATBaseDir
 	}
-	return a.sat.RunMemoryAcceptancePack(ctx, baseDir, logFunc)
+	return a.sat.RunMemoryAcceptancePack(ctx, baseDir, sizeMB, passes, logFunc)
 }
 
 func (a *App) RunMemoryAcceptancePackResult(baseDir string) (ActionResult, error) {
@@ -570,14 +634,14 @@ func (a *App) RunCPUAcceptancePackResult(baseDir string, durationSec int) (Actio
 }
 
 func (a *App) RunStorageAcceptancePack(baseDir string, logFunc func(string)) (string, error) {
-	return a.RunStorageAcceptancePackCtx(context.Background(), baseDir, logFunc)
+	return a.RunStorageAcceptancePackCtx(context.Background(), baseDir, false, logFunc)
 }
 
-func (a *App) RunStorageAcceptancePackCtx(ctx context.Context, baseDir string, logFunc func(string)) (string, error) {
+func (a *App) RunStorageAcceptancePackCtx(ctx context.Context, baseDir string, extended bool, logFunc func(string)) (string, error) {
 	if strings.TrimSpace(baseDir) == "" {
 		baseDir = DefaultSATBaseDir
 	}
-	return a.sat.RunStorageAcceptancePack(ctx, baseDir, logFunc)
+	return a.sat.RunStorageAcceptancePack(ctx, baseDir, extended, logFunc)
 }
 
 func (a *App) RunStorageAcceptancePackResult(baseDir string) (ActionResult, error) {
@@ -880,6 +944,12 @@ func latestSATSummaries() []string {
 		prefix string
 	}{
 		{label: "NVIDIA SAT", prefix: "gpu-nvidia-"},
+		{label: "NVIDIA Targeted Stress Validate (dcgmi diag targeted_stress)", prefix: "gpu-nvidia-targeted-stress-"},
+		{label: "NVIDIA Max Compute Load (dcgmproftester)", prefix: "gpu-nvidia-compute-"},
+		{label: "NVIDIA Targeted Power (dcgmi diag targeted_power)", prefix: "gpu-nvidia-targeted-power-"},
+		{label: "NVIDIA Pulse Test (dcgmi diag pulse_test)", prefix: "gpu-nvidia-pulse-"},
+		{label: "NVIDIA Interconnect Test (NCCL all_reduce_perf)", prefix: "gpu-nvidia-nccl-"},
+		{label: "NVIDIA Bandwidth Test (NVBandwidth)", prefix: "gpu-nvidia-bandwidth-"},
 		{label: "Memory SAT", prefix: "memory-"},
 		{label: "Storage SAT", prefix: "storage-"},
 		{label: "CPU SAT", prefix: "cpu-"},

audit/internal/app/app_test.go

@@ -120,15 +120,23 @@ func (f fakeTools) CheckTools(names []string) []platform.ToolStatus {
 }
 
 type fakeSAT struct {
-	runNvidiaFn       func(string) (string, error)
-	runNvidiaStressFn func(string, platform.NvidiaStressOptions) (string, error)
-	runMemoryFn       func(string) (string, error)
-	runStorageFn      func(string) (string, error)
-	runCPUFn          func(string, int) (string, error)
-	detectVendorFn    func() string
-	listAMDGPUsFn     func() ([]platform.AMDGPUInfo, error)
-	runAMDPackFn      func(string) (string, error)
-	listNvidiaGPUsFn  func() ([]platform.NvidiaGPU, error)
+	runNvidiaFn               func(string) (string, error)
+	runNvidiaBenchmarkFn      func(string, platform.NvidiaBenchmarkOptions) (string, error)
+	runNvidiaStressFn         func(string, platform.NvidiaStressOptions) (string, error)
+	runNvidiaComputeFn        func(string, int, []int) (string, error)
+	runNvidiaPowerFn          func(string, int, []int) (string, error)
+	runNvidiaPulseFn          func(string, int, []int) (string, error)
+	runNvidiaBandwidthFn      func(string, []int) (string, error)
+	runNvidiaTargetedStressFn func(string, int, []int) (string, error)
+	runMemoryFn               func(string) (string, error)
+	runStorageFn              func(string) (string, error)
+	runCPUFn                  func(string, int) (string, error)
+	detectVendorFn            func() string
+	listAMDGPUsFn             func() ([]platform.AMDGPUInfo, error)
+	runAMDPackFn              func(string) (string, error)
+	listNvidiaGPUsFn          func() ([]platform.NvidiaGPU, error)
+	listNvidiaGPUStatusesFn   func() ([]platform.NvidiaGPUStatus, error)
+	resetNvidiaGPUFn          func(int) (string, error)
 }
 
 func (f fakeSAT) RunNvidiaAcceptancePack(baseDir string, _ func(string)) (string, error) {
@@ -139,6 +147,48 @@ func (f fakeSAT) RunNvidiaAcceptancePackWithOptions(_ context.Context, baseDir s
 	return f.runNvidiaFn(baseDir)
 }
 
+func (f fakeSAT) RunNvidiaBenchmark(_ context.Context, baseDir string, opts platform.NvidiaBenchmarkOptions, _ func(string)) (string, error) {
+	if f.runNvidiaBenchmarkFn != nil {
+		return f.runNvidiaBenchmarkFn(baseDir, opts)
+	}
+	return f.runNvidiaFn(baseDir)
+}
+
+func (f fakeSAT) RunNvidiaTargetedStressValidatePack(_ context.Context, baseDir string, durationSec int, gpuIndices []int, _ func(string)) (string, error) {
+	if f.runNvidiaTargetedStressFn != nil {
+		return f.runNvidiaTargetedStressFn(baseDir, durationSec, gpuIndices)
+	}
+	return f.runNvidiaFn(baseDir)
+}
+
+func (f fakeSAT) RunNvidiaOfficialComputePack(_ context.Context, baseDir string, durationSec int, gpuIndices []int, _ int, _ func(string)) (string, error) {
+	if f.runNvidiaComputeFn != nil {
+		return f.runNvidiaComputeFn(baseDir, durationSec, gpuIndices)
+	}
+	return f.runNvidiaFn(baseDir)
+}
+
+func (f fakeSAT) RunNvidiaTargetedPowerPack(_ context.Context, baseDir string, durationSec int, gpuIndices []int, _ func(string)) (string, error) {
+	if f.runNvidiaPowerFn != nil {
+		return f.runNvidiaPowerFn(baseDir, durationSec, gpuIndices)
+	}
+	return f.runNvidiaFn(baseDir)
+}
+
+func (f fakeSAT) RunNvidiaPulseTestPack(_ context.Context, baseDir string, durationSec int, gpuIndices []int, _ func(string)) (string, error) {
+	if f.runNvidiaPulseFn != nil {
+		return f.runNvidiaPulseFn(baseDir, durationSec, gpuIndices)
+	}
+	return f.runNvidiaFn(baseDir)
+}
+
+func (f fakeSAT) RunNvidiaBandwidthPack(_ context.Context, baseDir string, gpuIndices []int, _ func(string)) (string, error) {
+	if f.runNvidiaBandwidthFn != nil {
+		return f.runNvidiaBandwidthFn(baseDir, gpuIndices)
+	}
+	return f.runNvidiaFn(baseDir)
+}
+
 func (f fakeSAT) RunNvidiaStressPack(_ context.Context, baseDir string, opts platform.NvidiaStressOptions, _ func(string)) (string, error) {
 	if f.runNvidiaStressFn != nil {
 		return f.runNvidiaStressFn(baseDir, opts)
@@ -153,11 +203,25 @@ func (f fakeSAT) ListNvidiaGPUs() ([]platform.NvidiaGPU, error) {
 	return nil, nil
 }
 
-func (f fakeSAT) RunMemoryAcceptancePack(_ context.Context, baseDir string, _ func(string)) (string, error) {
+func (f fakeSAT) ListNvidiaGPUStatuses() ([]platform.NvidiaGPUStatus, error) {
+	if f.listNvidiaGPUStatusesFn != nil {
+		return f.listNvidiaGPUStatusesFn()
+	}
+	return nil, nil
+}
+
+func (f fakeSAT) ResetNvidiaGPU(index int) (string, error) {
+	if f.resetNvidiaGPUFn != nil {
+		return f.resetNvidiaGPUFn(index)
+	}
+	return "", nil
+}
+
+func (f fakeSAT) RunMemoryAcceptancePack(_ context.Context, baseDir string, _, _ int, _ func(string)) (string, error) {
 	return f.runMemoryFn(baseDir)
 }
 
-func (f fakeSAT) RunStorageAcceptancePack(_ context.Context, baseDir string, _ func(string)) (string, error) {
+func (f fakeSAT) RunStorageAcceptancePack(_ context.Context, baseDir string, _ bool, _ func(string)) (string, error) {
 	return f.runStorageFn(baseDir)
 }
 
@@ -478,8 +542,6 @@ func TestActionResultsUseFallbackBody(t *testing.T) {
 }
 
 func TestExportSupportBundleResultMentionsUnmountedUSB(t *testing.T) {
-	t.Parallel()
-
 	tmp := t.TempDir()
 	oldExportDir := DefaultExportDir
 	DefaultExportDir = tmp
@@ -516,8 +578,6 @@ func TestExportSupportBundleResultMentionsUnmountedUSB(t *testing.T) {
 }
 
 func TestExportSupportBundleResultDoesNotPretendSuccessOnError(t *testing.T) {
-	t.Parallel()
-
 	tmp := t.TempDir()
 	oldExportDir := DefaultExportDir
 	DefaultExportDir = tmp
@@ -579,8 +639,6 @@ func TestRunNvidiaAcceptancePackResult(t *testing.T) {
 }
 
 func TestRunSATDefaultsToExportDir(t *testing.T) {
-	t.Parallel()
-
 	oldSATBaseDir := DefaultSATBaseDir
 	DefaultSATBaseDir = "/tmp/export/bee-sat"
 	t.Cleanup(func() { DefaultSATBaseDir = oldSATBaseDir })
@@ -754,6 +812,29 @@ func TestBuildSupportBundleIncludesExportDirContents(t *testing.T) {
 		}
 	}
 
+	for _, want := range []string{
+		"/system/ip-link.txt",
+		"/system/ip-link-stats.txt",
+		"/system/kernel-aer-nvidia.txt",
+		"/system/lspci-nvidia-bridges-vv.txt",
+		"/system/pcie-aer-sysfs.txt",
+		"/system/ethtool-info.txt",
+		"/system/ethtool-link.txt",
+		"/system/ethtool-module.txt",
+		"/system/mstflint-query.txt",
+	} {
+		var found bool
+		for _, name := range names {
+			if contains(name, want) {
+				found = true
+				break
+			}
+		}
+		if !found {
+			t.Fatalf("support bundle missing %s, names=%v", want, names)
+		}
+	}
+
 	var foundRaw bool
 	for _, name := range names {
 		if contains(name, "/export/bee-sat/memory-run/verbose.log") {

audit/internal/app/atomic_write.go

@@ -0,0 +1,48 @@
+package app
+
+import (
+	"fmt"
+	"os"
+	"path/filepath"
+)
+
+func atomicWriteFile(path string, data []byte, perm os.FileMode) error {
+	if err := os.MkdirAll(filepath.Dir(path), 0755); err != nil {
+		return fmt.Errorf("mkdir %s: %w", filepath.Dir(path), err)
+	}
+
+	tmpPath := path + ".tmp"
+	f, err := os.OpenFile(tmpPath, os.O_CREATE|os.O_WRONLY|os.O_TRUNC, perm)
+	if err != nil {
+		return fmt.Errorf("open temp %s: %w", tmpPath, err)
+	}
+
+	success := false
+	defer func() {
+		_ = f.Close()
+		if !success {
+			_ = os.Remove(tmpPath)
+		}
+	}()
+
+	if _, err := f.Write(data); err != nil {
+		return fmt.Errorf("write temp %s: %w", tmpPath, err)
+	}
+	if err := f.Sync(); err != nil {
+		return fmt.Errorf("sync temp %s: %w", tmpPath, err)
+	}
+	if err := f.Close(); err != nil {
+		return fmt.Errorf("close temp %s: %w", tmpPath, err)
+	}
+	if err := os.Rename(tmpPath, path); err != nil {
+		return fmt.Errorf("rename %s -> %s: %w", tmpPath, path, err)
+	}
+
+	if dir, err := os.Open(filepath.Dir(path)); err == nil {
+		_ = dir.Sync()
+		_ = dir.Close()
+	}
+
+	success = true
+	return nil
+}

audit/internal/app/atomic_write_test.go

@@ -0,0 +1,71 @@
+package app
+
+import (
+	"encoding/json"
+	"os"
+	"path/filepath"
+	"testing"
+
+	"bee/audit/internal/schema"
+)
+
+func TestAtomicWriteFileReplacesTargetWithoutLeavingTmp(t *testing.T) {
+	path := filepath.Join(t.TempDir(), "bee-audit.json")
+	if err := os.WriteFile(path, []byte("old\n"), 0644); err != nil {
+		t.Fatalf("seed file: %v", err)
+	}
+
+	if err := atomicWriteFile(path, []byte("new\n"), 0644); err != nil {
+		t.Fatalf("atomicWriteFile: %v", err)
+	}
+
+	raw, err := os.ReadFile(path)
+	if err != nil {
+		t.Fatalf("read final: %v", err)
+	}
+	if string(raw) != "new\n" {
+		t.Fatalf("final content=%q want %q", string(raw), "new\n")
+	}
+	if _, err := os.Stat(path + ".tmp"); !os.IsNotExist(err) {
+		t.Fatalf("tmp file should be absent after success, err=%v", err)
+	}
+}
+
+func TestRunRuntimePreflightWritesAtomically(t *testing.T) {
+	path := filepath.Join(t.TempDir(), "runtime-health.json")
+	a := &App{
+		runtime: fakeRuntime{
+			collectFn: func(exportDir string) (schema.RuntimeHealth, error) {
+				return schema.RuntimeHealth{
+					Status:      "OK",
+					ExportDir:   exportDir,
+					DriverReady: true,
+					CUDAReady:   true,
+				}, nil
+			},
+		},
+	}
+
+	got, err := a.RunRuntimePreflight("file:" + path)
+	if err != nil {
+		t.Fatalf("RunRuntimePreflight: %v", err)
+	}
+	if got != path {
+		t.Fatalf("path=%q want %q", got, path)
+	}
+	if _, err := os.Stat(path + ".tmp"); !os.IsNotExist(err) {
+		t.Fatalf("tmp file should be absent after success, err=%v", err)
+	}
+
+	raw, err := os.ReadFile(path)
+	if err != nil {
+		t.Fatalf("read runtime file: %v", err)
+	}
+	var health schema.RuntimeHealth
+	if err := json.Unmarshal(raw, &health); err != nil {
+		t.Fatalf("json unmarshal: %v", err)
+	}
+	if health.Status != "OK" {
+		t.Fatalf("status=%q want OK", health.Status)
+	}
+}

audit/internal/app/component_status_db.go

@@ -0,0 +1,268 @@
+package app
+
+import (
+	"encoding/json"
+	"os"
+	"path/filepath"
+	"strings"
+	"sync"
+	"time"
+)
+
+// ComponentStatusDB is a persistent, append-only store of hardware component health records.
+// Records are keyed by component identity strings (e.g. "pcie:0000:c8:00.0", "storage:nvme0n1").
+// Once a component is marked Warning or Critical, subsequent OK entries do not downgrade it —
+// the component stays at the highest observed severity until explicitly reset.
+type ComponentStatusDB struct {
+	path    string
+	mu      sync.Mutex
+	records map[string]*ComponentStatusRecord
+}
+
+// ComponentStatusRecord holds the current and historical health of one hardware component.
+type ComponentStatusRecord struct {
+	ComponentKey  string                 `json:"component_key"`
+	Status        string                 `json:"status"` // "OK", "Warning", "Critical", "Unknown"
+	LastCheckedAt time.Time              `json:"last_checked_at"`
+	LastChangedAt time.Time              `json:"last_changed_at"`
+	ErrorSummary  string                 `json:"error_summary,omitempty"`
+	History       []ComponentStatusEntry `json:"history"`
+}
+
+// ComponentStatusEntry is one observation written to a component's history.
+type ComponentStatusEntry struct {
+	At     time.Time `json:"at"`
+	Status string    `json:"status"`
+	Source string    `json:"source"` // e.g. "sat:nvidia", "sat:memory", "watchdog:kmsg"
+	Detail string    `json:"detail,omitempty"`
+}
+
+// OpenComponentStatusDB opens (or creates) the JSON status DB at path.
+func OpenComponentStatusDB(path string) (*ComponentStatusDB, error) {
+	db := &ComponentStatusDB{
+		path:    path,
+		records: make(map[string]*ComponentStatusRecord),
+	}
+	if err := os.MkdirAll(filepath.Dir(path), 0755); err != nil {
+		return nil, err
+	}
+	data, err := os.ReadFile(path)
+	if err != nil && !os.IsNotExist(err) {
+		return nil, err
+	}
+	if len(data) > 0 {
+		var records []ComponentStatusRecord
+		if err := json.Unmarshal(data, &records); err == nil {
+			for i := range records {
+				db.records[records[i].ComponentKey] = &records[i]
+			}
+		}
+	}
+	return db, nil
+}
+
+// Record writes one observation for the given component key.
+// source is a short label like "sat:nvidia" or "watchdog:kmsg".
+// status is "OK", "Warning", "Critical", or "Unknown".
+// OK never downgrades an existing Warning or Critical status.
+func (db *ComponentStatusDB) Record(key, source, status, detail string) {
+	if db == nil || strings.TrimSpace(key) == "" {
+		return
+	}
+	db.mu.Lock()
+	defer db.mu.Unlock()
+
+	now := time.Now().UTC()
+	rec, exists := db.records[key]
+	if !exists {
+		rec = &ComponentStatusRecord{ComponentKey: key}
+		db.records[key] = rec
+	}
+	rec.LastCheckedAt = now
+
+	entry := ComponentStatusEntry{At: now, Status: status, Source: source, Detail: detail}
+	rec.History = append(rec.History, entry)
+
+	// Status merge: OK never downgrades Warning/Critical.
+	newSev := componentSeverity(status)
+	curSev := componentSeverity(rec.Status)
+	if newSev > curSev {
+		rec.Status = status
+		rec.LastChangedAt = now
+		rec.ErrorSummary = detail
+	} else if rec.Status == "" {
+		rec.Status = status
+		rec.LastChangedAt = now
+	}
+
+	_ = db.saveLocked()
+}
+
+// Get returns the current record for a component key.
+func (db *ComponentStatusDB) Get(key string) (ComponentStatusRecord, bool) {
+	if db == nil {
+		return ComponentStatusRecord{}, false
+	}
+	db.mu.Lock()
+	defer db.mu.Unlock()
+	r, ok := db.records[key]
+	if !ok {
+		return ComponentStatusRecord{}, false
+	}
+	return *r, true
+}
+
+// All returns a snapshot of all records.
+func (db *ComponentStatusDB) All() []ComponentStatusRecord {
+	if db == nil {
+		return nil
+	}
+	db.mu.Lock()
+	defer db.mu.Unlock()
+	out := make([]ComponentStatusRecord, 0, len(db.records))
+	for _, r := range db.records {
+		out = append(out, *r)
+	}
+	return out
+}
+
+func (db *ComponentStatusDB) saveLocked() error {
+	records := make([]ComponentStatusRecord, 0, len(db.records))
+	for _, r := range db.records {
+		records = append(records, *r)
+	}
+	data, err := json.MarshalIndent(records, "", "  ")
+	if err != nil {
+		return err
+	}
+	return os.WriteFile(db.path, data, 0644)
+}
+
+// componentSeverity returns a numeric severity so higher values win.
+func componentSeverity(status string) int {
+	switch strings.TrimSpace(status) {
+	case "Critical":
+		return 3
+	case "Warning":
+		return 2
+	case "OK":
+		return 1
+	default:
+		return 0
+	}
+}
+
+// ApplySATResultToDB reads a SAT summary.txt from the run directory next to archivePath
+// and writes component status records to db for the given SAT target.
+// archivePath may be either a bare .tar.gz path or "Archive written to /path/foo.tar.gz".
+func ApplySATResultToDB(db *ComponentStatusDB, target, archivePath string) {
+	if db == nil || strings.TrimSpace(archivePath) == "" {
+		return
+	}
+	archivePath = extractArchivePath(archivePath)
+	if archivePath == "" {
+		return
+	}
+	runDir := strings.TrimSuffix(archivePath, ".tar.gz")
+	data, err := os.ReadFile(filepath.Join(runDir, "summary.txt"))
+	if err != nil {
+		return
+	}
+	kv := parseSATKV(string(data))
+	overall := strings.ToUpper(strings.TrimSpace(kv["overall_status"]))
+	if overall == "" {
+		return
+	}
+
+	source := "sat:" + target
+	dbStatus := satStatusToDBStatus(overall)
+
+	// Map SAT target to component keys.
+	switch target {
+	case "nvidia", "nvidia-targeted-stress", "nvidia-compute", "nvidia-targeted-power", "nvidia-pulse",
+		"nvidia-interconnect", "nvidia-bandwidth", "amd", "nvidia-stress",
+		"amd-stress", "amd-mem", "amd-bandwidth":
+		db.Record("pcie:gpu:"+target, source, dbStatus, target+" SAT: "+overall)
+	case "memory", "memory-stress", "sat-stress":
+		db.Record("memory:all", source, dbStatus, target+" SAT: "+overall)
+	case "cpu", "platform-stress":
+		db.Record("cpu:all", source, dbStatus, target+" SAT: "+overall)
+	case "storage":
+		// Try to record per-device if available in summary.
+		recordedAny := false
+		for key, val := range kv {
+			if !strings.HasSuffix(key, "_status") || key == "overall_status" {
+				continue
+			}
+			base := strings.TrimSuffix(key, "_status")
+			idx := strings.Index(base, "_")
+			if idx <= 0 {
+				continue
+			}
+			devName := base[:idx]
+			devStatus := satStatusToDBStatus(strings.ToUpper(strings.TrimSpace(val)))
+			db.Record("storage:"+devName, source, devStatus, "storage SAT: "+val)
+			recordedAny = true
+		}
+		if !recordedAny {
+			db.Record("storage:all", source, dbStatus, "storage SAT: "+overall)
+		}
+	}
+}
+
+func satStatusToDBStatus(overall string) string {
+	switch overall {
+	case "OK":
+		return "OK"
+	case "FAILED":
+		return "Warning"
+	case "PARTIAL", "UNSUPPORTED":
+		return "Unknown"
+	default:
+		return "Unknown"
+	}
+}
+
+// ExtractArchivePath extracts a bare .tar.gz path from a string that may be
+// "Archive written to /path/foo.tar.gz" or already a bare path.
+func ExtractArchivePath(s string) string {
+	return extractArchivePath(s)
+}
+
+// ReadSATOverallStatus reads the overall_status value from the summary.txt
+// file located in the run directory alongside archivePath.
+// Returns "" if the file cannot be read.
+func ReadSATOverallStatus(archivePath string) string {
+	if strings.TrimSpace(archivePath) == "" {
+		return ""
+	}
+	runDir := strings.TrimSuffix(archivePath, ".tar.gz")
+	data, err := os.ReadFile(filepath.Join(runDir, "summary.txt"))
+	if err != nil {
+		return ""
+	}
+	kv := parseSATKV(string(data))
+	return strings.ToUpper(strings.TrimSpace(kv["overall_status"]))
+}
+
+func extractArchivePath(s string) string {
+	s = strings.TrimSpace(s)
+	if strings.HasSuffix(s, ".tar.gz") {
+		parts := strings.Fields(s)
+		if len(parts) > 0 {
+			return parts[len(parts)-1]
+		}
+	}
+	return s
+}
+
+func parseSATKV(raw string) map[string]string {
+	kv := make(map[string]string)
+	for _, line := range strings.Split(raw, "\n") {
+		k, v, ok := strings.Cut(strings.TrimSpace(line), "=")
+		if ok {
+			kv[strings.TrimSpace(k)] = strings.TrimSpace(v)
+		}
+	}
+	return kv
+}

audit/internal/app/sat_overlay.go

@@ -3,13 +3,14 @@ package app
 import (
 	"os"
 	"path/filepath"
+	"strconv"
 	"sort"
 	"strings"
 
 	"bee/audit/internal/schema"
 )
 
-func applyLatestSATStatuses(snap *schema.HardwareSnapshot, baseDir string) {
+func applyLatestSATStatuses(snap *schema.HardwareSnapshot, baseDir string, db *ComponentStatusDB) {
 	if snap == nil || strings.TrimSpace(baseDir) == "" {
 		return
 	}
@@ -18,6 +19,7 @@ func applyLatestSATStatuses(snap *schema.HardwareSnapshot, baseDir string) {
 	}
 	if summary, ok := loadLatestSATSummary(baseDir, "gpu-nvidia-"); ok {
 		applyGPUVendorSAT(snap.PCIeDevices, "nvidia", summary)
+		applyNvidiaPerGPUStatus(snap.PCIeDevices, baseDir)
 	}
 	if summary, ok := loadLatestSATSummary(baseDir, "memory-"); ok {
 		applyMemorySAT(snap.Memory, summary)
@@ -28,6 +30,102 @@ func applyLatestSATStatuses(snap *schema.HardwareSnapshot, baseDir string) {
 	if summary, ok := loadLatestSATSummary(baseDir, "storage-"); ok {
 		applyStorageSAT(snap.Storage, summary)
 	}
+	// Apply unified component status DB — overlaid last so it can only upgrade severity.
+	applyComponentStatusDB(snap, db)
+}
+
+type nvidiaPerGPUStatus struct {
+	runStatus string
+	reason    string
+}
+
+func applyNvidiaPerGPUStatus(devs []schema.HardwarePCIeDevice, baseDir string) {
+	statusByIndex, ts, ok := loadLatestNvidiaPerGPUStatus(baseDir)
+	if !ok {
+		return
+	}
+	for i := range devs {
+		if devs[i].Telemetry == nil {
+			continue
+		}
+		rawIdx, ok := devs[i].Telemetry["nvidia_gpu_index"]
+		if !ok {
+			continue
+		}
+		idx, ok := telemetryInt(rawIdx)
+		if !ok {
+			continue
+		}
+		st, ok := statusByIndex[idx]
+		if !ok {
+			continue
+		}
+		status, description, ok := satKeyStatus(st.runStatus, firstNonEmpty(strings.TrimSpace(st.reason), "nvidia GPU SAT"))
+		if !ok {
+			continue
+		}
+		mergeComponentStatusPreferDetail(&devs[i].HardwareComponentStatus, ts, status, description)
+	}
+}
+
+func loadLatestNvidiaPerGPUStatus(baseDir string) (map[int]nvidiaPerGPUStatus, string, bool) {
+	matches, err := filepath.Glob(filepath.Join(baseDir, "gpu-nvidia-*"))
+	if err != nil || len(matches) == 0 {
+		return nil, "", false
+	}
+	sort.Strings(matches)
+	runDir := matches[len(matches)-1]
+	summaryRaw, err := os.ReadFile(filepath.Join(runDir, "summary.txt"))
+	if err != nil {
+		return nil, "", false
+	}
+	summaryKV := parseKeyValueSummary(string(summaryRaw))
+	runAtUTC := strings.TrimSpace(summaryKV["run_at_utc"])
+	files, err := filepath.Glob(filepath.Join(runDir, "gpu-*-status.txt"))
+	if err != nil || len(files) == 0 {
+		return nil, "", false
+	}
+	out := make(map[int]nvidiaPerGPUStatus, len(files))
+	for _, file := range files {
+		raw, err := os.ReadFile(file)
+		if err != nil {
+			continue
+		}
+		kv := parseKeyValueSummary(string(raw))
+		idx, err := strconv.Atoi(strings.TrimSpace(kv["gpu_index"]))
+		if err != nil {
+			continue
+		}
+		out[idx] = nvidiaPerGPUStatus{
+			runStatus: strings.ToUpper(strings.TrimSpace(kv["run_status"])),
+			reason:    strings.TrimSpace(kv["reason"]),
+		}
+	}
+	if len(out) == 0 {
+		return nil, "", false
+	}
+	return out, runAtUTC, true
+}
+
+func telemetryInt(v any) (int, bool) {
+	switch value := v.(type) {
+	case int:
+		return value, true
+	case int32:
+		return int(value), true
+	case int64:
+		return int(value), true
+	case float64:
+		return int(value), true
+	case string:
+		n, err := strconv.Atoi(strings.TrimSpace(value))
+		if err != nil {
+			return 0, false
+		}
+		return n, true
+	default:
+		return 0, false
+	}
 }
 
 type satSummary struct {
@@ -174,6 +272,31 @@ func mergeComponentStatus(component *schema.HardwareComponentStatus, changedAt,
 	}
 }
 
+func mergeComponentStatusPreferDetail(component *schema.HardwareComponentStatus, changedAt, satStatus, description string) {
+	if component == nil || satStatus == "" {
+		return
+	}
+	current := strings.TrimSpace(ptrString(component.Status))
+	newSeverity := statusSeverity(satStatus)
+	currentSeverity := statusSeverity(current)
+	if current == "" || current == "Unknown" || newSeverity > currentSeverity {
+		mergeComponentStatus(component, changedAt, satStatus, description)
+		return
+	}
+	if newSeverity == currentSeverity && strings.TrimSpace(description) != "" {
+		component.Status = appStringPtr(satStatus)
+		component.ErrorDescription = appStringPtr(description)
+		if strings.TrimSpace(changedAt) != "" {
+			component.StatusChangedAt = appStringPtr(changedAt)
+			component.StatusHistory = append(component.StatusHistory, schema.HardwareStatusHistory{
+				Status:    satStatus,
+				ChangedAt: changedAt,
+				Details:   appStringPtr(description),
+			})
+		}
+	}
+}
+
 func statusSeverity(status string) int {
 	switch strings.TrimSpace(status) {
 	case "Critical":
@@ -206,6 +329,86 @@ func matchesGPUVendor(dev schema.HardwarePCIeDevice, vendor string) bool {
 	}
 }
 
+func applyComponentStatusDB(snap *schema.HardwareSnapshot, db *ComponentStatusDB) {
+	if snap == nil || db == nil {
+		return
+	}
+	for _, rec := range db.All() {
+		key := rec.ComponentKey
+		status := dbStatusToSATStatus(rec.Status)
+		if status == "" {
+			continue
+		}
+		detail := rec.ErrorSummary
+		ts := rec.LastChangedAt.UTC().Format("2006-01-02T15:04:05Z")
+
+		switch {
+		case strings.HasPrefix(key, "pcie:"):
+			bdf := strings.TrimPrefix(key, "pcie:")
+			bdf = strings.TrimPrefix(bdf, "gpu:") // strip sub-type if present
+			// bdf may be empty (e.g. "pcie:gpu:nvidia") — skip BDF matching
+			if sanitizeBDFForLookup(bdf) == "" {
+				break
+			}
+			normalized := sanitizeBDFForLookup(bdf)
+			for i := range snap.PCIeDevices {
+				if snap.PCIeDevices[i].BDF == nil {
+					continue
+				}
+				if sanitizeBDFForLookup(*snap.PCIeDevices[i].BDF) == normalized {
+					mergeComponentStatus(&snap.PCIeDevices[i].HardwareComponentStatus, ts, status, detail)
+				}
+			}
+		case strings.HasPrefix(key, "storage:"):
+			devName := strings.TrimPrefix(key, "storage:")
+			if devName == "all" {
+				for i := range snap.Storage {
+					mergeComponentStatus(&snap.Storage[i].HardwareComponentStatus, ts, status, detail)
+				}
+			} else {
+				for i := range snap.Storage {
+					linuxDev, _ := snap.Storage[i].Telemetry["linux_device"].(string)
+					if filepath.Base(strings.TrimSpace(linuxDev)) == devName {
+						mergeComponentStatus(&snap.Storage[i].HardwareComponentStatus, ts, status, detail)
+					}
+				}
+			}
+		case strings.HasPrefix(key, "memory:"):
+			for i := range snap.Memory {
+				mergeComponentStatus(&snap.Memory[i].HardwareComponentStatus, ts, status, detail)
+			}
+		case strings.HasPrefix(key, "cpu:"):
+			for i := range snap.CPUs {
+				mergeComponentStatus(&snap.CPUs[i].HardwareComponentStatus, ts, status, detail)
+			}
+		}
+	}
+}
+
+// dbStatusToSATStatus converts ComponentStatusDB status strings to the format
+// expected by mergeComponentStatus (which uses "OK", "Warning", "Critical", "Unknown").
+func dbStatusToSATStatus(s string) string {
+	switch strings.TrimSpace(s) {
+	case "OK", "Warning", "Critical", "Unknown":
+		return s
+	default:
+		return ""
+	}
+}
+
+// sanitizeBDFForLookup normalises a PCIe BDF address to a canonical lower-case form
+// suitable for comparison. "c8:00.0" → "0000:c8:00.0"; already-full BDFs are left as-is.
+func sanitizeBDFForLookup(bdf string) string {
+	bdf = strings.ToLower(strings.TrimSpace(bdf))
+	if bdf == "" || bdf == "gpu" || strings.ContainsAny(bdf, " \t") {
+		return ""
+	}
+	if strings.Count(bdf, ":") == 1 {
+		bdf = "0000:" + bdf
+	}
+	return bdf
+}
+
 func ptrString(v *string) string {
 	if v == nil {
 		return ""

audit/internal/app/sat_overlay_test.go

@@ -23,7 +23,7 @@ func TestApplyLatestSATStatusesMarksStorageByDevice(t *testing.T) {
 	usb := schema.HardwareStorage{Telemetry: map[string]any{"linux_device": "/dev/sda"}}
 	snap := schema.HardwareSnapshot{Storage: []schema.HardwareStorage{nvme, usb}}
 
-	applyLatestSATStatuses(&snap, baseDir)
+	applyLatestSATStatuses(&snap, baseDir, nil)
 
 	if snap.Storage[0].Status == nil || *snap.Storage[0].Status != "OK" {
 		t.Fatalf("nvme status=%v want OK", snap.Storage[0].Status)
@@ -53,9 +53,57 @@ func TestApplyLatestSATStatusesMarksAMDGPUs(t *testing.T) {
 		}},
 	}
 
-	applyLatestSATStatuses(&snap, baseDir)
+	applyLatestSATStatuses(&snap, baseDir, nil)
 
 	if snap.PCIeDevices[0].Status == nil || *snap.PCIeDevices[0].Status != "Critical" {
 		t.Fatalf("gpu status=%v want Critical", snap.PCIeDevices[0].Status)
 	}
 }
+
+func TestApplyLatestSATStatusesMarksNvidiaGPUByPerGPUStatusFile(t *testing.T) {
+	baseDir := t.TempDir()
+	runDir := filepath.Join(baseDir, "gpu-nvidia-20260407-162123")
+	if err := os.MkdirAll(runDir, 0755); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(filepath.Join(runDir, "summary.txt"), []byte("run_at_utc=2026-04-07T16:21:23Z\noverall_status=FAILED\n"), 0644); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(filepath.Join(runDir, "gpu-1-status.txt"), []byte("gpu_index=1\ngpu_name=NVIDIA H100 PCIe\nrun_status=FAILED\nreason=GPU requires reset\n"), 0644); err != nil {
+		t.Fatal(err)
+	}
+
+	class := "VideoController"
+	manufacturer := "NVIDIA Corporation"
+	bdf0 := "0000:4b:00.0"
+	bdf1 := "0000:4f:00.0"
+	snap := schema.HardwareSnapshot{
+		PCIeDevices: []schema.HardwarePCIeDevice{
+			{
+				DeviceClass:  &class,
+				Manufacturer: &manufacturer,
+				BDF:          &bdf0,
+				Telemetry:    map[string]any{"nvidia_gpu_index": 0},
+			},
+			{
+				DeviceClass:  &class,
+				Manufacturer: &manufacturer,
+				BDF:          &bdf1,
+				Telemetry:    map[string]any{"nvidia_gpu_index": 1},
+			},
+		},
+	}
+
+	applyLatestSATStatuses(&snap, baseDir, nil)
+
+	if snap.PCIeDevices[1].Status == nil || *snap.PCIeDevices[1].Status != "Critical" {
+		t.Fatalf("gpu1 status=%v want Critical", snap.PCIeDevices[1].Status)
+	}
+	if snap.PCIeDevices[1].ErrorDescription == nil || *snap.PCIeDevices[1].ErrorDescription != "GPU requires reset failed" {
+		got := "<nil>"
+		if snap.PCIeDevices[1].ErrorDescription != nil {
+			got = *snap.PCIeDevices[1].ErrorDescription
+		}
+		t.Fatalf("gpu1 error=%q want per-gpu reason", got)
+	}
+}

audit/internal/app/support_bundle.go

@@ -19,6 +19,8 @@ var supportBundleServices = []string{
 	"bee-network.service",
 	"bee-nvidia.service",
 	"bee-preflight.service",
+	"bee-selfheal.service",
+	"bee-selfheal.timer",
 	"bee-sshsetup.service",
 }
 
@@ -32,21 +34,158 @@ var supportBundleCommands = []struct {
 	{name: "system/lspci-nn.txt", cmd: []string{"lspci", "-nn"}},
 	{name: "system/lspci-vvv.txt", cmd: []string{"lspci", "-vvv"}},
 	{name: "system/ip-addr.txt", cmd: []string{"ip", "addr"}},
+	{name: "system/ip-link.txt", cmd: []string{"ip", "-details", "link", "show"}},
+	{name: "system/ip-link-stats.txt", cmd: []string{"ip", "-s", "link", "show"}},
 	{name: "system/ip-route.txt", cmd: []string{"ip", "route"}},
 	{name: "system/mount.txt", cmd: []string{"mount"}},
 	{name: "system/df-h.txt", cmd: []string{"df", "-h"}},
 	{name: "system/dmesg.txt", cmd: []string{"dmesg"}},
+	{name: "system/kernel-aer-nvidia.txt", cmd: []string{"sh", "-c", `
+if command -v dmesg >/dev/null 2>&1; then
+  dmesg | grep -iE 'AER|NVRM|Xid|pcieport|nvidia' || echo "no AER/NVRM/Xid kernel messages found"
+else
+  echo "dmesg not found"
+fi
+`}},
 	{name: "system/nvidia-smi-q.txt", cmd: []string{"nvidia-smi", "-q"}},
+	{name: "system/lspci-nvidia-bridges-vv.txt", cmd: []string{"sh", "-c", `
+if ! command -v lspci >/dev/null 2>&1; then
+  echo "lspci not found"
+  exit 0
+fi
+found=0
+	for gpu in $(lspci -Dn | awk '$2 ~ /^03(00|02):$/ && $3 ~ /^10de:/ {print $1}'); do
+  found=1
+  echo "=== GPU $gpu ==="
+  lspci -s "$gpu" -vv 2>&1 || true
+  bridge=$(basename "$(readlink -f "/sys/bus/pci/devices/$gpu/.." 2>/dev/null)" 2>/dev/null)
+  if [ -n "$bridge" ] && [ "$bridge" != "$gpu" ]; then
+    echo
+    echo "=== UPSTREAM $bridge for $gpu ==="
+    lspci -s "$bridge" -vv 2>&1 || true
+  fi
+  echo
+done
+if [ "$found" -eq 0 ]; then
+  echo "no NVIDIA PCI devices found"
+fi
+`}},
 	{name: "system/pcie-nvidia-link.txt", cmd: []string{"sh", "-c", `
 for d in /sys/bus/pci/devices/*/; do
   vendor=$(cat "$d/vendor" 2>/dev/null)
-  [ "$vendor" = "0x10de" ] || continue
-  dev=$(basename "$d")
+	  [ "$vendor" = "0x10de" ] || continue
+	  class=$(cat "$d/class" 2>/dev/null)
+	  case "$class" in
+	    0x030000|0x030200) ;;
+	    *) continue ;;
+	  esac
+	  dev=$(basename "$d")
   echo "=== $dev ==="
   for f in current_link_speed current_link_width max_link_speed max_link_width; do
     printf "  %-22s %s\n" "$f" "$(cat "$d/$f" 2>/dev/null)"
   done
 done
+`}},
+	{name: "system/pcie-aer-sysfs.txt", cmd: []string{"sh", "-c", `
+found=0
+for dev in /sys/bus/pci/devices/*; do
+  [ -e "$dev" ] || continue
+  bdf=$(basename "$dev")
+  block=""
+  for f in aer_dev_correctable aer_dev_fatal aer_dev_nonfatal aer_rootport_total_err_cor aer_rootport_total_err_fatal aer_rootport_total_err_nonfatal; do
+    if [ -r "$dev/$f" ]; then
+      if [ -z "$block" ]; then
+        block=1
+        found=1
+        echo "=== $bdf ==="
+      fi
+      printf "  %-30s %s\n" "$f" "$(cat "$dev/$f" 2>/dev/null)"
+    fi
+  done
+  if [ -n "$block" ]; then
+    echo
+  fi
+done
+if [ "$found" -eq 0 ]; then
+  echo "no PCIe AER sysfs counters found"
+fi
+`}},
+	{name: "system/ethtool-info.txt", cmd: []string{"sh", "-c", `
+if ! command -v ethtool >/dev/null 2>&1; then
+  echo "ethtool not found"
+  exit 0
+fi
+found=0
+for path in /sys/class/net/*; do
+  [ -e "$path" ] || continue
+  iface=$(basename "$path")
+  [ "$iface" = "lo" ] && continue
+  found=1
+  echo "=== $iface ==="
+  ethtool -i "$iface" 2>&1 || true
+  echo
+done
+if [ "$found" -eq 0 ]; then
+  echo "no interfaces found"
+fi
+`}},
+	{name: "system/ethtool-link.txt", cmd: []string{"sh", "-c", `
+if ! command -v ethtool >/dev/null 2>&1; then
+  echo "ethtool not found"
+  exit 0
+fi
+found=0
+for path in /sys/class/net/*; do
+  [ -e "$path" ] || continue
+  iface=$(basename "$path")
+  [ "$iface" = "lo" ] && continue
+  found=1
+  echo "=== $iface ==="
+  ethtool "$iface" 2>&1 || true
+  echo
+done
+if [ "$found" -eq 0 ]; then
+  echo "no interfaces found"
+fi
+`}},
+	{name: "system/ethtool-module.txt", cmd: []string{"sh", "-c", `
+if ! command -v ethtool >/dev/null 2>&1; then
+  echo "ethtool not found"
+  exit 0
+fi
+found=0
+for path in /sys/class/net/*; do
+  [ -e "$path" ] || continue
+  iface=$(basename "$path")
+  [ "$iface" = "lo" ] && continue
+  found=1
+  echo "=== $iface ==="
+  ethtool -m "$iface" 2>&1 || true
+  echo
+done
+if [ "$found" -eq 0 ]; then
+  echo "no interfaces found"
+fi
+`}},
+	{name: "system/mstflint-query.txt", cmd: []string{"sh", "-c", `
+if ! command -v mstflint >/dev/null 2>&1; then
+  echo "mstflint not found"
+  exit 0
+fi
+found=0
+for path in /sys/bus/pci/devices/*; do
+  [ -e "$path/vendor" ] || continue
+  vendor=$(cat "$path/vendor" 2>/dev/null)
+  [ "$vendor" = "0x15b3" ] || continue
+  bdf=$(basename "$path")
+  found=1
+  echo "=== $bdf ==="
+  mstflint -d "$bdf" q 2>&1 || true
+  echo
+done
+if [ "$found" -eq 0 ]; then
+  echo "no Mellanox/NVIDIA networking devices found"
+fi
 `}},
 }
 
@@ -58,7 +197,7 @@ var supportBundleOptionalFiles = []struct {
 	{name: "system/syslog.txt", src: "/var/log/syslog"},
 }
 
-const supportBundleGlob = "bee-support-*.tar.gz"
+const supportBundleGlob = "????-??-?? (BEE-SP*)*.tar.gz"
 
 func BuildSupportBundle(exportDir string) (string, error) {
 	exportDir = strings.TrimSpace(exportDir)
@@ -72,9 +211,14 @@ func BuildSupportBundle(exportDir string) (string, error) {
 		return "", err
 	}
 
-	host := sanitizeFilename(hostnameOr("unknown"))
-	ts := time.Now().UTC().Format("20060102-150405")
-	stageRoot := filepath.Join(os.TempDir(), fmt.Sprintf("bee-support-%s-%s", host, ts))
+	now := time.Now().UTC()
+	date := now.Format("2006-01-02")
+	tod := now.Format("15:04:05")
+	ver := bundleVersion()
+	model := serverModelForBundle()
+	sn := serverSerialForBundle()
+
+	stageRoot := filepath.Join(os.TempDir(), fmt.Sprintf("bee-support-stage-%s-%s", sanitizeFilename(hostnameOr("unknown")), now.Format("20060102-150405")))
 	if err := os.MkdirAll(stageRoot, 0755); err != nil {
 		return "", err
 	}
@@ -106,7 +250,8 @@ func BuildSupportBundle(exportDir string) (string, error) {
 		return "", err
 	}
 
-	archivePath := filepath.Join(os.TempDir(), fmt.Sprintf("bee-support-%s-%s.tar.gz", host, ts))
+	archiveName := fmt.Sprintf("%s (BEE-SP v%s) %s %s %s.tar.gz", date, ver, model, sn, tod)
+	archivePath := filepath.Join(os.TempDir(), archiveName)
 	if err := createSupportTarGz(archivePath, stageRoot); err != nil {
 		return "", err
 	}
@@ -263,6 +408,60 @@ func writeManifest(dst, exportDir, stageRoot string) error {
 	return os.WriteFile(dst, []byte(body.String()), 0644)
 }
 
+func bundleVersion() string {
+	v := buildVersion()
+	v = strings.TrimPrefix(v, "v")
+	v = strings.TrimPrefix(v, "V")
+	if v == "" || v == "unknown" {
+		return "0.0"
+	}
+	return v
+}
+
+func serverModelForBundle() string {
+	raw, err := exec.Command("dmidecode", "-t", "1").Output()
+	if err != nil {
+		return "unknown"
+	}
+	for _, line := range strings.Split(string(raw), "\n") {
+		line = strings.TrimSpace(line)
+		key, val, ok := strings.Cut(line, ": ")
+		if !ok {
+			continue
+		}
+		if strings.TrimSpace(key) == "Product Name" {
+			val = strings.TrimSpace(val)
+			if val == "" {
+				return "unknown"
+			}
+			return strings.ReplaceAll(val, " ", "_")
+		}
+	}
+	return "unknown"
+}
+
+func serverSerialForBundle() string {
+	raw, err := exec.Command("dmidecode", "-t", "1").Output()
+	if err != nil {
+		return "unknown"
+	}
+	for _, line := range strings.Split(string(raw), "\n") {
+		line = strings.TrimSpace(line)
+		key, val, ok := strings.Cut(line, ": ")
+		if !ok {
+			continue
+		}
+		if strings.TrimSpace(key) == "Serial Number" {
+			val = strings.TrimSpace(val)
+			if val == "" {
+				return "unknown"
+			}
+			return val
+		}
+	}
+	return "unknown"
+}
+
 func buildVersion() string {
 	raw, err := exec.Command("bee", "version").CombinedOutput()
 	if err != nil {

audit/internal/collector/nic_mellanox.go

@@ -2,18 +2,21 @@ package collector
 
 import (
 	"bee/audit/internal/schema"
+	"context"
 	"log/slog"
 	"os"
 	"os/exec"
 	"path/filepath"
 	"strings"
+	"time"
 )
 
 const mellanoxVendorID = 0x15b3
+const nicProbeTimeout = 2 * time.Second
 
 var (
 	mstflintQuery = func(bdf string) (string, error) {
-		out, err := exec.Command("mstflint", "-d", bdf, "q").Output()
+		out, err := commandOutputWithTimeout(nicProbeTimeout, "mstflint", "-d", bdf, "q")
 		if err != nil {
 			return "", err
 		}
@@ -21,7 +24,7 @@ var (
 	}
 
 	ethtoolInfoQuery = func(iface string) (string, error) {
-		out, err := exec.Command("ethtool", "-i", iface).Output()
+		out, err := commandOutputWithTimeout(nicProbeTimeout, "ethtool", "-i", iface)
 		if err != nil {
 			return "", err
 		}
@@ -29,6 +32,14 @@ var (
 	}
 
 	netIfacesByBDF = listNetIfacesByBDF
+	readNetCarrierFile = func(iface string) (string, error) {
+		path := filepath.Join("/sys/class/net", iface, "carrier")
+		raw, err := os.ReadFile(path)
+		if err != nil {
+			return "", err
+		}
+		return strings.TrimSpace(string(raw)), nil
+	}
 )
 
 // enrichPCIeWithMellanox enriches Mellanox/NVIDIA Networking devices with
@@ -162,3 +173,17 @@ func listNetIfacesByBDF(bdf string) []string {
 	}
 	return ifaces
 }
+
+func commandOutputWithTimeout(timeout time.Duration, name string, args ...string) ([]byte, error) {
+	ctx, cancel := context.WithTimeout(context.Background(), timeout)
+	defer cancel()
+	return exec.CommandContext(ctx, name, args...).Output()
+}
+
+func interfaceHasCarrier(iface string) bool {
+	raw, err := readNetCarrierFile(iface)
+	if err != nil {
+		return false
+	}
+	return strings.TrimSpace(raw) == "1"
+}

audit/internal/collector/nic_telemetry.go

@@ -12,7 +12,7 @@ import (
 
 var (
 	ethtoolModuleQuery = func(iface string) (string, error) {
-		out, err := raidToolQuery("ethtool", "-m", iface)
+		out, err := commandOutputWithTimeout(nicProbeTimeout, "ethtool", "-m", iface)
 		if err != nil {
 			return "", err
 		}
@@ -58,10 +58,12 @@ func enrichPCIeWithNICTelemetry(devs []schema.HardwarePCIeDevice) []schema.Hardw
 			}
 		}
 
-		if out, err := ethtoolModuleQuery(iface); err == nil {
-			if injectSFPDOMTelemetry(&devs[i], out) {
-				enriched++
-				continue
+		if interfaceHasCarrier(iface) {
+			if out, err := ethtoolModuleQuery(iface); err == nil {
+				if injectSFPDOMTelemetry(&devs[i], out) {
+					enriched++
+					continue
+				}
 			}
 		}
 		if len(devs[i].MacAddresses) > 0 || devs[i].Firmware != nil {

audit/internal/collector/nic_telemetry_test.go

@@ -57,6 +57,7 @@ func TestEnrichPCIeWithNICTelemetryAddsSerialFallback(t *testing.T) {
 	origReadMAC := readNetAddressFile
 	origEth := ethtoolInfoQuery
 	origModule := ethtoolModuleQuery
+	origCarrier := readNetCarrierFile
 	t.Cleanup(func() {
 		queryPCILSPCIDetail = origDetail
 		readPCIVPDFile = origVPD
@@ -64,6 +65,7 @@ func TestEnrichPCIeWithNICTelemetryAddsSerialFallback(t *testing.T) {
 		readNetAddressFile = origReadMAC
 		ethtoolInfoQuery = origEth
 		ethtoolModuleQuery = origModule
+		readNetCarrierFile = origCarrier
 	})
 
 	queryPCILSPCIDetail = func(bdf string) (string, error) {
@@ -82,6 +84,7 @@ func TestEnrichPCIeWithNICTelemetryAddsSerialFallback(t *testing.T) {
 		}
 		return "aa:bb:cc:dd:ee:ff", nil
 	}
+	readNetCarrierFile = func(string) (string, error) { return "1", nil }
 	ethtoolInfoQuery = func(string) (string, error) { return "", fmt.Errorf("skip firmware") }
 	ethtoolModuleQuery = func(string) (string, error) { return "", fmt.Errorf("skip optics") }
 
@@ -101,6 +104,42 @@ func TestEnrichPCIeWithNICTelemetryAddsSerialFallback(t *testing.T) {
 	}
 }
 
+func TestEnrichPCIeWithNICTelemetrySkipsModuleQueryWithoutCarrier(t *testing.T) {
+	origIfaces := netIfacesByBDF
+	origReadMAC := readNetAddressFile
+	origEth := ethtoolInfoQuery
+	origModule := ethtoolModuleQuery
+	origCarrier := readNetCarrierFile
+	t.Cleanup(func() {
+		netIfacesByBDF = origIfaces
+		readNetAddressFile = origReadMAC
+		ethtoolInfoQuery = origEth
+		ethtoolModuleQuery = origModule
+		readNetCarrierFile = origCarrier
+	})
+
+	netIfacesByBDF = func(string) []string { return []string{"eth0"} }
+	readNetAddressFile = func(string) (string, error) { return "aa:bb:cc:dd:ee:ff", nil }
+	readNetCarrierFile = func(string) (string, error) { return "0", nil }
+	ethtoolInfoQuery = func(string) (string, error) { return "", fmt.Errorf("skip firmware") }
+	ethtoolModuleQuery = func(string) (string, error) {
+		t.Fatal("ethtool -m should not be called without carrier")
+		return "", nil
+	}
+
+	class := "EthernetController"
+	bdf := "0000:18:00.0"
+	devs := []schema.HardwarePCIeDevice{{
+		DeviceClass: &class,
+		BDF:         &bdf,
+	}}
+
+	out := enrichPCIeWithNICTelemetry(devs)
+	if len(out[0].MacAddresses) != 1 || out[0].MacAddresses[0] != "aa:bb:cc:dd:ee:ff" {
+		t.Fatalf("mac_addresses=%v", out[0].MacAddresses)
+	}
+}
+
 func TestDBMValue(t *testing.T) {
 	tests := []struct {
 		in   string

audit/internal/collector/nvidia.go

@@ -13,6 +13,7 @@ import (
 const nvidiaVendorID = 0x10de
 
 type nvidiaGPUInfo struct {
+	Index              int
 	BDF                string
 	Serial             string
 	VBIOS              string
@@ -132,6 +133,7 @@ func parseNVIDIASMIQuery(raw string) (map[string]nvidiaGPUInfo, error) {
 		}
 
 		info := nvidiaGPUInfo{
+			Index:              parseRequiredInt(rec[0]),
 			BDF:                bdf,
 			Serial:             strings.TrimSpace(rec[2]),
 			VBIOS:              strings.TrimSpace(rec[3]),
@@ -187,6 +189,14 @@ func parseMaybeInt(v string) *int {
 	return &n
 }
 
+func parseRequiredInt(v string) int {
+	n, err := strconv.Atoi(strings.TrimSpace(v))
+	if err != nil {
+		return 0
+	}
+	return n
+}
+
 func pcieLinkGenLabel(gen int) string {
 	return fmt.Sprintf("Gen%d", gen)
 }
@@ -240,6 +250,10 @@ func setPCIeFallback(dev *schema.HardwarePCIeDevice) {
 }
 
 func injectNVIDIATelemetry(dev *schema.HardwarePCIeDevice, info nvidiaGPUInfo) {
+	if dev.Telemetry == nil {
+		dev.Telemetry = map[string]any{}
+	}
+	dev.Telemetry["nvidia_gpu_index"] = info.Index
 	if info.TemperatureC != nil {
 		dev.TemperatureC = info.TemperatureC
 	}

audit/internal/collector/nvidia_test.go

@@ -86,6 +86,9 @@ func TestEnrichPCIeWithNVIDIAData_driverLoaded(t *testing.T) {
 	if out[0].Firmware == nil || *out[0].Firmware != "96.00.1F.00.02" {
 		t.Fatalf("firmware: got %v", out[0].Firmware)
 	}
+	if out[0].Telemetry == nil || out[0].Telemetry["nvidia_gpu_index"] != 0 {
+		t.Fatalf("telemetry nvidia_gpu_index: got %#v", out[0].Telemetry)
+	}
 	if out[0].Status == nil || *out[0].Status != statusWarning {
 		t.Fatalf("status: got %v", out[0].Status)
 	}

audit/internal/platform/benchmark_report.go

@@ -0,0 +1,344 @@
+package platform
+
+import (
+	"fmt"
+	"os"
+	"path/filepath"
+	"regexp"
+	"strings"
+	"time"
+)
+
+func renderBenchmarkReport(result NvidiaBenchmarkResult) string {
+	return renderBenchmarkReportWithCharts(result, nil)
+}
+
+type benchmarkReportChart struct {
+	Title   string
+	Content string
+}
+
+var ansiEscapePattern = regexp.MustCompile(`\x1b\[[0-9;]*m`)
+
+func renderBenchmarkReportWithCharts(result NvidiaBenchmarkResult, charts []benchmarkReportChart) string {
+	var b strings.Builder
+
+	// ── Header ────────────────────────────────────────────────────────────────
+	b.WriteString("# Bee NVIDIA Benchmark Report\n\n")
+
+	// System identity block
+	if result.ServerModel != "" {
+		fmt.Fprintf(&b, "**Server:** %s  \n", result.ServerModel)
+	}
+	if result.Hostname != "" {
+		fmt.Fprintf(&b, "**Host:** %s  \n", result.Hostname)
+	}
+	// GPU models summary
+	if len(result.GPUs) > 0 {
+		modelCount := make(map[string]int)
+		var modelOrder []string
+		for _, g := range result.GPUs {
+			m := strings.TrimSpace(g.Name)
+			if m == "" {
+				m = "Unknown GPU"
+			}
+			if modelCount[m] == 0 {
+				modelOrder = append(modelOrder, m)
+			}
+			modelCount[m]++
+		}
+		var parts []string
+		for _, m := range modelOrder {
+			if modelCount[m] == 1 {
+				parts = append(parts, m)
+			} else {
+				parts = append(parts, fmt.Sprintf("%d× %s", modelCount[m], m))
+			}
+		}
+		fmt.Fprintf(&b, "**GPU(s):** %s  \n", strings.Join(parts, ", "))
+	}
+	fmt.Fprintf(&b, "**Profile:** %s  \n", result.BenchmarkProfile)
+	fmt.Fprintf(&b, "**App version:** %s  \n", result.BenchmarkVersion)
+	fmt.Fprintf(&b, "**Generated:** %s  \n", result.GeneratedAt.Format("2006-01-02 15:04:05 UTC"))
+	if result.ParallelGPUs {
+		fmt.Fprintf(&b, "**Mode:** parallel (all GPUs simultaneously)  \n")
+	}
+	fmt.Fprintf(&b, "**Overall status:** %s  \n", result.OverallStatus)
+	b.WriteString("\n")
+
+	// ── Executive Summary ─────────────────────────────────────────────────────
+	if len(result.Findings) > 0 {
+		b.WriteString("## Executive Summary\n\n")
+		for _, finding := range result.Findings {
+			fmt.Fprintf(&b, "- %s\n", finding)
+		}
+		b.WriteString("\n")
+	}
+
+	if len(result.Warnings) > 0 {
+		b.WriteString("## Warnings\n\n")
+		for _, warning := range result.Warnings {
+			fmt.Fprintf(&b, "- %s\n", warning)
+		}
+		b.WriteString("\n")
+	}
+
+	// ── Scorecard table ───────────────────────────────────────────────────────
+	b.WriteString("## Scorecard\n\n")
+	b.WriteString("| GPU | Status | Composite | Compute | TOPS/SM/GHz | Power Sustain | Thermal Sustain | Stability | Interconnect |\n")
+	b.WriteString("|-----|--------|-----------|---------|-------------|---------------|-----------------|-----------|-------------|\n")
+	for _, gpu := range result.GPUs {
+		name := strings.TrimSpace(gpu.Name)
+		if name == "" {
+			name = "Unknown"
+		}
+		interconnect := "-"
+		if gpu.Scores.InterconnectScore > 0 {
+			interconnect = fmt.Sprintf("%.1f", gpu.Scores.InterconnectScore)
+		}
+		topsPerSM := "-"
+		if gpu.Scores.TOPSPerSMPerGHz > 0 {
+			topsPerSM = fmt.Sprintf("%.3f", gpu.Scores.TOPSPerSMPerGHz)
+		}
+		fmt.Fprintf(&b, "| GPU %d %s | %s | **%.2f** | %.2f | %s | %.1f | %.1f | %.1f | %s |\n",
+			gpu.Index, name,
+			gpu.Status,
+			gpu.Scores.CompositeScore,
+			gpu.Scores.ComputeScore,
+			topsPerSM,
+			gpu.Scores.PowerSustainScore,
+			gpu.Scores.ThermalSustainScore,
+			gpu.Scores.StabilityScore,
+			interconnect,
+		)
+	}
+	b.WriteString("\n")
+
+	// ── Per GPU detail ────────────────────────────────────────────────────────
+	b.WriteString("## Per-GPU Details\n\n")
+	for _, gpu := range result.GPUs {
+		name := strings.TrimSpace(gpu.Name)
+		if name == "" {
+			name = "Unknown GPU"
+		}
+		fmt.Fprintf(&b, "### GPU %d — %s\n\n", gpu.Index, name)
+
+		// Identity
+		if gpu.BusID != "" {
+			fmt.Fprintf(&b, "- **Bus ID:** %s\n", gpu.BusID)
+		}
+		if gpu.VBIOS != "" {
+			fmt.Fprintf(&b, "- **vBIOS:** %s\n", gpu.VBIOS)
+		}
+		if gpu.ComputeCapability != "" {
+			fmt.Fprintf(&b, "- **Compute capability:** %s\n", gpu.ComputeCapability)
+		}
+		if gpu.MultiprocessorCount > 0 {
+			fmt.Fprintf(&b, "- **SMs:** %d\n", gpu.MultiprocessorCount)
+		}
+		if gpu.PowerLimitW > 0 {
+			fmt.Fprintf(&b, "- **Power limit:** %.0f W (default %.0f W)\n", gpu.PowerLimitW, gpu.DefaultPowerLimitW)
+		}
+		if gpu.LockedGraphicsClockMHz > 0 {
+			fmt.Fprintf(&b, "- **Locked clocks:** GPU %.0f MHz / Mem %.0f MHz\n", gpu.LockedGraphicsClockMHz, gpu.LockedMemoryClockMHz)
+		}
+		b.WriteString("\n")
+
+		// Steady-state telemetry
+		fmt.Fprintf(&b, "**Steady-state telemetry** (%ds):\n\n", int(gpu.Steady.DurationSec))
+		b.WriteString("| | Avg | P95 |\n|---|---|---|\n")
+		fmt.Fprintf(&b, "| Power | %.1f W | %.1f W |\n", gpu.Steady.AvgPowerW, gpu.Steady.P95PowerW)
+		fmt.Fprintf(&b, "| Temperature | %.1f °C | %.1f °C |\n", gpu.Steady.AvgTempC, gpu.Steady.P95TempC)
+		fmt.Fprintf(&b, "| GPU clock | %.0f MHz | %.0f MHz |\n", gpu.Steady.AvgGraphicsClockMHz, gpu.Steady.P95GraphicsClockMHz)
+		fmt.Fprintf(&b, "| Memory clock | %.0f MHz | %.0f MHz |\n", gpu.Steady.AvgMemoryClockMHz, gpu.Steady.P95MemoryClockMHz)
+		fmt.Fprintf(&b, "| GPU utilisation | %.1f %% | — |\n", gpu.Steady.AvgUsagePct)
+		b.WriteString("\n")
+
+		// Throttle
+		throttle := formatThrottleLine(gpu.Throttle, gpu.Steady.DurationSec)
+		if throttle != "none" {
+			fmt.Fprintf(&b, "**Throttle:** %s\n\n", throttle)
+		}
+
+		// Precision results
+		if len(gpu.PrecisionResults) > 0 {
+			b.WriteString("**Precision results:**\n\n")
+			b.WriteString("| Precision | TOPS | Lanes | Iterations |\n|-----------|------|-------|------------|\n")
+			for _, p := range gpu.PrecisionResults {
+				if p.Supported {
+					fmt.Fprintf(&b, "| %s | %.2f | %d | %d |\n", p.Name, p.TeraOpsPerSec, p.Lanes, p.Iterations)
+				} else {
+					fmt.Fprintf(&b, "| %s | — (unsupported) | — | — |\n", p.Name)
+				}
+			}
+			b.WriteString("\n")
+		}
+
+		// Degradation / Notes
+		if len(gpu.DegradationReasons) > 0 {
+			fmt.Fprintf(&b, "**Degradation reasons:** %s\n\n", strings.Join(gpu.DegradationReasons, ", "))
+		}
+		if len(gpu.Notes) > 0 {
+			b.WriteString("**Notes:**\n\n")
+			for _, note := range gpu.Notes {
+				fmt.Fprintf(&b, "- %s\n", note)
+			}
+			b.WriteString("\n")
+		}
+	}
+
+	// ── Interconnect ──────────────────────────────────────────────────────────
+	if result.Interconnect != nil {
+		b.WriteString("## Interconnect (NCCL)\n\n")
+		fmt.Fprintf(&b, "**Status:** %s\n\n", result.Interconnect.Status)
+		if result.Interconnect.Supported {
+			b.WriteString("| Metric | Avg | Max |\n|--------|-----|-----|\n")
+			fmt.Fprintf(&b, "| Alg BW | %.1f GB/s | %.1f GB/s |\n", result.Interconnect.AvgAlgBWGBps, result.Interconnect.MaxAlgBWGBps)
+			fmt.Fprintf(&b, "| Bus BW | %.1f GB/s | %.1f GB/s |\n", result.Interconnect.AvgBusBWGBps, result.Interconnect.MaxBusBWGBps)
+			b.WriteString("\n")
+		}
+		for _, note := range result.Interconnect.Notes {
+			fmt.Fprintf(&b, "- %s\n", note)
+		}
+		if len(result.Interconnect.Notes) > 0 {
+			b.WriteString("\n")
+		}
+	}
+
+	// ── Server Power (IPMI) ───────────────────────────────────────────────────
+	if sp := result.ServerPower; sp != nil {
+		b.WriteString("## Server Power (IPMI)\n\n")
+		if !sp.Available {
+			b.WriteString("IPMI power measurement unavailable.\n\n")
+		} else {
+			b.WriteString("| | Value |\n|---|---|\n")
+			fmt.Fprintf(&b, "| Server idle | %.0f W |\n", sp.IdleW)
+			fmt.Fprintf(&b, "| Server under load | %.0f W |\n", sp.LoadedW)
+			fmt.Fprintf(&b, "| Server delta (load − idle) | %.0f W |\n", sp.DeltaW)
+			fmt.Fprintf(&b, "| GPU-reported sum | %.0f W |\n", sp.GPUReportedSumW)
+			if sp.ReportingRatio > 0 {
+				fmt.Fprintf(&b, "| Reporting ratio | %.2f (1.0 = accurate, <0.75 = GPU over-reports) |\n", sp.ReportingRatio)
+			}
+			b.WriteString("\n")
+		}
+		for _, note := range sp.Notes {
+			fmt.Fprintf(&b, "- %s\n", note)
+		}
+		if len(sp.Notes) > 0 {
+			b.WriteString("\n")
+		}
+	}
+
+	// ── Terminal charts (steady-state only) ───────────────────────────────────
+	if len(charts) > 0 {
+		b.WriteString("## Steady-State Charts\n\n")
+		for _, chart := range charts {
+			content := strings.TrimSpace(stripANSIEscapeSequences(chart.Content))
+			if content == "" {
+				continue
+			}
+			fmt.Fprintf(&b, "### %s\n\n```\n%s\n```\n\n", chart.Title, content)
+		}
+	}
+
+	// ── Methodology ───────────────────────────────────────────────────────────
+	b.WriteString("## Methodology\n\n")
+	fmt.Fprintf(&b, "- Profile `%s` uses standardized baseline → warmup → steady-state → interconnect → cooldown phases.\n", result.BenchmarkProfile)
+	b.WriteString("- Single-GPU compute score from bee-gpu-burn cuBLASLt when available.\n")
+	b.WriteString("- Thermal and power limitations inferred from NVIDIA clock event reason counters and sustained telemetry.\n")
+	b.WriteString("- `result.json` is the canonical machine-readable source for this benchmark run.\n\n")
+
+	// ── Raw files ─────────────────────────────────────────────────────────────
+	b.WriteString("## Raw Files\n\n")
+	b.WriteString("- `result.json`\n- `report.md`\n- `summary.txt`\n- `verbose.log`\n")
+	b.WriteString("- `gpu-*-baseline-metrics.csv/html/term.txt`\n")
+	b.WriteString("- `gpu-*-warmup.log`\n")
+	b.WriteString("- `gpu-*-steady.log`\n")
+	b.WriteString("- `gpu-*-steady-metrics.csv/html/term.txt`\n")
+	b.WriteString("- `gpu-*-cooldown-metrics.csv/html/term.txt`\n")
+	if result.Interconnect != nil {
+		b.WriteString("- `nccl-all-reduce.log`\n")
+	}
+	return b.String()
+}
+
+// loadBenchmarkReportCharts loads only steady-state terminal charts (baseline and
+// cooldown charts are not useful for human review).
+func loadBenchmarkReportCharts(runDir string, gpuIndices []int) []benchmarkReportChart {
+	var charts []benchmarkReportChart
+	for _, idx := range gpuIndices {
+		path := filepath.Join(runDir, fmt.Sprintf("gpu-%d-steady-metrics-term.txt", idx))
+		raw, err := os.ReadFile(path)
+		if err != nil || len(raw) == 0 {
+			continue
+		}
+		charts = append(charts, benchmarkReportChart{
+			Title:   fmt.Sprintf("GPU %d — Steady State", idx),
+			Content: string(raw),
+		})
+	}
+	return charts
+}
+
+func stripANSIEscapeSequences(raw string) string {
+	return ansiEscapePattern.ReplaceAllString(raw, "")
+}
+
+// formatThrottleLine renders throttle counters as human-readable percentages of
+// the steady-state window.  Only non-zero counters are shown.  When the steady
+// duration is unknown (0), raw seconds are shown instead.
+func formatThrottleLine(t BenchmarkThrottleCounters, steadyDurationSec float64) string {
+	type counter struct {
+		label string
+		us    uint64
+	}
+	counters := []counter{
+		{"sw_power", t.SWPowerCapUS},
+		{"sw_thermal", t.SWThermalSlowdownUS},
+		{"sync_boost", t.SyncBoostUS},
+		{"hw_thermal", t.HWThermalSlowdownUS},
+		{"hw_power_brake", t.HWPowerBrakeSlowdownUS},
+	}
+	var parts []string
+	for _, c := range counters {
+		if c.us == 0 {
+			continue
+		}
+		sec := float64(c.us) / 1e6
+		if steadyDurationSec > 0 {
+			pct := sec / steadyDurationSec * 100
+			parts = append(parts, fmt.Sprintf("%s=%.1f%% (%.0fs)", c.label, pct, sec))
+		} else if sec < 1 {
+			parts = append(parts, fmt.Sprintf("%s=%.0fms", c.label, sec*1000))
+		} else {
+			parts = append(parts, fmt.Sprintf("%s=%.1fs", c.label, sec))
+		}
+	}
+	if len(parts) == 0 {
+		return "none"
+	}
+	return strings.Join(parts, "  ")
+}
+
+func renderBenchmarkSummary(result NvidiaBenchmarkResult) string {
+	var b strings.Builder
+	fmt.Fprintf(&b, "run_at_utc=%s\n", result.GeneratedAt.Format(time.RFC3339))
+	fmt.Fprintf(&b, "benchmark_profile=%s\n", result.BenchmarkProfile)
+	fmt.Fprintf(&b, "overall_status=%s\n", result.OverallStatus)
+	fmt.Fprintf(&b, "gpu_count=%d\n", len(result.GPUs))
+	fmt.Fprintf(&b, "normalization_status=%s\n", result.Normalization.Status)
+	var best float64
+	for i, gpu := range result.GPUs {
+		fmt.Fprintf(&b, "gpu_%d_status=%s\n", gpu.Index, gpu.Status)
+		fmt.Fprintf(&b, "gpu_%d_composite_score=%.2f\n", gpu.Index, gpu.Scores.CompositeScore)
+		if i == 0 || gpu.Scores.CompositeScore > best {
+			best = gpu.Scores.CompositeScore
+		}
+	}
+	fmt.Fprintf(&b, "best_composite_score=%.2f\n", best)
+	if result.Interconnect != nil {
+		fmt.Fprintf(&b, "interconnect_status=%s\n", result.Interconnect.Status)
+		fmt.Fprintf(&b, "interconnect_max_busbw_gbps=%.1f\n", result.Interconnect.MaxBusBWGBps)
+	}
+	return b.String()
+}

audit/internal/platform/benchmark_test.go

@@ -0,0 +1,180 @@
+package platform
+
+import (
+	"strings"
+	"testing"
+)
+
+func TestResolveBenchmarkProfile(t *testing.T) {
+	t.Parallel()
+
+	cases := []struct {
+		name    string
+		profile string
+		want    benchmarkProfileSpec
+	}{
+		{
+			name:    "default",
+			profile: "",
+			want:    benchmarkProfileSpec{Name: NvidiaBenchmarkProfileStandard, BaselineSec: 15, WarmupSec: 120, SteadySec: 480, NCCLSec: 180, CooldownSec: 120},
+		},
+		{
+			name:    "stability",
+			profile: "stability",
+			want:    benchmarkProfileSpec{Name: NvidiaBenchmarkProfileStability, BaselineSec: 30, WarmupSec: 300, SteadySec: 3600, NCCLSec: 300, CooldownSec: 300},
+		},
+		{
+			name:    "overnight",
+			profile: "overnight",
+			want:    benchmarkProfileSpec{Name: NvidiaBenchmarkProfileOvernight, BaselineSec: 60, WarmupSec: 600, SteadySec: 27000, NCCLSec: 600, CooldownSec: 300},
+		},
+	}
+
+	for _, tc := range cases {
+		tc := tc
+		t.Run(tc.name, func(t *testing.T) {
+			got := resolveBenchmarkProfile(tc.profile)
+			if got != tc.want {
+				t.Fatalf("profile=%q got %+v want %+v", tc.profile, got, tc.want)
+			}
+		})
+	}
+}
+
+func TestNormalizeNvidiaBenchmarkOptionsPreservesRunNCCLChoice(t *testing.T) {
+	t.Parallel()
+
+	opts := normalizeNvidiaBenchmarkOptionsForBenchmark(NvidiaBenchmarkOptions{
+		Profile: "stability",
+		RunNCCL: false,
+	})
+	if opts.Profile != NvidiaBenchmarkProfileStability {
+		t.Fatalf("profile=%q want %q", opts.Profile, NvidiaBenchmarkProfileStability)
+	}
+	if opts.RunNCCL {
+		t.Fatalf("RunNCCL should stay false when explicitly disabled")
+	}
+}
+
+func TestParseBenchmarkBurnLog(t *testing.T) {
+	t.Parallel()
+
+	raw := strings.Join([]string{
+		"loader=bee-gpu-burn",
+		"[gpu 0] device=NVIDIA H100",
+		"[gpu 0] compute_capability=9.0",
+		"[gpu 0] backend=cublasLt",
+		"[gpu 0] duration_s=10",
+		"[gpu 0] fp16_tensor[0]=READY dim=4096x4096x4096 block=128 stream=0",
+		"[gpu 0] fp8_e4m3[0]=READY dim=8192x8192x4096 block=128 stream=0",
+		"[gpu 0] fp16_tensor_iterations=200",
+		"[gpu 0] fp8_e4m3_iterations=50",
+		"[gpu 0] status=OK",
+	}, "\n")
+
+	got := parseBenchmarkBurnLog(raw)
+	if got.Backend != "cublasLt" {
+		t.Fatalf("backend=%q want cublasLt", got.Backend)
+	}
+	if got.ComputeCapability != "9.0" {
+		t.Fatalf("compute capability=%q want 9.0", got.ComputeCapability)
+	}
+	if len(got.Profiles) != 2 {
+		t.Fatalf("profiles=%d want 2", len(got.Profiles))
+	}
+	if got.Profiles[0].TeraOpsPerSec <= 0 {
+		t.Fatalf("profile[0] teraops=%f want >0", got.Profiles[0].TeraOpsPerSec)
+	}
+	if got.Profiles[1].Category != "fp8" {
+		t.Fatalf("profile[1] category=%q want fp8", got.Profiles[1].Category)
+	}
+}
+
+func TestRenderBenchmarkReportIncludesFindingsAndScores(t *testing.T) {
+	t.Parallel()
+
+	result := NvidiaBenchmarkResult{
+		BenchmarkVersion:   benchmarkVersion,
+		BenchmarkProfile:   NvidiaBenchmarkProfileStandard,
+		OverallStatus:      "PARTIAL",
+		SelectedGPUIndices: []int{0},
+		Normalization: BenchmarkNormalization{
+			Status: "partial",
+		},
+		Findings: []string{"GPU 0 spent measurable time under SW power cap."},
+		GPUs: []BenchmarkGPUResult{
+			{
+				Index:  0,
+				Name:   "NVIDIA H100",
+				Status: "OK",
+				Steady: BenchmarkTelemetrySummary{
+					AvgPowerW:           680,
+					AvgTempC:            79,
+					AvgGraphicsClockMHz: 1725,
+					P95PowerW:           700,
+					P95TempC:            82,
+					P95GraphicsClockMHz: 1800,
+				},
+				Scores: BenchmarkScorecard{
+					ComputeScore:        1200,
+					PowerSustainScore:   96,
+					ThermalSustainScore: 88,
+					StabilityScore:      92,
+					CompositeScore:      1176,
+				},
+				PrecisionResults: []BenchmarkPrecisionResult{
+					{Name: "fp16_tensor", Supported: true, TeraOpsPerSec: 700},
+				},
+				Throttle: BenchmarkThrottleCounters{
+					SWPowerCapUS: 1000000,
+				},
+				DegradationReasons: []string{"power_capped"},
+			},
+		},
+	}
+
+	report := renderBenchmarkReport(result)
+	for _, needle := range []string{
+		"Executive Summary",
+		"GPU 0 spent measurable time under SW power cap.",
+		"1176.00",
+		"fp16_tensor",
+		"700.00",
+	} {
+		if !strings.Contains(report, needle) {
+			t.Fatalf("report missing %q\n%s", needle, report)
+		}
+	}
+}
+
+func TestRenderBenchmarkReportIncludesTerminalChartsWithoutANSI(t *testing.T) {
+	t.Parallel()
+
+	report := renderBenchmarkReportWithCharts(NvidiaBenchmarkResult{
+		BenchmarkProfile:   NvidiaBenchmarkProfileStandard,
+		OverallStatus:      "OK",
+		SelectedGPUIndices: []int{0},
+		Normalization: BenchmarkNormalization{
+			Status: "full",
+		},
+	}, []benchmarkReportChart{
+		{
+			Title:   "GPU 0 Steady State",
+			Content: "\x1b[31mGPU 0 chart\x1b[0m\n 42┤───",
+		},
+	})
+
+	for _, needle := range []string{
+		"Steady-State Charts",
+		"GPU 0 Steady State",
+		"GPU 0 chart",
+		"42┤───",
+	} {
+		if !strings.Contains(report, needle) {
+			t.Fatalf("report missing %q\n%s", needle, report)
+		}
+	}
+	if strings.Contains(report, "\x1b[31m") {
+		t.Fatalf("report should not contain ANSI escapes\n%s", report)
+	}
+}

audit/internal/platform/benchmark_types.go

@@ -0,0 +1,158 @@
+package platform
+
+import "time"
+
+const (
+	NvidiaBenchmarkProfileStandard  = "standard"
+	NvidiaBenchmarkProfileStability = "stability"
+	NvidiaBenchmarkProfileOvernight = "overnight"
+)
+
+type NvidiaBenchmarkOptions struct {
+	Profile           string
+	SizeMB            int
+	GPUIndices        []int
+	ExcludeGPUIndices []int
+	RunNCCL           bool
+	ParallelGPUs      bool // run all selected GPUs simultaneously instead of sequentially
+}
+
+
+type NvidiaBenchmarkResult struct {
+	BenchmarkVersion   string                       `json:"benchmark_version"`
+	GeneratedAt        time.Time                    `json:"generated_at"`
+	Hostname           string                       `json:"hostname,omitempty"`
+	ServerModel        string                       `json:"server_model,omitempty"`
+	BenchmarkProfile   string                       `json:"benchmark_profile"`
+	ParallelGPUs       bool                         `json:"parallel_gpus,omitempty"`
+	OverallStatus      string                       `json:"overall_status"`
+	SelectedGPUIndices []int                        `json:"selected_gpu_indices"`
+	Findings           []string                     `json:"findings,omitempty"`
+	Warnings           []string                     `json:"warnings,omitempty"`
+	Normalization      BenchmarkNormalization       `json:"normalization"`
+	GPUs               []BenchmarkGPUResult         `json:"gpus"`
+	Interconnect       *BenchmarkInterconnectResult `json:"interconnect,omitempty"`
+	ServerPower        *BenchmarkServerPower        `json:"server_power,omitempty"`
+}
+
+type BenchmarkNormalization struct {
+	Status string                      `json:"status"`
+	Notes  []string                    `json:"notes,omitempty"`
+	GPUs   []BenchmarkNormalizationGPU `json:"gpus,omitempty"`
+}
+
+type BenchmarkNormalizationGPU struct {
+	Index                 int      `json:"index"`
+	PersistenceMode       string   `json:"persistence_mode,omitempty"`
+	GPUClockLockMHz       float64  `json:"gpu_clock_lock_mhz,omitempty"`
+	GPUClockLockStatus    string   `json:"gpu_clock_lock_status,omitempty"`
+	MemoryClockLockMHz    float64  `json:"memory_clock_lock_mhz,omitempty"`
+	MemoryClockLockStatus string   `json:"memory_clock_lock_status,omitempty"`
+	Notes                 []string `json:"notes,omitempty"`
+}
+
+type BenchmarkGPUResult struct {
+	Index                  int                        `json:"index"`
+	UUID                   string                     `json:"uuid,omitempty"`
+	Name                   string                     `json:"name,omitempty"`
+	BusID                  string                     `json:"bus_id,omitempty"`
+	VBIOS                  string                     `json:"vbios,omitempty"`
+	ComputeCapability      string                     `json:"compute_capability,omitempty"`
+	Backend                string                     `json:"backend,omitempty"`
+	Status                 string                     `json:"status"`
+	PowerLimitW            float64                    `json:"power_limit_w,omitempty"`
+	MultiprocessorCount    int                        `json:"multiprocessor_count,omitempty"`
+	DefaultPowerLimitW     float64                    `json:"default_power_limit_w,omitempty"`
+	MaxGraphicsClockMHz    float64                    `json:"max_graphics_clock_mhz,omitempty"`
+	BaseGraphicsClockMHz   float64                    `json:"base_graphics_clock_mhz,omitempty"`
+	MaxMemoryClockMHz      float64                    `json:"max_memory_clock_mhz,omitempty"`
+	LockedGraphicsClockMHz float64                    `json:"locked_graphics_clock_mhz,omitempty"`
+	LockedMemoryClockMHz   float64                    `json:"locked_memory_clock_mhz,omitempty"`
+	Baseline               BenchmarkTelemetrySummary  `json:"baseline"`
+	Steady                 BenchmarkTelemetrySummary  `json:"steady"`
+	Cooldown               BenchmarkTelemetrySummary  `json:"cooldown"`
+	Throttle               BenchmarkThrottleCounters  `json:"throttle_counters"`
+	PrecisionResults       []BenchmarkPrecisionResult `json:"precision_results,omitempty"`
+	Scores                 BenchmarkScorecard         `json:"scores"`
+	DegradationReasons     []string                   `json:"degradation_reasons,omitempty"`
+	Notes                  []string                   `json:"notes,omitempty"`
+}
+
+type BenchmarkTelemetrySummary struct {
+	DurationSec         float64 `json:"duration_sec"`
+	Samples             int     `json:"samples"`
+	AvgTempC            float64 `json:"avg_temp_c"`
+	P95TempC            float64 `json:"p95_temp_c"`
+	AvgPowerW           float64 `json:"avg_power_w"`
+	P95PowerW           float64 `json:"p95_power_w"`
+	AvgGraphicsClockMHz float64 `json:"avg_graphics_clock_mhz"`
+	P95GraphicsClockMHz float64 `json:"p95_graphics_clock_mhz"`
+	AvgMemoryClockMHz   float64 `json:"avg_memory_clock_mhz"`
+	P95MemoryClockMHz   float64 `json:"p95_memory_clock_mhz"`
+	AvgUsagePct         float64 `json:"avg_usage_pct"`
+	AvgMemUsagePct      float64 `json:"avg_mem_usage_pct"`
+	ClockCVPct          float64 `json:"clock_cv_pct"`
+	PowerCVPct          float64 `json:"power_cv_pct"`
+	TempCVPct           float64 `json:"temp_cv_pct"`
+	ClockDriftPct       float64 `json:"clock_drift_pct"`
+}
+
+type BenchmarkThrottleCounters struct {
+	SWPowerCapUS           uint64 `json:"sw_power_cap_us"`
+	SWThermalSlowdownUS    uint64 `json:"sw_thermal_slowdown_us"`
+	SyncBoostUS            uint64 `json:"sync_boost_us"`
+	HWThermalSlowdownUS    uint64 `json:"hw_thermal_slowdown_us"`
+	HWPowerBrakeSlowdownUS uint64 `json:"hw_power_brake_slowdown_us"`
+}
+
+type BenchmarkPrecisionResult struct {
+	Name          string  `json:"name"`
+	Category      string  `json:"category"`
+	Supported     bool    `json:"supported"`
+	Lanes         int     `json:"lanes,omitempty"`
+	M             uint64  `json:"m,omitempty"`
+	N             uint64  `json:"n,omitempty"`
+	K             uint64  `json:"k,omitempty"`
+	Iterations    uint64  `json:"iterations,omitempty"`
+	TeraOpsPerSec float64 `json:"teraops_per_sec,omitempty"`
+	Notes         string  `json:"notes,omitempty"`
+}
+
+type BenchmarkScorecard struct {
+	ComputeScore        float64 `json:"compute_score"`
+	PowerSustainScore   float64 `json:"power_sustain_score"`
+	ThermalSustainScore float64 `json:"thermal_sustain_score"`
+	StabilityScore      float64 `json:"stability_score"`
+	InterconnectScore   float64 `json:"interconnect_score"`
+	CompositeScore      float64 `json:"composite_score"`
+	// TOPSPerSMPerGHz is compute efficiency independent of clock speed and SM count.
+	// Comparable across throttle levels and GPU generations. Low value at normal
+	// clocks indicates silicon degradation.
+	TOPSPerSMPerGHz float64 `json:"tops_per_sm_per_ghz,omitempty"`
+}
+
+// BenchmarkServerPower captures server-side power via IPMI alongside GPU-reported
+// power. The reporting_ratio (delta / gpu_reported_sum) near 1.0 means GPU power
+// telemetry is accurate; a ratio well below 1.0 (e.g. 0.5) means the GPU is
+// over-reporting its power consumption.
+type BenchmarkServerPower struct {
+	Available       bool     `json:"available"`
+	IdleW           float64  `json:"idle_w,omitempty"`
+	LoadedW         float64  `json:"loaded_w,omitempty"`
+	DeltaW          float64  `json:"delta_w,omitempty"`
+	GPUReportedSumW float64  `json:"gpu_reported_sum_w,omitempty"`
+	ReportingRatio  float64  `json:"reporting_ratio,omitempty"`
+	Notes           []string `json:"notes,omitempty"`
+}
+
+type BenchmarkInterconnectResult struct {
+	Status             string   `json:"status"`
+	Attempted          bool     `json:"attempted"`
+	Supported          bool     `json:"supported"`
+	SelectedGPUIndices []int    `json:"selected_gpu_indices,omitempty"`
+	AvgAlgBWGBps       float64  `json:"avg_algbw_gbps,omitempty"`
+	MaxAlgBWGBps       float64  `json:"max_algbw_gbps,omitempty"`
+	AvgBusBWGBps       float64  `json:"avg_busbw_gbps,omitempty"`
+	MaxBusBWGBps       float64  `json:"max_busbw_gbps,omitempty"`
+	Notes              []string `json:"notes,omitempty"`
+}

audit/internal/platform/error_patterns.go

@@ -0,0 +1,139 @@
+package platform
+
+import "regexp"
+
+// ErrorPattern describes a kernel log pattern that indicates a hardware error.
+// Add new patterns by appending to HardwareErrorPatterns — no other code changes needed.
+type ErrorPattern struct {
+	// Name is a short machine-readable label for logging and deduplication.
+	Name string
+	// Re is the compiled regular expression matched against a single kmsg line.
+	Re *regexp.Regexp
+	// Category groups related errors: "gpu", "pcie", "storage", "mce", "memory", "cpu".
+	Category string
+	// Severity is "warning" for recoverable/uncertain faults, "critical" for definitive failures.
+	Severity string
+	// BDFGroup is the capture group index (1-based) that contains a PCIe BDF address
+	// (e.g. "0000:c8:00.0"). 0 means no BDF is captured by this pattern.
+	BDFGroup int
+	// DevGroup is the capture group index (1-based) that contains a device name
+	// (e.g. "sda", "nvme0"). 0 means no device name is captured by this pattern.
+	DevGroup int
+}
+
+// HardwareErrorPatterns is the global list of kernel log patterns that indicate hardware faults.
+// To add a new pattern: append a new ErrorPattern struct to this slice.
+var HardwareErrorPatterns = []ErrorPattern{
+	// ── GPU / NVIDIA ────────────────────────────────────────────────────────────
+	{
+		Name:     "nvidia-rminitadapter",
+		Re:       mustPat(`(?i)NVRM:.*GPU\s+([\da-f]{4}:[\da-f]{2}:[\da-f]{2}\.\d)`),
+		Category: "gpu",
+		Severity: "warning",
+		BDFGroup: 1,
+	},
+	{
+		Name:     "nvidia-msi-fail",
+		Re:       mustPat(`(?i)NVRM:.*Failed to enable MSI`),
+		Category: "gpu",
+		Severity: "warning",
+	},
+	{
+		Name:     "nvidia-aer",
+		Re:       mustPat(`(?i)nvidia\s+([\da-f]{4}:[\da-f]{2}:[\da-f]{2}\.\d).*AER`),
+		Category: "gpu",
+		Severity: "warning",
+		BDFGroup: 1,
+	},
+	{
+		Name:     "nvidia-xid",
+		Re:       mustPat(`(?i)NVRM:.*Xid.*\b([\da-f]{4}:[\da-f]{2}:[\da-f]{2}\.\d)`),
+		Category: "gpu",
+		Severity: "warning",
+		BDFGroup: 1,
+	},
+
+	// ── PCIe AER (generic) ──────────────────────────────────────────────────────
+	{
+		Name:     "pcie-aer",
+		Re:       mustPat(`(?i)pcieport\s+([\da-f]{4}:[\da-f]{2}:[\da-f]{2}\.\d).*AER`),
+		Category: "pcie",
+		Severity: "warning",
+		BDFGroup: 1,
+	},
+	{
+		Name:     "pcie-uncorrectable",
+		Re:       mustPat(`(?i)([\da-f]{4}:[\da-f]{2}:[\da-f]{2}\.\d).*[Uu]ncorrectable`),
+		Category: "pcie",
+		Severity: "warning",
+		BDFGroup: 1,
+	},
+	{
+		Name:     "pcie-link-down",
+		Re:       mustPat(`(?i)pcieport\s+([\da-f]{4}:[\da-f]{2}:[\da-f]{2}\.\d).*[Ll]ink.*[Dd]own`),
+		Category: "pcie",
+		Severity: "warning",
+		BDFGroup: 1,
+	},
+
+	// ── Storage ─────────────────────────────────────────────────────────────────
+	{
+		Name:     "blk-io-error",
+		Re:       mustPat(`(?i)blk_update_request.*I/O error.*dev\s+(\w+)`),
+		Category: "storage",
+		Severity: "warning",
+		DevGroup: 1,
+	},
+	{
+		Name:     "nvme-timeout",
+		Re:       mustPat(`(?i)nvme\s+(\w+):.*timeout`),
+		Category: "storage",
+		Severity: "warning",
+		DevGroup: 1,
+	},
+	{
+		Name:     "scsi-failed",
+		Re:       mustPat(`(?i)sd\s+[\da-f:]+:.*FAILED`),
+		Category: "storage",
+		Severity: "warning",
+	},
+	{
+		Name:     "nvme-reset",
+		Re:       mustPat(`(?i)nvme\s+(\w+):.*reset`),
+		Category: "storage",
+		Severity: "warning",
+		DevGroup: 1,
+	},
+
+	// ── Machine Check Exceptions ────────────────────────────────────────────────
+	{
+		Name:     "mce-hardware-error",
+		Re:       mustPat(`(?i)mce:.*[Hh]ardware [Ee]rror`),
+		Category: "mce",
+		Severity: "warning",
+	},
+	{
+		Name:     "mce-corrected",
+		Re:       mustPat(`(?i)mce:.*[Cc]orrected`),
+		Category: "mce",
+		Severity: "warning",
+	},
+
+	// ── Memory ─────────────────────────────────────────────────────────────────
+	{
+		Name:     "edac-ue",
+		Re:       mustPat(`(?i)EDAC.*[Uu]ncorrectable`),
+		Category: "memory",
+		Severity: "warning",
+	},
+	{
+		Name:     "edac-ce",
+		Re:       mustPat(`(?i)EDAC.*[Cc]orrectable`),
+		Category: "memory",
+		Severity: "warning",
+	},
+}
+
+func mustPat(s string) *regexp.Regexp {
+	return regexp.MustCompile(s)
+}

audit/internal/platform/gpu_metrics.go

@@ -20,12 +20,13 @@ type GPUMetricRow struct {
 	MemUsagePct float64 `json:"mem_usage_pct"`
 	PowerW      float64 `json:"power_w"`
 	ClockMHz    float64 `json:"clock_mhz"`
+	MemClockMHz float64 `json:"mem_clock_mhz"`
 }
 
 // sampleGPUMetrics runs nvidia-smi once and returns current metrics for each GPU.
 func sampleGPUMetrics(gpuIndices []int) ([]GPUMetricRow, error) {
 	args := []string{
-		"--query-gpu=index,temperature.gpu,utilization.gpu,utilization.memory,power.draw,clocks.current.graphics",
+		"--query-gpu=index,temperature.gpu,utilization.gpu,utilization.memory,power.draw,clocks.current.graphics,clocks.current.memory",
 		"--format=csv,noheader,nounits",
 	}
 	if len(gpuIndices) > 0 {
@@ -46,7 +47,7 @@ func sampleGPUMetrics(gpuIndices []int) ([]GPUMetricRow, error) {
 			continue
 		}
 		parts := strings.Split(line, ", ")
-		if len(parts) < 6 {
+		if len(parts) < 7 {
 			continue
 		}
 		idx, _ := strconv.Atoi(strings.TrimSpace(parts[0]))
@@ -57,6 +58,7 @@ func sampleGPUMetrics(gpuIndices []int) ([]GPUMetricRow, error) {
 			MemUsagePct: parseGPUFloat(parts[3]),
 			PowerW:      parseGPUFloat(parts[4]),
 			ClockMHz:    parseGPUFloat(parts[5]),
+			MemClockMHz: parseGPUFloat(parts[6]),
 		})
 	}
 	return rows, nil
@@ -139,10 +141,10 @@ func sampleAMDGPUMetrics() ([]GPUMetricRow, error) {
 // WriteGPUMetricsCSV writes collected rows as a CSV file.
 func WriteGPUMetricsCSV(path string, rows []GPUMetricRow) error {
 	var b bytes.Buffer
-	b.WriteString("elapsed_sec,gpu_index,temperature_c,usage_pct,power_w,clock_mhz\n")
+	b.WriteString("elapsed_sec,gpu_index,temperature_c,usage_pct,mem_usage_pct,power_w,clock_mhz,mem_clock_mhz\n")
 	for _, r := range rows {
-		fmt.Fprintf(&b, "%.1f,%d,%.1f,%.1f,%.1f,%.0f\n",
-			r.ElapsedSec, r.GPUIndex, r.TempC, r.UsagePct, r.PowerW, r.ClockMHz)
+		fmt.Fprintf(&b, "%.1f,%d,%.1f,%.1f,%.1f,%.1f,%.0f,%.0f\n",
+			r.ElapsedSec, r.GPUIndex, r.TempC, r.UsagePct, r.MemUsagePct, r.PowerW, r.ClockMHz, r.MemClockMHz)
 	}
 	return os.WriteFile(path, b.Bytes(), 0644)
 }
@@ -197,7 +199,7 @@ func drawGPUChartSVG(rows []GPUMetricRow, gpuIdx int) string {
 	const PW = plotX2 - plotX1
 	const PH = plotY2 - plotY1
 	// Outer axes
-	const tempAxisX = 60  // temp axis line
+	const tempAxisX = 60   // temp axis line
 	const clockAxisX = 900 // clock axis line
 
 	colors := [4]string{"#e74c3c", "#3498db", "#2ecc71", "#f39c12"}
@@ -381,10 +383,7 @@ func drawGPUChartSVG(rows []GPUMetricRow, gpuIdx int) string {
 }
 
 const (
-	ansiRed    = "\033[31m"
-	ansiBlue   = "\033[34m"
-	ansiGreen  = "\033[32m"
-	ansiYellow = "\033[33m"
+	ansiAmber  = "\033[38;5;214m"
 	ansiReset  = "\033[0m"
 )
 
@@ -413,10 +412,10 @@ func RenderGPUTerminalChart(rows []GPUMetricRow) string {
 		fn      func(GPUMetricRow) float64
 	}
 	defs := []seriesDef{
-		{"Temperature (°C)", ansiRed, func(r GPUMetricRow) float64 { return r.TempC }},
-		{"GPU Usage (%)", ansiBlue, func(r GPUMetricRow) float64 { return r.UsagePct }},
-		{"Power (W)", ansiGreen, func(r GPUMetricRow) float64 { return r.PowerW }},
-		{"Clock (MHz)", ansiYellow, func(r GPUMetricRow) float64 { return r.ClockMHz }},
+		{"Temperature (°C)", ansiAmber, func(r GPUMetricRow) float64 { return r.TempC }},
+		{"GPU Usage (%)", ansiAmber, func(r GPUMetricRow) float64 { return r.UsagePct }},
+		{"Power (W)", ansiAmber, func(r GPUMetricRow) float64 { return r.PowerW }},
+		{"Clock (MHz)", ansiAmber, func(r GPUMetricRow) float64 { return r.ClockMHz }},
 	}
 
 	var b strings.Builder

audit/internal/platform/install_to_ram.go

@@ -116,14 +116,71 @@ func (s *System) RunInstallToRAM(ctx context.Context, logFunc func(string)) erro
 	if err := ctx.Err(); err != nil {
 		return err
 	}
-	if err := exec.Command("mount", "--bind", dstDir, "/run/live/medium").Run(); err != nil {
-		log(fmt.Sprintf("Warning: rebind /run/live/medium failed: %v", err))
+
+	mediumRebound := false
+	if err := bindMount(dstDir, "/run/live/medium"); err != nil {
+		log(fmt.Sprintf("Warning: rebind /run/live/medium → %s failed: %v", dstDir, err))
+	} else {
+		mediumRebound = true
 	}
 
-	log("Done. Installation media can be safely disconnected.")
+	log("Verifying live medium now served from RAM...")
+	status := s.LiveBootSource()
+	if err := verifyInstallToRAMStatus(status, dstDir, mediumRebound, log); err != nil {
+		return err
+	}
+	if status.InRAM {
+		log(fmt.Sprintf("Verification passed: live medium now served from %s.", describeLiveBootSource(status)))
+	}
+	log("Done. Squashfs files are in RAM. Installation media can be safely disconnected.")
 	return nil
 }
 
+func verifyInstallToRAMStatus(status LiveBootSource, dstDir string, mediumRebound bool, log func(string)) error {
+	if status.InRAM {
+		return nil
+	}
+
+	// The live medium mount was not redirected to RAM. This is expected when
+	// booting from an ISO/CD-ROM: the squashfs loop device has a non-zero
+	// offset and LOOP_CHANGE_FD cannot be used; the bind mount also fails
+	// because the CD-ROM mount is in use. Check whether files were at least
+	// copied to the tmpfs directory — that is sufficient for safe disconnection
+	// once the kernel has paged in all actively-used data.
+	files, _ := filepath.Glob(filepath.Join(dstDir, "*.squashfs"))
+	if len(files) > 0 {
+		if !mediumRebound {
+			log(fmt.Sprintf("Note: squashfs copied to RAM (%s) but /run/live/medium still shows the original source.", dstDir))
+			log("This is normal for CD-ROM boots. For a fully transparent RAM boot, add 'toram' to the kernel parameters.")
+		}
+		return nil
+	}
+
+	return fmt.Errorf("install to RAM verification failed: live medium still mounted from %s and no squashfs found in %s", describeLiveBootSource(status), dstDir)
+}
+
+func describeLiveBootSource(status LiveBootSource) string {
+	source := strings.TrimSpace(status.Device)
+	if source == "" {
+		source = strings.TrimSpace(status.Source)
+	}
+	if source == "" {
+		source = "unknown source"
+	}
+	switch strings.TrimSpace(status.Kind) {
+	case "ram":
+		return "RAM"
+	case "usb":
+		return "USB (" + source + ")"
+	case "cdrom":
+		return "CD-ROM (" + source + ")"
+	case "disk":
+		return "disk (" + source + ")"
+	default:
+		return source
+	}
+}
+
 func copyFileLarge(ctx context.Context, src, dst string, logFunc func(string)) error {
 	in, err := os.Open(src)
 	if err != nil {
@@ -212,7 +269,31 @@ func findLoopForFile(backingFile string) (string, error) {
 	return "", fmt.Errorf("no loop device found for %s", backingFile)
 }
 
+// loopDeviceOffset returns the byte offset configured for the loop device,
+// or -1 if it cannot be determined.
+func loopDeviceOffset(loopDev string) int64 {
+	out, err := exec.Command("losetup", "--json", loopDev).Output()
+	if err != nil {
+		return -1
+	}
+	var result struct {
+		Loopdevices []struct {
+			Offset int64 `json:"offset"`
+		} `json:"loopdevices"`
+	}
+	if err := json.Unmarshal(out, &result); err != nil || len(result.Loopdevices) == 0 {
+		return -1
+	}
+	return result.Loopdevices[0].Offset
+}
+
 func reassociateLoopDevice(loopDev, newFile string) error {
+	// LOOP_CHANGE_FD requires lo_offset == 0. ISO/CD-ROM loop devices are
+	// typically set up with a non-zero offset (squashfs lives inside the ISO),
+	// so the ioctl returns EINVAL. Detect this early for a clear error message.
+	if off := loopDeviceOffset(loopDev); off > 0 {
+		return fmt.Errorf("loop device has non-zero offset (%d bytes, typical for ISO/CD-ROM) — LOOP_CHANGE_FD not supported; use 'toram' kernel parameter for RAM boot", off)
+	}
 	if err := exec.Command("losetup", "--replace", loopDev, newFile).Run(); err == nil {
 		return nil
 	}

audit/internal/platform/install_to_ram_linux.go

@@ -26,3 +26,8 @@ func loopChangeFD(loopDev, newFile string) error {
 	}
 	return nil
 }
+
+// bindMount binds src over dst using the syscall directly (avoids exec PATH issues).
+func bindMount(src, dst string) error {
+	return syscall.Mount(src, dst, "", syscall.MS_BIND, "")
+}

audit/internal/platform/install_to_ram_other.go

@@ -7,3 +7,7 @@ import "errors"
 func loopChangeFD(loopDev, newFile string) error {
 	return errors.New("LOOP_CHANGE_FD not available on this platform")
 }
+
+func bindMount(src, dst string) error {
+	return errors.New("bind mount not available on this platform")
+}

audit/internal/platform/install_to_ram_test.go

@@ -3,6 +3,8 @@ package platform
 import "testing"
 
 func TestInferLiveBootKind(t *testing.T) {
+	t.Parallel()
+
 	tests := []struct {
 		name       string
 		fsType     string
@@ -18,6 +20,7 @@ func TestInferLiveBootKind(t *testing.T) {
 		{name: "unknown", source: "overlay", want: "unknown"},
 	}
 	for _, tc := range tests {
+		tc := tc
 		t.Run(tc.name, func(t *testing.T) {
 			got := inferLiveBootKind(tc.fsType, tc.source, tc.deviceType, tc.transport)
 			if got != tc.want {
@@ -26,3 +29,32 @@ func TestInferLiveBootKind(t *testing.T) {
 		})
 	}
 }
+
+func TestVerifyInstallToRAMStatus(t *testing.T) {
+	t.Parallel()
+
+	dstDir := t.TempDir()
+
+	if err := verifyInstallToRAMStatus(LiveBootSource{InRAM: true, Kind: "ram", Source: "tmpfs"}, dstDir, false, nil); err != nil {
+		t.Fatalf("expected success for RAM-backed status, got %v", err)
+	}
+
+	err := verifyInstallToRAMStatus(LiveBootSource{InRAM: false, Kind: "usb", Device: "/dev/sdb1"}, dstDir, false, nil)
+	if err == nil {
+		t.Fatal("expected verification failure when media is still on USB")
+	}
+	if got := err.Error(); got != "install to RAM verification failed: live medium still mounted from USB (/dev/sdb1) and no squashfs found in "+dstDir {
+		t.Fatalf("error=%q", got)
+	}
+}
+
+func TestDescribeLiveBootSource(t *testing.T) {
+	t.Parallel()
+
+	if got := describeLiveBootSource(LiveBootSource{InRAM: true, Kind: "ram"}); got != "RAM" {
+		t.Fatalf("got %q want RAM", got)
+	}
+	if got := describeLiveBootSource(LiveBootSource{Kind: "unknown", Source: "/run/live/medium"}); got != "/run/live/medium" {
+		t.Fatalf("got %q want /run/live/medium", got)
+	}
+}

audit/internal/platform/kill_workers.go

@@ -15,6 +15,10 @@ var workerPatterns = []string{
 	"stress-ng",
 	"stressapptest",
 	"memtester",
+	// DCGM diagnostic workers — nvvs is spawned by dcgmi diag and survives
+	// if dcgmi is killed mid-run, leaving the GPU occupied (DCGM_ST_IN_USE).
+	"nvvs",
+	"dcgmi",
 }
 
 // KilledProcess describes a process that was sent SIGKILL.

audit/internal/platform/nvidia_stress.go

@@ -16,12 +16,12 @@ func (s *System) RunNvidiaStressPack(ctx context.Context, baseDir string, opts N
 		return "", err
 	}
 
-	return runAcceptancePackCtx(ctx, baseDir, nvidiaStressArchivePrefix(opts.Loader), []satJob{
-		{name: "01-nvidia-smi-q.log", cmd: []string{"nvidia-smi", "-q"}},
-		{name: "02-nvidia-smi-list.log", cmd: []string{"nvidia-smi", "-L"}},
+	return runAcceptancePackCtx(ctx, baseDir, nvidiaStressArchivePrefix(opts.Loader), withNvidiaPersistenceMode(
+		satJob{name: "01-nvidia-smi-q.log", cmd: []string{"nvidia-smi", "-q"}},
+		satJob{name: "02-nvidia-smi-list.log", cmd: []string{"nvidia-smi", "-L"}},
 		job,
-		{name: "04-nvidia-smi-after.log", cmd: []string{"nvidia-smi", "--query-gpu=index,name,temperature.gpu,power.draw,utilization.gpu,memory.used,memory.total", "--format=csv,noheader,nounits"}},
-	}, logFunc)
+		satJob{name: "04-nvidia-smi-after.log", cmd: []string{"nvidia-smi", "--query-gpu=index,name,temperature.gpu,power.draw,utilization.gpu,memory.used,memory.total", "--format=csv,noheader,nounits"}},
+	), logFunc)
 }
 
 func nvidiaStressArchivePrefix(loader string) string {
@@ -49,6 +49,9 @@ func buildNvidiaStressJob(opts NvidiaStressOptions) (satJob, error) {
 			"--seconds", strconv.Itoa(opts.DurationSec),
 			"--size-mb", strconv.Itoa(opts.SizeMB),
 		}
+		if opts.StaggerSeconds > 0 && len(selected) > 1 {
+			cmd = append(cmd, "--stagger-seconds", strconv.Itoa(opts.StaggerSeconds))
+		}
 		if len(selected) > 0 {
 			cmd = append(cmd, "--devices", joinIndexList(selected))
 		}
@@ -63,6 +66,9 @@ func buildNvidiaStressJob(opts NvidiaStressOptions) (satJob, error) {
 			"bee-john-gpu-stress",
 			"--seconds", strconv.Itoa(opts.DurationSec),
 		}
+		if opts.StaggerSeconds > 0 && len(selected) > 1 {
+			cmd = append(cmd, "--stagger-seconds", strconv.Itoa(opts.StaggerSeconds))
+		}
 		if len(selected) > 0 {
 			cmd = append(cmd, "--devices", joinIndexList(selected))
 		}

audit/internal/platform/platform_stress.go

@@ -110,7 +110,7 @@ func (s *System) RunPlatformStress(
 			wg.Add(1)
 			go func() {
 				defer wg.Done()
-				gpuCmd := buildGPUStressCmd(loadCtx, vendor)
+				gpuCmd := buildGPUStressCmd(loadCtx, vendor, cycle.LoadSec)
 				if gpuCmd == nil {
 					return
 				}
@@ -392,6 +392,13 @@ func buildCPUStressCmd(ctx context.Context) (*exec.Cmd, error) {
 		cmdArgs = append(cmdArgs, "-M", strconv.Itoa(mb))
 	}
 	cmd := exec.CommandContext(ctx, path, cmdArgs...)
+	cmd.SysProcAttr = &syscall.SysProcAttr{Setpgid: true}
+	cmd.Cancel = func() error {
+		if cmd.Process != nil {
+			_ = syscall.Kill(-cmd.Process.Pid, syscall.SIGKILL)
+		}
+		return nil
+	}
 	cmd.Stdout = nil
 	cmd.Stderr = nil
 	if err := startLowPriorityCmd(cmd, 15); err != nil {
@@ -402,28 +409,28 @@ func buildCPUStressCmd(ctx context.Context) (*exec.Cmd, error) {
 
 // buildGPUStressCmd creates a GPU stress command appropriate for the detected vendor.
 // Returns nil if no GPU stress tool is available (CPU-only cycling still useful).
-func buildGPUStressCmd(ctx context.Context, vendor string) *exec.Cmd {
+func buildGPUStressCmd(ctx context.Context, vendor string, durSec int) *exec.Cmd {
 	switch strings.ToLower(vendor) {
 	case "amd":
-		return buildAMDGPUStressCmd(ctx)
+		return buildAMDGPUStressCmd(ctx, durSec)
 	case "nvidia":
-		return buildNvidiaGPUStressCmd(ctx)
+		return buildNvidiaGPUStressCmd(ctx, durSec)
 	}
 	return nil
 }
 
-func buildAMDGPUStressCmd(ctx context.Context) *exec.Cmd {
+func buildAMDGPUStressCmd(ctx context.Context, durSec int) *exec.Cmd {
 	rvsArgs, err := resolveRVSCommand()
 	if err != nil {
 		return nil
 	}
 	rvsPath := rvsArgs[0]
-	cfg := `actions:
+	cfg := fmt.Sprintf(`actions:
 - name: gst_platform
   device: all
   module: gst
   parallel: true
-  duration: 86400000
+  duration: %d`, durSec*1000) + `
   copy_matrix: false
   target_stress: 90
   matrix_size_a: 8640
@@ -433,13 +440,20 @@ func buildAMDGPUStressCmd(ctx context.Context) *exec.Cmd {
 	cfgFile := "/tmp/bee-platform-gst.conf"
 	_ = os.WriteFile(cfgFile, []byte(cfg), 0644)
 	cmd := exec.CommandContext(ctx, rvsPath, "-c", cfgFile)
+	cmd.SysProcAttr = &syscall.SysProcAttr{Setpgid: true}
+	cmd.Cancel = func() error {
+		if cmd.Process != nil {
+			_ = syscall.Kill(-cmd.Process.Pid, syscall.SIGKILL)
+		}
+		return nil
+	}
 	cmd.Stdout = nil
 	cmd.Stderr = nil
 	_ = startLowPriorityCmd(cmd, 10)
 	return cmd
 }
 
-func buildNvidiaGPUStressCmd(ctx context.Context) *exec.Cmd {
+func buildNvidiaGPUStressCmd(ctx context.Context, durSec int) *exec.Cmd {
 	path, err := satLookPath("bee-gpu-burn")
 	if err != nil {
 		path, err = satLookPath("bee-gpu-stress")
@@ -447,7 +461,17 @@ func buildNvidiaGPUStressCmd(ctx context.Context) *exec.Cmd {
 	if err != nil {
 		return nil
 	}
-	cmd := exec.CommandContext(ctx, path, "--seconds", "86400")
+	// Pass exact duration so bee-gpu-burn exits on its own when the cycle ends.
+	// Process group kill via Setpgid+Cancel is kept as a safety net for cases
+	// where the context is cancelled early (user stop, parent timeout).
+	cmd := exec.CommandContext(ctx, path, "--seconds", strconv.Itoa(durSec))
+	cmd.SysProcAttr = &syscall.SysProcAttr{Setpgid: true}
+	cmd.Cancel = func() error {
+		if cmd.Process != nil {
+			_ = syscall.Kill(-cmd.Process.Pid, syscall.SIGKILL)
+		}
+		return nil
+	}
 	cmd.Stdout = nil
 	cmd.Stderr = nil
 	_ = startLowPriorityCmd(cmd, 10)

audit/internal/platform/runtime.go

@@ -135,12 +135,15 @@ func (s *System) runtimeToolStatuses(vendor string) []ToolStatus {
 	case "nvidia":
 		tools = append(tools, s.CheckTools([]string{
 			"nvidia-smi",
+			"dcgmi",
+			"nv-hostengine",
 			"nvidia-bug-report.sh",
 			"bee-gpu-burn",
 			"bee-john-gpu-stress",
 			"bee-nccl-gpu-stress",
 			"all_reduce_perf",
 		})...)
+		tools = append(tools, resolvedToolStatus("dcgmproftester", dcgmProfTesterCandidates...))
 	case "amd":
 		tool := ToolStatus{Name: "rocm-smi"}
 		if cmd, err := resolveROCmSMICommand(); err == nil && len(cmd) > 0 {
@@ -155,11 +158,37 @@ func (s *System) runtimeToolStatuses(vendor string) []ToolStatus {
 	return tools
 }
 
+func resolvedToolStatus(display string, candidates ...string) ToolStatus {
+	for _, candidate := range candidates {
+		path, err := exec.LookPath(candidate)
+		if err == nil {
+			return ToolStatus{Name: display, Path: path, OK: true}
+		}
+	}
+	return ToolStatus{Name: display}
+}
+
 func (s *System) collectGPURuntimeHealth(vendor string, health *schema.RuntimeHealth) {
 	lsmodText := commandText("lsmod")
 
 	switch vendor {
 	case "nvidia":
+		if raw, err := os.ReadFile("/run/bee-nvidia-mode"); err == nil {
+			health.NvidiaGSPMode = strings.TrimSpace(string(raw))
+			if health.NvidiaGSPMode == "gsp-stuck" {
+				health.Issues = append(health.Issues, schema.RuntimeIssue{
+					Code:        "nvidia_gsp_stuck",
+					Severity:    "critical",
+					Description: "NVIDIA GSP firmware init timed out and the kernel module is stuck. Reboot and select 'GSP=off' in the boot menu.",
+				})
+			} else if health.NvidiaGSPMode == "gsp-off" {
+				health.Issues = append(health.Issues, schema.RuntimeIssue{
+					Code:        "nvidia_gsp_disabled",
+					Severity:    "warning",
+					Description: "NVIDIA GSP firmware disabled (fallback). Power management runs via CPU path — power draw readings may differ from reference hardware.",
+				})
+			}
+		}
 		health.DriverReady = strings.Contains(lsmodText, "nvidia ")
 		if !health.DriverReady {
 			health.Issues = append(health.Issues, schema.RuntimeIssue{

audit/internal/platform/sat.go

@@ -12,19 +12,20 @@ import (
 	"os"
 	"os/exec"
 	"path/filepath"
-	"syscall"
 	"sort"
 	"strconv"
 	"strings"
 	"sync"
+	"syscall"
 	"time"
 )
 
 var (
-	satExecCommand = exec.Command
-	satLookPath    = exec.LookPath
-	satGlob        = filepath.Glob
-	satStat        = os.Stat
+	satExecCommand  = exec.Command
+	satLookPath     = exec.LookPath
+	satGlob         = filepath.Glob
+	satStat         = os.Stat
+	satFreeMemBytes = freeMemBytes
 
 	rocmSMIExecutableGlobs = []string{
 		"/opt/rocm/bin/rocm-smi",
@@ -38,6 +39,12 @@ var (
 		"/opt/rocm/bin/rvs",
 		"/opt/rocm-*/bin/rvs",
 	}
+	dcgmProfTesterCandidates = []string{
+		"dcgmproftester",
+		"dcgmproftester13",
+		"dcgmproftester12",
+		"dcgmproftester11",
+	}
 )
 
 // streamExecOutput runs cmd and streams each output line to logFunc (if non-nil).
@@ -76,15 +83,46 @@ func streamExecOutput(cmd *exec.Cmd, logFunc func(string)) ([]byte, error) {
 
 // NvidiaGPU holds basic GPU info from nvidia-smi.
 type NvidiaGPU struct {
-	Index    int
-	Name     string
-	MemoryMB int
+	Index    int    `json:"index"`
+	Name     string `json:"name"`
+	MemoryMB int    `json:"memory_mb"`
+}
+
+type NvidiaGPUStatus struct {
+	Index        int    `json:"index"`
+	Name         string `json:"name"`
+	BDF          string `json:"bdf,omitempty"`
+	Serial       string `json:"serial,omitempty"`
+	Status       string `json:"status"`
+	RawLine      string `json:"raw_line,omitempty"`
+	NeedsReset   bool   `json:"needs_reset"`
+	ParseFailure bool   `json:"parse_failure,omitempty"`
+}
+
+type nvidiaGPUHealth struct {
+	Index        int
+	Name         string
+	NeedsReset   bool
+	RawLine      string
+	ParseFailure bool
+}
+
+type nvidiaGPUStatusFile struct {
+	Index       int
+	Name        string
+	RunStatus   string
+	Reason      string
+	Health      string
+	HealthRaw   string
+	Observed    bool
+	Selected    bool
+	FailingJob  string
 }
 
 // AMDGPUInfo holds basic info about an AMD GPU from rocm-smi.
 type AMDGPUInfo struct {
-	Index int
-	Name  string
+	Index int    `json:"index"`
+	Name  string `json:"name"`
 }
 
 // DetectGPUVendor returns "nvidia" if /dev/nvidia0 exists, "amd" if /dev/kfd exists, or "" otherwise.
@@ -256,9 +294,78 @@ func (s *System) ListNvidiaGPUs() ([]NvidiaGPU, error) {
 			MemoryMB: memMB,
 		})
 	}
+	sort.Slice(gpus, func(i, j int) bool {
+		return gpus[i].Index < gpus[j].Index
+	})
 	return gpus, nil
 }
 
+func (s *System) ListNvidiaGPUStatuses() ([]NvidiaGPUStatus, error) {
+	out, err := satExecCommand(
+		"nvidia-smi",
+		"--query-gpu=index,name,pci.bus_id,serial,temperature.gpu,power.draw,utilization.gpu,memory.used,memory.total",
+		"--format=csv,noheader,nounits",
+	).Output()
+	if err != nil {
+		return nil, fmt.Errorf("nvidia-smi: %w", err)
+	}
+	var gpus []NvidiaGPUStatus
+	for _, line := range strings.Split(strings.TrimSpace(string(out)), "\n") {
+		line = strings.TrimSpace(line)
+		if line == "" {
+			continue
+		}
+		parts := strings.Split(line, ",")
+		if len(parts) < 4 {
+			gpus = append(gpus, NvidiaGPUStatus{RawLine: line, Status: "UNKNOWN", ParseFailure: true})
+			continue
+		}
+		idx, err := strconv.Atoi(strings.TrimSpace(parts[0]))
+		if err != nil {
+			gpus = append(gpus, NvidiaGPUStatus{RawLine: line, Status: "UNKNOWN", ParseFailure: true})
+			continue
+		}
+		upper := strings.ToUpper(line)
+		needsReset := strings.Contains(upper, "GPU REQUIRES RESET")
+		status := "OK"
+		if needsReset {
+			status = "RESET_REQUIRED"
+		}
+		gpus = append(gpus, NvidiaGPUStatus{
+			Index:      idx,
+			Name:       strings.TrimSpace(parts[1]),
+			BDF:        normalizeNvidiaBusID(strings.TrimSpace(parts[2])),
+			Serial:     strings.TrimSpace(parts[3]),
+			Status:     status,
+			RawLine:    line,
+			NeedsReset: needsReset,
+		})
+	}
+	sort.Slice(gpus, func(i, j int) bool { return gpus[i].Index < gpus[j].Index })
+	return gpus, nil
+}
+
+func normalizeNvidiaBusID(v string) string {
+	v = strings.TrimSpace(strings.ToLower(v))
+	parts := strings.Split(v, ":")
+	if len(parts) == 3 && len(parts[0]) > 4 {
+		parts[0] = parts[0][len(parts[0])-4:]
+		return strings.Join(parts, ":")
+	}
+	return v
+}
+
+func (s *System) ResetNvidiaGPU(index int) (string, error) {
+	if index < 0 {
+		return "", fmt.Errorf("gpu index must be >= 0")
+	}
+	raw, err := satExecCommand("nvidia-smi", "-r", "-i", strconv.Itoa(index)).CombinedOutput()
+	if len(raw) == 0 && err == nil {
+		raw = []byte("GPU reset completed.\n")
+	}
+	return string(raw), err
+}
+
 // RunNCCLTests runs nccl-tests all_reduce_perf across all NVIDIA GPUs.
 // Measures collective communication bandwidth over NVLink/PCIe.
 func (s *System) RunNCCLTests(ctx context.Context, baseDir string, logFunc func(string)) (string, error) {
@@ -268,13 +375,101 @@ func (s *System) RunNCCLTests(ctx context.Context, baseDir string, logFunc func(
 	if gpuCount < 1 {
 		gpuCount = 1
 	}
-	return runAcceptancePackCtx(ctx, baseDir, "nccl-tests", []satJob{
-		{name: "01-nvidia-smi-q.log", cmd: []string{"nvidia-smi", "-q"}},
-		{name: "02-all-reduce-perf.log", cmd: []string{
+	return runAcceptancePackCtx(ctx, baseDir, "nccl-tests", withNvidiaPersistenceMode(
+		satJob{name: "01-nvidia-smi-q.log", cmd: []string{"nvidia-smi", "-q"}},
+		satJob{name: "02-all-reduce-perf.log", cmd: []string{
 			"all_reduce_perf", "-b", "512M", "-e", "4G", "-f", "2",
 			"-g", strconv.Itoa(gpuCount), "--iters", "20",
 		}},
-	}, logFunc)
+	), logFunc)
+}
+
+func (s *System) RunNvidiaOfficialComputePack(ctx context.Context, baseDir string, durationSec int, gpuIndices []int, staggerSec int, logFunc func(string)) (string, error) {
+	selected, err := resolveDCGMGPUIndices(gpuIndices)
+	if err != nil {
+		return "", err
+	}
+	var (
+		profCmd []string
+		profEnv []string
+	)
+	if staggerSec > 0 && len(selected) > 1 {
+		profCmd = []string{
+			"bee-dcgmproftester-staggered",
+			"--seconds", strconv.Itoa(normalizeNvidiaBurnDuration(durationSec)),
+			"--stagger-seconds", strconv.Itoa(staggerSec),
+			"--devices", joinIndexList(selected),
+		}
+	} else {
+		profCmd, err = resolveDCGMProfTesterCommand("--no-dcgm-validation", "-t", "1004", "-d", strconv.Itoa(normalizeNvidiaBurnDuration(durationSec)))
+		if err != nil {
+			return "", err
+		}
+		profEnv = nvidiaVisibleDevicesEnv(selected)
+	}
+	return runAcceptancePackCtx(ctx, baseDir, "gpu-nvidia-compute", withNvidiaPersistenceMode(
+		satJob{name: "01-nvidia-smi-q.log", cmd: []string{"nvidia-smi", "-q"}},
+		satJob{name: "02-dcgmi-version.log", cmd: []string{"dcgmi", "-v"}},
+			satJob{
+				name:       "03-dcgmproftester.log",
+				cmd:        profCmd,
+				env:        profEnv,
+				collectGPU: true,
+				gpuIndices: selected,
+			},
+		satJob{name: "04-nvidia-smi-after.log", cmd: []string{"nvidia-smi", "--query-gpu=index,name,temperature.gpu,power.draw,utilization.gpu,memory.used,memory.total", "--format=csv,noheader,nounits"}},
+	), logFunc)
+}
+
+func (s *System) RunNvidiaTargetedPowerPack(ctx context.Context, baseDir string, durationSec int, gpuIndices []int, logFunc func(string)) (string, error) {
+	selected, err := resolveDCGMGPUIndices(gpuIndices)
+	if err != nil {
+		return "", err
+	}
+	return runAcceptancePackCtx(ctx, baseDir, "gpu-nvidia-targeted-power", withNvidiaPersistenceMode(
+		satJob{name: "01-nvidia-smi-q.log", cmd: []string{"nvidia-smi", "-q"}},
+		satJob{
+			name:       "02-dcgmi-targeted-power.log",
+			cmd:        nvidiaDCGMNamedDiagCommand("targeted_power", normalizeNvidiaBurnDuration(durationSec), selected),
+			collectGPU: true,
+			gpuIndices: selected,
+		},
+		satJob{name: "03-nvidia-smi-after.log", cmd: []string{"nvidia-smi", "--query-gpu=index,name,temperature.gpu,power.draw,utilization.gpu,memory.used,memory.total", "--format=csv,noheader,nounits"}},
+	), logFunc)
+}
+
+func (s *System) RunNvidiaPulseTestPack(ctx context.Context, baseDir string, durationSec int, gpuIndices []int, logFunc func(string)) (string, error) {
+	selected, err := resolveDCGMGPUIndices(gpuIndices)
+	if err != nil {
+		return "", err
+	}
+	return runAcceptancePackCtx(ctx, baseDir, "gpu-nvidia-pulse", withNvidiaPersistenceMode(
+		satJob{name: "01-nvidia-smi-q.log", cmd: []string{"nvidia-smi", "-q"}},
+		satJob{
+			name:       "02-dcgmi-pulse-test.log",
+			cmd:        nvidiaDCGMNamedDiagCommand("pulse_test", normalizeNvidiaBurnDuration(durationSec), selected),
+			collectGPU: true,
+			gpuIndices: selected,
+		},
+		satJob{name: "03-nvidia-smi-after.log", cmd: []string{"nvidia-smi", "--query-gpu=index,name,temperature.gpu,power.draw,utilization.gpu,memory.used,memory.total", "--format=csv,noheader,nounits"}},
+	), logFunc)
+}
+
+func (s *System) RunNvidiaBandwidthPack(ctx context.Context, baseDir string, gpuIndices []int, logFunc func(string)) (string, error) {
+	selected, err := resolveDCGMGPUIndices(gpuIndices)
+	if err != nil {
+		return "", err
+	}
+	return runAcceptancePackCtx(ctx, baseDir, "gpu-nvidia-bandwidth", withNvidiaPersistenceMode(
+		satJob{name: "01-nvidia-smi-q.log", cmd: []string{"nvidia-smi", "-q"}},
+		satJob{
+			name:       "02-dcgmi-nvbandwidth.log",
+			cmd:        nvidiaDCGMNamedDiagCommand("nvbandwidth", 0, selected),
+			collectGPU: true,
+			gpuIndices: selected,
+		},
+		satJob{name: "03-nvidia-smi-after.log", cmd: []string{"nvidia-smi", "--query-gpu=index,name,temperature.gpu,power.draw,utilization.gpu,memory.used,memory.total", "--format=csv,noheader,nounits"}},
+	), logFunc)
 }
 
 func (s *System) RunNvidiaAcceptancePack(baseDir string, logFunc func(string)) (string, error) {
@@ -286,12 +481,77 @@ func (s *System) RunNvidiaAcceptancePack(baseDir string, logFunc func(string)) (
 // gpuIndices: specific GPU indices to test (empty = all GPUs).
 // ctx cancellation kills the running job.
 func (s *System) RunNvidiaAcceptancePackWithOptions(ctx context.Context, baseDir string, diagLevel int, gpuIndices []int, logFunc func(string)) (string, error) {
-	return runAcceptancePackCtx(ctx, baseDir, "gpu-nvidia", nvidiaDCGMJobs(diagLevel, gpuIndices), logFunc)
+	resolvedGPUIndices, err := resolveDCGMGPUIndices(gpuIndices)
+	if err != nil {
+		return "", err
+	}
+	return runAcceptancePackCtx(ctx, baseDir, "gpu-nvidia", nvidiaDCGMJobs(diagLevel, resolvedGPUIndices), logFunc)
 }
 
-func (s *System) RunMemoryAcceptancePack(ctx context.Context, baseDir string, logFunc func(string)) (string, error) {
-	sizeMB := envInt("BEE_MEMTESTER_SIZE_MB", 128)
-	passes := envInt("BEE_MEMTESTER_PASSES", 1)
+func (s *System) RunNvidiaTargetedStressValidatePack(ctx context.Context, baseDir string, durationSec int, gpuIndices []int, logFunc func(string)) (string, error) {
+	selected, err := resolveDCGMGPUIndices(gpuIndices)
+	if err != nil {
+		return "", err
+	}
+	// Kill any lingering nvvs/dcgmi processes from a previous interrupted run
+	// before starting — otherwise dcgmi diag fails with DCGM_ST_IN_USE (-34).
+	if killed := KillTestWorkers(); len(killed) > 0 && logFunc != nil {
+		for _, p := range killed {
+			logFunc(fmt.Sprintf("pre-flight: killed stale worker pid=%d name=%s", p.PID, p.Name))
+		}
+	}
+	return runAcceptancePackCtx(ctx, baseDir, "gpu-nvidia-targeted-stress", withNvidiaPersistenceMode(
+		satJob{name: "01-nvidia-smi-q.log", cmd: []string{"nvidia-smi", "-q"}},
+		satJob{
+			name:       "02-dcgmi-targeted-stress.log",
+			cmd:        nvidiaDCGMNamedDiagCommand("targeted_stress", normalizeNvidiaBurnDuration(durationSec), selected),
+			collectGPU: true,
+			gpuIndices: selected,
+		},
+		satJob{name: "03-nvidia-smi-after.log", cmd: []string{"nvidia-smi", "--query-gpu=index,name,temperature.gpu,power.draw,utilization.gpu,memory.used,memory.total", "--format=csv,noheader,nounits"}},
+	), logFunc)
+}
+
+func resolveDCGMGPUIndices(gpuIndices []int) ([]int, error) {
+	if len(gpuIndices) > 0 {
+		return dedupeSortedIndices(gpuIndices), nil
+	}
+	all, err := listNvidiaGPUIndices()
+	if err != nil {
+		return nil, err
+	}
+	if len(all) == 0 {
+		return nil, fmt.Errorf("nvidia-smi found no NVIDIA GPUs")
+	}
+	return all, nil
+}
+
+func memoryStressSizeArg() string {
+	if mb := envInt("BEE_VM_STRESS_SIZE_MB", 0); mb > 0 {
+		return fmt.Sprintf("%dM", mb)
+	}
+	availBytes := satFreeMemBytes()
+	if availBytes <= 0 {
+		return "80%"
+	}
+	availMB := availBytes / (1024 * 1024)
+	targetMB := (availMB * 2) / 3
+	if targetMB >= 256 {
+		targetMB = (targetMB / 256) * 256
+	}
+	if targetMB <= 0 {
+		return "80%"
+	}
+	return fmt.Sprintf("%dM", targetMB)
+}
+
+func (s *System) RunMemoryAcceptancePack(ctx context.Context, baseDir string, sizeMB, passes int, logFunc func(string)) (string, error) {
+	if sizeMB <= 0 {
+		sizeMB = 256
+	}
+	if passes <= 0 {
+		passes = 1
+	}
 	return runAcceptancePackCtx(ctx, baseDir, "memory", []satJob{
 		{name: "01-free-before.log", cmd: []string{"free", "-h"}},
 		{name: "02-memtester.log", cmd: []string{"memtester", fmt.Sprintf("%dM", sizeMB), fmt.Sprintf("%d", passes)}},
@@ -304,11 +564,9 @@ func (s *System) RunMemoryStressPack(ctx context.Context, baseDir string, durati
 	if seconds <= 0 {
 		seconds = envInt("BEE_VM_STRESS_SECONDS", 300)
 	}
-	// Use 80% of RAM by default; override with BEE_VM_STRESS_SIZE_MB.
-	sizeArg := "80%"
-	if mb := envInt("BEE_VM_STRESS_SIZE_MB", 0); mb > 0 {
-		sizeArg = fmt.Sprintf("%dM", mb)
-	}
+	// Base the default on current MemAvailable and keep headroom for the OS and
+	// concurrent stressors so mixed burn runs do not trip the OOM killer.
+	sizeArg := memoryStressSizeArg()
 	return runAcceptancePackCtx(ctx, baseDir, "memory-stress", []satJob{
 		{name: "01-free-before.log", cmd: []string{"free", "-h"}},
 		{name: "02-stress-ng-vm.log", cmd: []string{
@@ -350,7 +608,7 @@ func (s *System) RunCPUAcceptancePack(ctx context.Context, baseDir string, durat
 	}, logFunc)
 }
 
-func (s *System) RunStorageAcceptancePack(ctx context.Context, baseDir string, logFunc func(string)) (string, error) {
+func (s *System) RunStorageAcceptancePack(ctx context.Context, baseDir string, extended bool, logFunc func(string)) (string, error) {
 	if baseDir == "" {
 		baseDir = "/var/log/bee-sat"
 	}
@@ -382,7 +640,7 @@ func (s *System) RunStorageAcceptancePack(ctx context.Context, baseDir string, l
 			break
 		}
 		prefix := fmt.Sprintf("%02d-%s", index+1, filepath.Base(devPath))
-		commands := storageSATCommands(devPath)
+		commands := storageSATCommands(devPath, extended)
 		for cmdIndex, job := range commands {
 			if ctx.Err() != nil {
 				break
@@ -425,14 +683,24 @@ type satStats struct {
 	Unsupported int
 }
 
+func withNvidiaPersistenceMode(jobs ...satJob) []satJob {
+	out := make([]satJob, 0, len(jobs)+1)
+	out = append(out, satJob{
+		name: "00-nvidia-smi-persistence-mode.log",
+		cmd:  []string{"nvidia-smi", "-pm", "1"},
+	})
+	out = append(out, jobs...)
+	return out
+}
+
 func nvidiaSATJobs() []satJob {
-	return []satJob{
-		{name: "01-nvidia-smi-q.log", cmd: []string{"nvidia-smi", "-q"}},
-		{name: "02-dmidecode-baseboard.log", cmd: []string{"dmidecode", "-t", "baseboard"}},
-		{name: "03-dmidecode-system.log", cmd: []string{"dmidecode", "-t", "system"}},
-		{name: "04-nvidia-bug-report.log", cmd: []string{"nvidia-bug-report.sh", "--output-file", "{{run_dir}}/nvidia-bug-report.log"}},
-		{name: "05-bee-gpu-burn.log", cmd: []string{"bee-gpu-burn", "--seconds", "5", "--size-mb", "64"}},
-	}
+	return withNvidiaPersistenceMode(
+		satJob{name: "01-nvidia-smi-q.log", cmd: []string{"nvidia-smi", "-q"}},
+		satJob{name: "02-dmidecode-baseboard.log", cmd: []string{"dmidecode", "-t", "baseboard"}},
+		satJob{name: "03-dmidecode-system.log", cmd: []string{"dmidecode", "-t", "system"}},
+		satJob{name: "04-nvidia-bug-report.log", cmd: []string{"nvidia-bug-report.sh", "--output-file", "{{run_dir}}/nvidia-bug-report.log"}},
+		satJob{name: "05-bee-gpu-burn.log", cmd: []string{"bee-gpu-burn", "--seconds", "5", "--size-mb", "64"}},
+	)
 }
 
 func nvidiaDCGMJobs(diagLevel int, gpuIndices []int) []satJob {
@@ -447,11 +715,39 @@ func nvidiaDCGMJobs(diagLevel int, gpuIndices []int) []satJob {
 		}
 		diagArgs = append(diagArgs, "-i", strings.Join(ids, ","))
 	}
-	return []satJob{
-		{name: "01-nvidia-smi-q.log", cmd: []string{"nvidia-smi", "-q"}},
-		{name: "02-dmidecode-baseboard.log", cmd: []string{"dmidecode", "-t", "baseboard"}},
-		{name: "03-dmidecode-system.log", cmd: []string{"dmidecode", "-t", "system"}},
-		{name: "04-dcgmi-diag.log", cmd: diagArgs},
+	return withNvidiaPersistenceMode(
+		satJob{name: "01-nvidia-smi-q.log", cmd: []string{"nvidia-smi", "-q"}},
+		satJob{name: "02-dmidecode-baseboard.log", cmd: []string{"dmidecode", "-t", "baseboard"}},
+		satJob{name: "03-dmidecode-system.log", cmd: []string{"dmidecode", "-t", "system"}},
+		satJob{name: "04-dcgmi-diag.log", cmd: diagArgs, gpuIndices: gpuIndices},
+	)
+}
+
+func nvidiaDCGMNamedDiagCommand(name string, durationSec int, gpuIndices []int) []string {
+	args := []string{"dcgmi", "diag", "-r", name}
+	if durationSec > 0 {
+		args = append(args, "-p", fmt.Sprintf("%s.test_duration=%d", name, durationSec))
+	}
+	if len(gpuIndices) > 0 {
+		args = append(args, "-i", joinIndexList(gpuIndices))
+	}
+	return args
+}
+
+func normalizeNvidiaBurnDuration(durationSec int) int {
+	if durationSec <= 0 {
+		return 300
+	}
+	return durationSec
+}
+
+func nvidiaVisibleDevicesEnv(gpuIndices []int) []string {
+	if len(gpuIndices) == 0 {
+		return nil
+	}
+	return []string{
+		"CUDA_DEVICE_ORDER=PCI_BUS_ID",
+		"CUDA_VISIBLE_DEVICES=" + joinIndexList(gpuIndices),
 	}
 }
 
@@ -471,11 +767,23 @@ func runAcceptancePackCtx(ctx context.Context, baseDir, prefix string, jobs []sa
 
 	var summary strings.Builder
 	stats := satStats{}
+	nvidiaPack := strings.HasPrefix(prefix, "gpu-nvidia")
+	perGPU := map[int]*nvidiaGPUStatusFile{}
+	selectedGPUIndices := map[int]struct{}{}
 	fmt.Fprintf(&summary, "run_at_utc=%s\n", time.Now().UTC().Format(time.RFC3339))
 	for _, job := range jobs {
 		if ctx.Err() != nil {
 			break
 		}
+		for _, idx := range job.gpuIndices {
+			selectedGPUIndices[idx] = struct{}{}
+			status := perGPU[idx]
+			if status == nil {
+				status = &nvidiaGPUStatusFile{Index: idx}
+				perGPU[idx] = status
+			}
+			status.Selected = true
+		}
 		cmd := make([]string, 0, len(job.cmd))
 		for _, arg := range job.cmd {
 			cmd = append(cmd, strings.ReplaceAll(arg, "{{run_dir}}", runDir))
@@ -484,17 +792,52 @@ func runAcceptancePackCtx(ctx context.Context, baseDir, prefix string, jobs []sa
 		var out []byte
 		var err error
 
-		if job.collectGPU {
-			out, err = runSATCommandWithMetrics(ctx, verboseLog, job.name, cmd, job.env, job.gpuIndices, runDir, logFunc)
-		} else {
-			out, err = runSATCommandCtx(ctx, verboseLog, job.name, cmd, job.env, logFunc)
+		if nvidiaPack && nvidiaJobNeedsHealthCheck(job) {
+			if msg, healthErr := checkNvidiaJobHealth(job.gpuIndices); healthErr != nil {
+				if logFunc != nil {
+					logFunc(msg)
+				}
+				out = []byte(msg + "\n")
+				err = healthErr
+			}
+		}
+
+		if err == nil {
+			if job.collectGPU {
+				out, err = runSATCommandWithMetrics(ctx, verboseLog, job.name, cmd, job.env, job.gpuIndices, runDir, logFunc)
+			} else {
+				out, err = runSATCommandCtx(ctx, verboseLog, job.name, cmd, job.env, logFunc)
+			}
+		}
+
+		if nvidiaPack && nvidiaJobNeedsHealthCheck(job) {
+			if msg, healthErr := checkNvidiaJobHealth(job.gpuIndices); healthErr != nil {
+				if logFunc != nil {
+					logFunc(msg)
+				}
+				if len(out) > 0 && !bytes.HasSuffix(out, []byte("\n")) {
+					out = append(out, '\n')
+				}
+				out = append(out, []byte(msg+"\n")...)
+				if err == nil {
+					err = healthErr
+				}
+			}
 		}
 
 		if writeErr := os.WriteFile(filepath.Join(runDir, job.name), out, 0644); writeErr != nil {
 			return "", writeErr
 		}
+		if ctx.Err() != nil {
+			return "", ctx.Err()
+		}
 		status, rc := classifySATResult(job.name, out, err)
 		stats.Add(status)
+		if nvidiaPack && len(job.gpuIndices) > 0 && nvidiaJobNeedsHealthCheck(job) {
+			for _, idx := range job.gpuIndices {
+				updateNvidiaGPUStatus(perGPU, idx, status, job.name, string(out))
+			}
+		}
 		key := strings.TrimSuffix(strings.TrimPrefix(job.name, "0"), ".log")
 		fmt.Fprintf(&summary, "%s_rc=%d\n", key, rc)
 		fmt.Fprintf(&summary, "%s_status=%s\n", key, status)
@@ -503,6 +846,11 @@ func runAcceptancePackCtx(ctx context.Context, baseDir, prefix string, jobs []sa
 	if err := os.WriteFile(filepath.Join(runDir, "summary.txt"), []byte(summary.String()), 0644); err != nil {
 		return "", err
 	}
+	if nvidiaPack {
+		if err := writeNvidiaGPUStatusFiles(runDir, stats.Overall(), perGPU, selectedGPUIndices); err != nil {
+			return "", err
+		}
+	}
 
 	archive := filepath.Join(baseDir, prefix+"-"+ts+".tar.gz")
 	if err := createTarGz(archive, runDir); err != nil {
@@ -511,6 +859,197 @@ func runAcceptancePackCtx(ctx context.Context, baseDir, prefix string, jobs []sa
 	return archive, nil
 }
 
+func updateNvidiaGPUStatus(perGPU map[int]*nvidiaGPUStatusFile, idx int, status, jobName, detail string) {
+	entry := perGPU[idx]
+	if entry == nil {
+		entry = &nvidiaGPUStatusFile{Index: idx}
+		perGPU[idx] = entry
+	}
+	if nvidiaSATStatusSeverity(status) >= nvidiaSATStatusSeverity(entry.RunStatus) {
+		entry.RunStatus = status
+		entry.FailingJob = jobName
+		entry.Reason = firstLine(detail)
+	}
+}
+
+func writeNvidiaGPUStatusFiles(runDir, overall string, perGPU map[int]*nvidiaGPUStatusFile, selected map[int]struct{}) error {
+	health, err := readNvidiaGPUHealth()
+	if err == nil {
+		for _, gpu := range health {
+			entry := perGPU[gpu.Index]
+			if entry == nil {
+				entry = &nvidiaGPUStatusFile{Index: gpu.Index}
+				perGPU[gpu.Index] = entry
+			}
+			entry.Name = gpu.Name
+			entry.Observed = true
+			entry.HealthRaw = gpu.RawLine
+			if gpu.NeedsReset {
+				entry.Health = "RESET_REQUIRED"
+				if entry.RunStatus == "" || nvidiaSATStatusSeverity("FAILED") >= nvidiaSATStatusSeverity(entry.RunStatus) {
+					entry.RunStatus = "FAILED"
+					if strings.TrimSpace(entry.Reason) == "" {
+						entry.Reason = "GPU requires reset"
+					}
+				}
+			} else {
+				entry.Health = "OK"
+			}
+		}
+	}
+	for idx := range selected {
+		entry := perGPU[idx]
+		if entry == nil {
+			entry = &nvidiaGPUStatusFile{Index: idx}
+			perGPU[idx] = entry
+		}
+		entry.Selected = true
+	}
+	var indices []int
+	for idx := range perGPU {
+		indices = append(indices, idx)
+	}
+	sort.Ints(indices)
+	for _, idx := range indices {
+		entry := perGPU[idx]
+		if entry.RunStatus == "" {
+			entry.RunStatus = overall
+		}
+		if entry.Health == "" {
+			entry.Health = "UNKNOWN"
+		}
+		if entry.Name == "" {
+			entry.Name = "unknown"
+		}
+		var body strings.Builder
+		fmt.Fprintf(&body, "gpu_index=%d\n", entry.Index)
+		fmt.Fprintf(&body, "gpu_name=%s\n", entry.Name)
+		fmt.Fprintf(&body, "selected=%t\n", entry.Selected)
+		fmt.Fprintf(&body, "observed=%t\n", entry.Observed)
+		fmt.Fprintf(&body, "run_status=%s\n", entry.RunStatus)
+		fmt.Fprintf(&body, "health_status=%s\n", entry.Health)
+		if strings.TrimSpace(entry.FailingJob) != "" {
+			fmt.Fprintf(&body, "failing_job=%s\n", entry.FailingJob)
+		}
+		if strings.TrimSpace(entry.Reason) != "" {
+			fmt.Fprintf(&body, "reason=%s\n", entry.Reason)
+		}
+		if strings.TrimSpace(entry.HealthRaw) != "" {
+			fmt.Fprintf(&body, "health_raw=%s\n", entry.HealthRaw)
+		}
+		if err := os.WriteFile(filepath.Join(runDir, fmt.Sprintf("gpu-%d-status.txt", idx)), []byte(body.String()), 0644); err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
+func nvidiaSATStatusSeverity(status string) int {
+	switch strings.ToUpper(strings.TrimSpace(status)) {
+	case "FAILED":
+		return 3
+	case "PARTIAL", "UNSUPPORTED":
+		return 2
+	case "OK":
+		return 1
+	default:
+		return 0
+	}
+}
+
+func firstLine(s string) string {
+	s = strings.TrimSpace(s)
+	if s == "" {
+		return ""
+	}
+	if idx := strings.IndexByte(s, '\n'); idx >= 0 {
+		return strings.TrimSpace(s[:idx])
+	}
+	return s
+}
+
+func nvidiaJobNeedsHealthCheck(job satJob) bool {
+	if job.collectGPU {
+		return true
+	}
+	name := strings.ToLower(strings.TrimSpace(job.name))
+	return strings.Contains(name, "dcgmi") ||
+		strings.Contains(name, "gpu-burn") ||
+		strings.Contains(name, "gpu-stress") ||
+		strings.Contains(name, "dcgmproftester")
+}
+
+func checkNvidiaJobHealth(selected []int) (string, error) {
+	health, err := readNvidiaGPUHealth()
+	if err != nil {
+		return "", nil
+	}
+	var bad []nvidiaGPUHealth
+	selectedSet := make(map[int]struct{}, len(selected))
+	for _, idx := range selected {
+		selectedSet[idx] = struct{}{}
+	}
+	for _, gpu := range health {
+		if len(selectedSet) > 0 {
+			if _, ok := selectedSet[gpu.Index]; !ok {
+				continue
+			}
+		}
+		if gpu.NeedsReset {
+			bad = append(bad, gpu)
+		}
+	}
+	if len(bad) == 0 {
+		return "", nil
+	}
+	lines := make([]string, 0, len(bad)+1)
+	lines = append(lines, "NVIDIA GPU health check failed:")
+	for _, gpu := range bad {
+		lines = append(lines, fmt.Sprintf("gpu %d (%s) requires reset: %s", gpu.Index, gpu.Name, gpu.RawLine))
+	}
+	return strings.Join(lines, "\n"), errors.New("nvidia gpu requires reset")
+}
+
+func readNvidiaGPUHealth() ([]nvidiaGPUHealth, error) {
+	out, err := satExecCommand(
+		"nvidia-smi",
+		"--query-gpu=index,name,temperature.gpu,power.draw,utilization.gpu,memory.used,memory.total",
+		"--format=csv,noheader,nounits",
+	).Output()
+	if err != nil {
+		return nil, fmt.Errorf("nvidia-smi: %w", err)
+	}
+	return parseNvidiaGPUHealth(string(out)), nil
+}
+
+func parseNvidiaGPUHealth(raw string) []nvidiaGPUHealth {
+	var gpus []nvidiaGPUHealth
+	for _, line := range strings.Split(strings.TrimSpace(raw), "\n") {
+		line = strings.TrimSpace(line)
+		if line == "" {
+			continue
+		}
+		parts := strings.Split(line, ",")
+		if len(parts) < 2 {
+			gpus = append(gpus, nvidiaGPUHealth{RawLine: line, ParseFailure: true})
+			continue
+		}
+		idx, err := strconv.Atoi(strings.TrimSpace(parts[0]))
+		if err != nil {
+			gpus = append(gpus, nvidiaGPUHealth{RawLine: line, ParseFailure: true})
+			continue
+		}
+		upper := strings.ToUpper(line)
+		gpus = append(gpus, nvidiaGPUHealth{
+			Index:      idx,
+			Name:       strings.TrimSpace(parts[1]),
+			NeedsReset: strings.Contains(upper, "GPU REQUIRES RESET"),
+			RawLine:    line,
+		})
+	}
+	return gpus
+}
+
 func runSATCommandCtx(ctx context.Context, verboseLog, name string, cmd []string, env []string, logFunc func(string)) ([]byte, error) {
 	start := time.Now().UTC()
 	resolvedCmd, err := resolveSATCommand(cmd)
@@ -565,17 +1104,25 @@ func listStorageDevices() ([]string, error) {
 	return parseStorageDevices(string(out)), nil
 }
 
-func storageSATCommands(devPath string) []satJob {
+func storageSATCommands(devPath string, extended bool) []satJob {
 	if strings.Contains(filepath.Base(devPath), "nvme") {
+		selfTestLevel := "1"
+		if extended {
+			selfTestLevel = "2"
+		}
 		return []satJob{
 			{name: "nvme-id-ctrl", cmd: []string{"nvme", "id-ctrl", devPath, "-o", "json"}},
 			{name: "nvme-smart-log", cmd: []string{"nvme", "smart-log", devPath, "-o", "json"}},
-			{name: "nvme-device-self-test", cmd: []string{"nvme", "device-self-test", devPath, "-s", "1", "--wait"}},
+			{name: "nvme-device-self-test", cmd: []string{"nvme", "device-self-test", devPath, "-s", selfTestLevel, "--wait"}},
 		}
 	}
+	smartTestType := "short"
+	if extended {
+		smartTestType = "long"
+	}
 	return []satJob{
 		{name: "smartctl-health", cmd: []string{"smartctl", "-H", "-A", devPath}},
-		{name: "smartctl-self-test-short", cmd: []string{"smartctl", "-t", "short", devPath}},
+		{name: "smartctl-self-test-short", cmd: []string{"smartctl", "-t", smartTestType, devPath}},
 	}
 }
 
@@ -624,6 +1171,7 @@ func classifySATResult(name string, out []byte, err error) (string, int) {
 	}
 	if strings.Contains(text, "unsupported") ||
 		strings.Contains(text, "not supported") ||
+		strings.Contains(text, "not found in path") ||
 		strings.Contains(text, "invalid opcode") ||
 		strings.Contains(text, "unknown command") ||
 		strings.Contains(text, "not implemented") ||
@@ -633,6 +1181,11 @@ func classifySATResult(name string, out []byte, err error) (string, int) {
 		// nvidia-smi on a machine with no NVIDIA GPU
 		strings.Contains(text, "couldn't communicate with the nvidia driver") ||
 		strings.Contains(text, "no nvidia gpu") ||
+		// Some NVMe firmwares start self-test but never expose progress to nvme-cli
+		// while waiting, so the CLI stops polling without proving device failure.
+		(strings.Contains(name, "self-test") &&
+			strings.Contains(text, "no progress for") &&
+			strings.Contains(text, "stop waiting")) ||
 		(strings.Contains(name, "self-test") && strings.Contains(text, "aborted")) {
 		return "UNSUPPORTED", rc
 	}
@@ -730,6 +1283,15 @@ func resolveROCmSMICommand(args ...string) ([]string, error) {
 	return nil, errors.New("rocm-smi not found in PATH or under /opt/rocm")
 }
 
+func resolveDCGMProfTesterCommand(args ...string) ([]string, error) {
+	for _, candidate := range dcgmProfTesterCandidates {
+		if path, err := satLookPath(candidate); err == nil {
+			return append([]string{path}, args...), nil
+		}
+	}
+	return nil, errors.New("dcgmproftester not found in PATH")
+}
+
 func ensureAMDRuntimeReady() error {
 	if _, err := os.Stat("/dev/kfd"); err == nil {
 		return nil

audit/internal/platform/sat_fan_stress.go

@@ -20,7 +20,7 @@ type FanStressOptions struct {
 	Phase1DurSec int   // first load phase duration in seconds (default 300)
 	PauseSec     int   // pause between the two load phases (default 60)
 	Phase2DurSec int   // second load phase duration in seconds (default 300)
-	SizeMB       int   // GPU memory to allocate per GPU during stress (default 64)
+	SizeMB       int   // GPU memory to allocate per GPU during stress (0 = auto: 95% of VRAM)
 	GPUIndices   []int // which GPU indices to stress (empty = all detected)
 }
 
@@ -243,9 +243,8 @@ func applyFanStressDefaults(opts *FanStressOptions) {
 	if opts.Phase2DurSec <= 0 {
 		opts.Phase2DurSec = 300
 	}
-	if opts.SizeMB <= 0 {
-		opts.SizeMB = 64
-	}
+	// SizeMB == 0 means "auto" (worker picks 95% of GPU VRAM for maximum power draw).
+	// Leave at 0 to avoid passing a too-small size that starves the tensor-core path.
 }
 
 // sampleFanStressRow collects all metrics for one telemetry sample.

audit/internal/platform/sat_test.go

@@ -1,23 +1,25 @@
 package platform
 
 import (
+	"context"
 	"errors"
 	"os"
 	"os/exec"
 	"path/filepath"
 	"strings"
 	"testing"
+	"time"
 )
 
 func TestStorageSATCommands(t *testing.T) {
 	t.Parallel()
 
-	nvme := storageSATCommands("/dev/nvme0n1")
+	nvme := storageSATCommands("/dev/nvme0n1", false)
 	if len(nvme) != 3 || nvme[2].cmd[0] != "nvme" {
 		t.Fatalf("unexpected nvme commands: %#v", nvme)
 	}
 
-	sata := storageSATCommands("/dev/sda")
+	sata := storageSATCommands("/dev/sda", false)
 	if len(sata) != 2 || sata[0].cmd[0] != "smartctl" {
 		t.Fatalf("unexpected sata commands: %#v", sata)
 	}
@@ -28,13 +30,19 @@ func TestRunNvidiaAcceptancePackIncludesGPUStress(t *testing.T) {
 
 	jobs := nvidiaSATJobs()
 
-	if len(jobs) != 5 {
-		t.Fatalf("jobs=%d want 5", len(jobs))
+	if len(jobs) != 6 {
+		t.Fatalf("jobs=%d want 6", len(jobs))
 	}
-	if got := jobs[4].cmd[0]; got != "bee-gpu-burn" {
+	if got := jobs[0].cmd[0]; got != "nvidia-smi" {
+		t.Fatalf("preflight command=%q want nvidia-smi", got)
+	}
+	if got := strings.Join(jobs[0].cmd, " "); got != "nvidia-smi -pm 1" {
+		t.Fatalf("preflight=%q want %q", got, "nvidia-smi -pm 1")
+	}
+	if got := jobs[5].cmd[0]; got != "bee-gpu-burn" {
 		t.Fatalf("gpu stress command=%q want bee-gpu-burn", got)
 	}
-	if got := jobs[3].cmd[1]; got != "--output-file" {
+	if got := jobs[4].cmd[1]; got != "--output-file" {
 		t.Fatalf("bug report flag=%q want --output-file", got)
 	}
 }
@@ -82,7 +90,7 @@ func TestAMDStressJobsIncludeBandwidthAndGST(t *testing.T) {
 
 func TestNvidiaSATJobsUseBuiltinBurnDefaults(t *testing.T) {
 	jobs := nvidiaSATJobs()
-	got := jobs[4].cmd
+	got := jobs[5].cmd
 	want := []string{"bee-gpu-burn", "--seconds", "5", "--size-mb", "64"}
 	if len(got) != len(want) {
 		t.Fatalf("cmd len=%d want %d", len(got), len(want))
@@ -94,6 +102,19 @@ func TestNvidiaSATJobsUseBuiltinBurnDefaults(t *testing.T) {
 	}
 }
 
+func TestNvidiaDCGMJobsEnablePersistenceModeBeforeDiag(t *testing.T) {
+	jobs := nvidiaDCGMJobs(3, []int{2, 0})
+	if len(jobs) != 5 {
+		t.Fatalf("jobs=%d want 5", len(jobs))
+	}
+	if got := strings.Join(jobs[0].cmd, " "); got != "nvidia-smi -pm 1" {
+		t.Fatalf("preflight=%q want %q", got, "nvidia-smi -pm 1")
+	}
+	if got := strings.Join(jobs[4].cmd, " "); got != "dcgmi diag -r 3 -i 2,0" {
+		t.Fatalf("diag=%q want %q", got, "dcgmi diag -r 3 -i 2,0")
+	}
+}
+
 func TestBuildNvidiaStressJobUsesSelectedLoaderAndDevices(t *testing.T) {
 	t.Parallel()
 
@@ -162,6 +183,157 @@ func TestBuildNvidiaStressJobUsesNCCLLoader(t *testing.T) {
 	}
 }
 
+func TestResolveDCGMGPUIndicesUsesDetectedGPUsWhenUnset(t *testing.T) {
+	t.Parallel()
+
+	oldExecCommand := satExecCommand
+	satExecCommand = func(name string, args ...string) *exec.Cmd {
+		if name == "nvidia-smi" {
+			return exec.Command("sh", "-c", "printf '2\n0\n1\n'")
+		}
+		return exec.Command(name, args...)
+	}
+	t.Cleanup(func() { satExecCommand = oldExecCommand })
+
+	got, err := resolveDCGMGPUIndices(nil)
+	if err != nil {
+		t.Fatalf("resolveDCGMGPUIndices error: %v", err)
+	}
+	if want := "0,1,2"; joinIndexList(got) != want {
+		t.Fatalf("gpuIndices=%q want %q", joinIndexList(got), want)
+	}
+}
+
+func TestResolveDCGMGPUIndicesKeepsExplicitSelection(t *testing.T) {
+	t.Parallel()
+
+	got, err := resolveDCGMGPUIndices([]int{3, 1, 3})
+	if err != nil {
+		t.Fatalf("resolveDCGMGPUIndices error: %v", err)
+	}
+	if want := "1,3"; joinIndexList(got) != want {
+		t.Fatalf("gpuIndices=%q want %q", joinIndexList(got), want)
+	}
+}
+
+func TestParseNvidiaGPUHealthDetectsResetRequired(t *testing.T) {
+	t.Parallel()
+
+	got := parseNvidiaGPUHealth("0, NVIDIA H100 PCIe, 38, 46.89, 0, 0, 81559\n1, NVIDIA H100 PCIe, [GPU requires reset], [N/A], [N/A], 0, 81559\n")
+	if len(got) != 2 {
+		t.Fatalf("len=%d want 2", len(got))
+	}
+	if got[0].NeedsReset {
+		t.Fatalf("gpu0 unexpectedly marked reset-required")
+	}
+	if !got[1].NeedsReset {
+		t.Fatalf("gpu1 should be marked reset-required: %#v", got[1])
+	}
+}
+
+func TestCheckNvidiaJobHealthReturnsErrorForSelectedResetRequiredGPU(t *testing.T) {
+	oldExecCommand := satExecCommand
+	satExecCommand = func(name string, args ...string) *exec.Cmd {
+		if name == "nvidia-smi" {
+			return exec.Command("sh", "-c", "printf '0, NVIDIA H100 PCIe, 38, 46.89, 0, 0, 81559\n1, NVIDIA H100 PCIe, [GPU requires reset], [N/A], [N/A], 0, 81559\n'")
+		}
+		return exec.Command(name, args...)
+	}
+	t.Cleanup(func() { satExecCommand = oldExecCommand })
+
+	msg, err := checkNvidiaJobHealth([]int{1})
+	if err == nil {
+		t.Fatal("expected health check error")
+	}
+	if !strings.Contains(msg, "gpu 1") || !strings.Contains(strings.ToLower(msg), "requires reset") {
+		t.Fatalf("unexpected message: %q", msg)
+	}
+}
+
+func TestWriteNvidiaGPUStatusFilesCreatesPerGPUFiles(t *testing.T) {
+	dir := t.TempDir()
+	oldExecCommand := satExecCommand
+	satExecCommand = func(name string, args ...string) *exec.Cmd {
+		if name == "nvidia-smi" {
+			return exec.Command("sh", "-c", "printf '0, NVIDIA H100 PCIe, 38, 46.89, 0, 0, 81559\n1, NVIDIA H100 PCIe, [GPU requires reset], [N/A], [N/A], 0, 81559\n'")
+		}
+		return exec.Command(name, args...)
+	}
+	t.Cleanup(func() { satExecCommand = oldExecCommand })
+
+	perGPU := map[int]*nvidiaGPUStatusFile{
+		0: {Index: 0, RunStatus: "OK"},
+		1: {Index: 1, RunStatus: "FAILED", FailingJob: "02-dcgmi-targeted-stress.log", Reason: "NVIDIA GPU health check failed:"},
+	}
+	if err := writeNvidiaGPUStatusFiles(dir, "FAILED", perGPU, map[int]struct{}{0: {}, 1: {}}); err != nil {
+		t.Fatalf("writeNvidiaGPUStatusFiles error: %v", err)
+	}
+	raw, err := os.ReadFile(filepath.Join(dir, "gpu-1-status.txt"))
+	if err != nil {
+		t.Fatalf("ReadFile gpu-1-status.txt: %v", err)
+	}
+	text := string(raw)
+	if !strings.Contains(text, "run_status=FAILED") {
+		t.Fatalf("missing run status:\n%s", text)
+	}
+	if !strings.Contains(text, "health_status=RESET_REQUIRED") {
+		t.Fatalf("missing health status:\n%s", text)
+	}
+	if !strings.Contains(text, "failing_job=02-dcgmi-targeted-stress.log") {
+		t.Fatalf("missing failing job:\n%s", text)
+	}
+}
+
+func TestResolveDCGMProfTesterCommandUsesVersionedBinary(t *testing.T) {
+	oldLookPath := satLookPath
+	satLookPath = func(file string) (string, error) {
+		switch file {
+		case "dcgmproftester13":
+			return "/usr/bin/dcgmproftester13", nil
+		default:
+			return "", exec.ErrNotFound
+		}
+	}
+	t.Cleanup(func() { satLookPath = oldLookPath })
+
+	cmd, err := resolveDCGMProfTesterCommand("--no-dcgm-validation", "-t", "1004")
+	if err != nil {
+		t.Fatalf("resolveDCGMProfTesterCommand error: %v", err)
+	}
+	if len(cmd) != 4 {
+		t.Fatalf("cmd len=%d want 4 (%v)", len(cmd), cmd)
+	}
+	if cmd[0] != "/usr/bin/dcgmproftester13" {
+		t.Fatalf("cmd[0]=%q want /usr/bin/dcgmproftester13", cmd[0])
+	}
+}
+
+func TestNvidiaDCGMNamedDiagCommandUsesDurationAndSelection(t *testing.T) {
+	cmd := nvidiaDCGMNamedDiagCommand("targeted_power", 900, []int{3, 1})
+	want := []string{"dcgmi", "diag", "-r", "targeted_power", "-p", "targeted_power.test_duration=900", "-i", "3,1"}
+	if len(cmd) != len(want) {
+		t.Fatalf("cmd len=%d want %d (%v)", len(cmd), len(want), cmd)
+	}
+	for i := range want {
+		if cmd[i] != want[i] {
+			t.Fatalf("cmd[%d]=%q want %q", i, cmd[i], want[i])
+		}
+	}
+}
+
+func TestNvidiaVisibleDevicesEnvUsesSelectedGPUs(t *testing.T) {
+	env := nvidiaVisibleDevicesEnv([]int{0, 2, 4})
+	if len(env) != 2 {
+		t.Fatalf("env len=%d want 2 (%v)", len(env), env)
+	}
+	if env[0] != "CUDA_DEVICE_ORDER=PCI_BUS_ID" {
+		t.Fatalf("env[0]=%q want CUDA_DEVICE_ORDER=PCI_BUS_ID", env[0])
+	}
+	if env[1] != "CUDA_VISIBLE_DEVICES=0,2,4" {
+		t.Fatalf("env[1]=%q want CUDA_VISIBLE_DEVICES=0,2,4", env[1])
+	}
+}
+
 func TestNvidiaStressArchivePrefixByLoader(t *testing.T) {
 	t.Parallel()
 
@@ -196,6 +368,37 @@ func TestEnvIntFallback(t *testing.T) {
 	}
 }
 
+func TestMemoryStressSizeArgUsesAvailableMemory(t *testing.T) {
+	oldFreeMemBytes := satFreeMemBytes
+	satFreeMemBytes = func() int64 { return 96 * 1024 * 1024 * 1024 }
+	t.Cleanup(func() { satFreeMemBytes = oldFreeMemBytes })
+
+	if got := memoryStressSizeArg(); got != "65536M" {
+		t.Fatalf("sizeArg=%q want 65536M", got)
+	}
+}
+
+func TestMemoryStressSizeArgRespectsOverride(t *testing.T) {
+	oldFreeMemBytes := satFreeMemBytes
+	satFreeMemBytes = func() int64 { return 96 * 1024 * 1024 * 1024 }
+	t.Cleanup(func() { satFreeMemBytes = oldFreeMemBytes })
+	t.Setenv("BEE_VM_STRESS_SIZE_MB", "4096")
+
+	if got := memoryStressSizeArg(); got != "4096M" {
+		t.Fatalf("sizeArg=%q want 4096M", got)
+	}
+}
+
+func TestMemoryStressSizeArgFallsBackWhenFreeMemoryUnknown(t *testing.T) {
+	oldFreeMemBytes := satFreeMemBytes
+	satFreeMemBytes = func() int64 { return 0 }
+	t.Cleanup(func() { satFreeMemBytes = oldFreeMemBytes })
+
+	if got := memoryStressSizeArg(); got != "80%" {
+		t.Fatalf("sizeArg=%q want 80%%", got)
+	}
+}
+
 func TestClassifySATResult(t *testing.T) {
 	tests := []struct {
 		name   string
@@ -206,6 +409,7 @@ func TestClassifySATResult(t *testing.T) {
 	}{
 		{name: "ok", job: "memtester", out: "done", err: nil, status: "OK"},
 		{name: "unsupported", job: "smartctl-self-test-short", out: "Self-test not supported", err: errors.New("rc 1"), status: "UNSUPPORTED"},
+		{name: "nvme wait timeout without progress", job: "nvme-device-self-test", out: "Short Device self-test started\nWaiting for self test completion...\nno progress for 78 seconds, stop waiting", err: errors.New("rc 1"), status: "UNSUPPORTED"},
 		{name: "failed", job: "bee-gpu-burn", out: "cuda error", err: errors.New("rc 1"), status: "FAILED"},
 		{name: "cuda not ready", job: "bee-gpu-burn", out: "cuInit failed: CUDA_ERROR_SYSTEM_NOT_READY", err: errors.New("rc 1"), status: "UNSUPPORTED"},
 	}
@@ -220,6 +424,38 @@ func TestClassifySATResult(t *testing.T) {
 	}
 }
 
+func TestRunAcceptancePackCtxReturnsContextErrorWithoutArchive(t *testing.T) {
+	dir := t.TempDir()
+	ctx, cancel := context.WithCancel(context.Background())
+	t.Cleanup(cancel)
+
+	done := make(chan struct{})
+	go func() {
+		time.Sleep(100 * time.Millisecond)
+		cancel()
+		close(done)
+	}()
+
+	archive, err := runAcceptancePackCtx(ctx, dir, "cancelled-pack", []satJob{
+		{name: "01-sleep.log", cmd: []string{"sh", "-c", "sleep 5"}},
+	}, nil)
+	<-done
+
+	if !errors.Is(err, context.Canceled) {
+		t.Fatalf("err=%v want context.Canceled", err)
+	}
+	if archive != "" {
+		t.Fatalf("archive=%q want empty", archive)
+	}
+	matches, globErr := filepath.Glob(filepath.Join(dir, "cancelled-pack-*.tar.gz"))
+	if globErr != nil {
+		t.Fatalf("Glob error: %v", globErr)
+	}
+	if len(matches) != 0 {
+		t.Fatalf("archives=%v want none", matches)
+	}
+}
+
 func TestParseStorageDevicesSkipsUSBDisks(t *testing.T) {
 	t.Parallel()
 

audit/internal/platform/services.go

@@ -10,17 +10,30 @@ import (
 func (s *System) ListBeeServices() ([]string, error) {
 	seen := map[string]bool{}
 	var out []string
-	for _, pattern := range []string{"/etc/systemd/system/bee-*.service", "/lib/systemd/system/bee-*.service"} {
+	for _, pattern := range []string{
+		"/etc/systemd/system/bee-*.service",
+		"/lib/systemd/system/bee-*.service",
+		"/etc/systemd/system/bee-*.timer",
+		"/lib/systemd/system/bee-*.timer",
+	} {
 		matches, err := filepath.Glob(pattern)
 		if err != nil {
 			return nil, err
 		}
 		for _, match := range matches {
-			name := strings.TrimSuffix(filepath.Base(match), ".service")
+			base := filepath.Base(match)
+			name := base
+			if strings.HasSuffix(base, ".service") {
+				name = strings.TrimSuffix(base, ".service")
+			}
 			// Skip template units (e.g. bee-journal-mirror@) — they have no instances to query.
 			if strings.HasSuffix(name, "@") {
 				continue
 			}
+			// bee-selfheal is timer-managed; showing the oneshot service as inactive is misleading.
+			if name == "bee-selfheal" && strings.HasSuffix(base, ".service") {
+				continue
+			}
 			if !seen[name] {
 				seen[name] = true
 				out = append(out, name)
@@ -48,7 +61,9 @@ func (s *System) ServiceState(name string) string {
 }
 
 func (s *System) ServiceDo(name string, action ServiceAction) (string, error) {
-	raw, err := exec.Command("systemctl", string(action), name).CombinedOutput()
+	// bee-web runs as the bee user; sudo is required to control system services.
+	// /etc/sudoers.d/bee grants bee NOPASSWD:ALL.
+	raw, err := exec.Command("sudo", "systemctl", string(action), name).CombinedOutput()
 	return string(raw), err
 }
 

audit/internal/platform/types.go

@@ -44,12 +44,12 @@ type StaticIPv4Config struct {
 }
 
 type RemovableTarget struct {
-	Device     string
-	FSType     string
-	Size       string
-	Label      string
-	Model      string
-	Mountpoint string
+	Device     string `json:"device"`
+	FSType     string `json:"fs_type"`
+	Size       string `json:"size"`
+	Label      string `json:"label"`
+	Model      string `json:"model"`
+	Mountpoint string `json:"mountpoint"`
 }
 
 type ToolStatus struct {
@@ -70,6 +70,7 @@ type NvidiaStressOptions struct {
 	Loader            string
 	GPUIndices        []int
 	ExcludeGPUIndices []int
+	StaggerSeconds    int
 }
 
 func New() *System {

audit/internal/platform/types_test.go

@@ -0,0 +1,31 @@
+package platform
+
+import (
+	"encoding/json"
+	"strings"
+	"testing"
+)
+
+func TestRemovableTargetJSONUsesFrontendFieldNames(t *testing.T) {
+	t.Parallel()
+
+	data, err := json.Marshal(RemovableTarget{
+		Device: "/dev/sdb1",
+		FSType: "exfat",
+		Size:   "1.8T",
+		Label:  "USB",
+		Model:  "Flash",
+	})
+	if err != nil {
+		t.Fatalf("marshal: %v", err)
+	}
+	raw := string(data)
+	for _, key := range []string{`"device"`, `"fs_type"`, `"size"`, `"label"`, `"model"`} {
+		if !strings.Contains(raw, key) {
+			t.Fatalf("json missing key %s: %s", key, raw)
+		}
+	}
+	if strings.Contains(raw, `"Device"`) || strings.Contains(raw, `"FSType"`) {
+		t.Fatalf("json still contains Go field names: %s", raw)
+	}
+}

audit/internal/schema/hardware.go

@@ -20,6 +20,7 @@ type RuntimeHealth struct {
 	ExportDir     string                 `json:"export_dir,omitempty"`
 	DriverReady   bool                   `json:"driver_ready,omitempty"`
 	CUDAReady     bool                   `json:"cuda_ready,omitempty"`
+	NvidiaGSPMode string                 `json:"nvidia_gsp_mode,omitempty"` // "gsp-on", "gsp-off", "gsp-stuck"
 	NetworkStatus string                 `json:"network_status,omitempty"`
 	Issues        []RuntimeIssue         `json:"issues,omitempty"`
 	Tools         []RuntimeToolStatus    `json:"tools,omitempty"`

audit/internal/webui/api.go

@@ -11,6 +11,7 @@ import (
 	"os/exec"
 	"path/filepath"
 	"regexp"
+	"sort"
 	"strings"
 	"sync/atomic"
 	"syscall"
@@ -21,13 +22,305 @@ import (
 )
 
 var ansiEscapeRE = regexp.MustCompile(`\x1b\[[0-9;]*[a-zA-Z]|\x1b[()][A-Z0-9]|\x1b[DABC]`)
+var apiListNvidiaGPUs = func(a *app.App) ([]platform.NvidiaGPU, error) {
+	if a == nil {
+		return nil, fmt.Errorf("app not configured")
+	}
+	return a.ListNvidiaGPUs()
+}
+var apiListNvidiaGPUStatuses = func(a *app.App) ([]platform.NvidiaGPUStatus, error) {
+	if a == nil {
+		return nil, fmt.Errorf("app not configured")
+	}
+	return a.ListNvidiaGPUStatuses()
+}
 
 // ── Job ID counter ────────────────────────────────────────────────────────────
 
 var jobCounter atomic.Uint64
 
-func newJobID(prefix string) string {
-	return fmt.Sprintf("%s-%d", prefix, jobCounter.Add(1))
+func newJobID(_ string) string {
+	start := int((jobCounter.Add(1) - 1) % 1000)
+	globalQueue.mu.Lock()
+	defer globalQueue.mu.Unlock()
+	for offset := 0; offset < 1000; offset++ {
+		n := (start + offset) % 1000
+		id := fmt.Sprintf("TASK-%03d", n)
+		if !taskIDInUseLocked(id) {
+			return id
+		}
+	}
+	return fmt.Sprintf("TASK-%03d", start)
+}
+
+func taskIDInUseLocked(id string) bool {
+	for _, t := range globalQueue.tasks {
+		if t != nil && t.ID == id {
+			return true
+		}
+	}
+	return false
+}
+
+type taskRunResponse struct {
+	TaskID    string   `json:"task_id,omitempty"`
+	JobID     string   `json:"job_id,omitempty"`
+	TaskIDs   []string `json:"task_ids,omitempty"`
+	JobIDs    []string `json:"job_ids,omitempty"`
+	TaskCount int      `json:"task_count,omitempty"`
+}
+
+type nvidiaTaskSelection struct {
+	GPUIndices []int
+	Label      string
+}
+
+func writeTaskRunResponse(w http.ResponseWriter, tasks []*Task) {
+	if len(tasks) == 0 {
+		writeJSON(w, taskRunResponse{})
+		return
+	}
+	ids := make([]string, 0, len(tasks))
+	for _, t := range tasks {
+		if t == nil || strings.TrimSpace(t.ID) == "" {
+			continue
+		}
+		ids = append(ids, t.ID)
+	}
+	resp := taskRunResponse{TaskCount: len(ids)}
+	if len(ids) > 0 {
+		resp.TaskID = ids[0]
+		resp.JobID = ids[0]
+		resp.TaskIDs = ids
+		resp.JobIDs = ids
+	}
+	writeJSON(w, resp)
+}
+
+func shouldSplitHomogeneousNvidiaTarget(target string) bool {
+	switch strings.TrimSpace(target) {
+	case "nvidia", "nvidia-targeted-stress", "nvidia-benchmark", "nvidia-compute",
+		"nvidia-targeted-power", "nvidia-pulse", "nvidia-interconnect",
+		"nvidia-bandwidth", "nvidia-stress":
+		return true
+	default:
+		return false
+	}
+}
+
+func expandHomogeneousNvidiaSelections(gpus []platform.NvidiaGPU, include, exclude []int) ([]nvidiaTaskSelection, error) {
+	if len(gpus) == 0 {
+		return nil, fmt.Errorf("no NVIDIA GPUs detected")
+	}
+	indexed := make(map[int]platform.NvidiaGPU, len(gpus))
+	allIndices := make([]int, 0, len(gpus))
+	for _, gpu := range gpus {
+		indexed[gpu.Index] = gpu
+		allIndices = append(allIndices, gpu.Index)
+	}
+	sort.Ints(allIndices)
+
+	selected := allIndices
+	if len(include) > 0 {
+		selected = make([]int, 0, len(include))
+		seen := make(map[int]struct{}, len(include))
+		for _, idx := range include {
+			if _, ok := indexed[idx]; !ok {
+				continue
+			}
+			if _, dup := seen[idx]; dup {
+				continue
+			}
+			seen[idx] = struct{}{}
+			selected = append(selected, idx)
+		}
+		sort.Ints(selected)
+	}
+	if len(exclude) > 0 {
+		skip := make(map[int]struct{}, len(exclude))
+		for _, idx := range exclude {
+			skip[idx] = struct{}{}
+		}
+		filtered := selected[:0]
+		for _, idx := range selected {
+			if _, ok := skip[idx]; ok {
+				continue
+			}
+			filtered = append(filtered, idx)
+		}
+		selected = filtered
+	}
+	if len(selected) == 0 {
+		return nil, fmt.Errorf("no NVIDIA GPUs selected")
+	}
+
+	modelGroups := make(map[string][]platform.NvidiaGPU)
+	modelOrder := make([]string, 0)
+	for _, idx := range selected {
+		gpu := indexed[idx]
+		model := strings.TrimSpace(gpu.Name)
+		if model == "" {
+			model = fmt.Sprintf("GPU %d", gpu.Index)
+		}
+		if _, ok := modelGroups[model]; !ok {
+			modelOrder = append(modelOrder, model)
+		}
+		modelGroups[model] = append(modelGroups[model], gpu)
+	}
+	sort.Slice(modelOrder, func(i, j int) bool {
+		left := modelGroups[modelOrder[i]]
+		right := modelGroups[modelOrder[j]]
+		if len(left) == 0 || len(right) == 0 {
+			return modelOrder[i] < modelOrder[j]
+		}
+		return left[0].Index < right[0].Index
+	})
+
+	var groups []nvidiaTaskSelection
+	var singles []nvidiaTaskSelection
+	for _, model := range modelOrder {
+		group := modelGroups[model]
+		sort.Slice(group, func(i, j int) bool { return group[i].Index < group[j].Index })
+		indices := make([]int, 0, len(group))
+		for _, gpu := range group {
+			indices = append(indices, gpu.Index)
+		}
+		if len(indices) >= 2 {
+			groups = append(groups, nvidiaTaskSelection{
+				GPUIndices: indices,
+				Label:      fmt.Sprintf("%s; GPUs %s", model, joinTaskIndices(indices)),
+			})
+			continue
+		}
+		gpu := group[0]
+		singles = append(singles, nvidiaTaskSelection{
+			GPUIndices: []int{gpu.Index},
+			Label:      fmt.Sprintf("GPU %d — %s", gpu.Index, model),
+		})
+	}
+	return append(groups, singles...), nil
+}
+
+func joinTaskIndices(indices []int) string {
+	parts := make([]string, 0, len(indices))
+	for _, idx := range indices {
+		parts = append(parts, fmt.Sprintf("%d", idx))
+	}
+	return strings.Join(parts, ",")
+}
+
+func formatSplitTaskName(baseName, selectionLabel string) string {
+	baseName = strings.TrimSpace(baseName)
+	selectionLabel = strings.TrimSpace(selectionLabel)
+	if baseName == "" {
+		return selectionLabel
+	}
+	if selectionLabel == "" {
+		return baseName
+	}
+	return baseName + " (" + selectionLabel + ")"
+}
+
+func buildNvidiaTaskSet(target string, priority int, createdAt time.Time, params taskParams, baseName string, appRef *app.App, idPrefix string) ([]*Task, error) {
+	if !shouldSplitHomogeneousNvidiaTarget(target) || params.ParallelGPUs {
+		// Parallel mode (or non-splittable target): one task for all selected GPUs.
+		if params.ParallelGPUs && shouldSplitHomogeneousNvidiaTarget(target) {
+			// Resolve the selected GPU indices so ExcludeGPUIndices is applied.
+			gpus, err := apiListNvidiaGPUs(appRef)
+			if err != nil {
+				return nil, err
+			}
+			resolved, err := expandSelectedGPUIndices(gpus, params.GPUIndices, params.ExcludeGPUIndices)
+			if err != nil {
+				return nil, err
+			}
+			params.GPUIndices = resolved
+			params.ExcludeGPUIndices = nil
+		}
+		t := &Task{
+			ID:        newJobID(idPrefix),
+			Name:      baseName,
+			Target:    target,
+			Priority:  priority,
+			Status:    TaskPending,
+			CreatedAt: createdAt,
+			params:    params,
+		}
+		return []*Task{t}, nil
+	}
+	gpus, err := apiListNvidiaGPUs(appRef)
+	if err != nil {
+		return nil, err
+	}
+	selections, err := expandHomogeneousNvidiaSelections(gpus, params.GPUIndices, params.ExcludeGPUIndices)
+	if err != nil {
+		return nil, err
+	}
+	tasks := make([]*Task, 0, len(selections))
+	for _, selection := range selections {
+		taskParamsCopy := params
+		taskParamsCopy.GPUIndices = append([]int(nil), selection.GPUIndices...)
+		taskParamsCopy.ExcludeGPUIndices = nil
+		displayName := formatSplitTaskName(baseName, selection.Label)
+		taskParamsCopy.DisplayName = displayName
+		tasks = append(tasks, &Task{
+			ID:        newJobID(idPrefix),
+			Name:      displayName,
+			Target:    target,
+			Priority:  priority,
+			Status:    TaskPending,
+			CreatedAt: createdAt,
+			params:    taskParamsCopy,
+		})
+	}
+	return tasks, nil
+}
+
+// expandSelectedGPUIndices returns the sorted list of selected GPU indices after
+// applying include/exclude filters, without splitting by model.
+func expandSelectedGPUIndices(gpus []platform.NvidiaGPU, include, exclude []int) ([]int, error) {
+	indexed := make(map[int]struct{}, len(gpus))
+	allIndices := make([]int, 0, len(gpus))
+	for _, gpu := range gpus {
+		indexed[gpu.Index] = struct{}{}
+		allIndices = append(allIndices, gpu.Index)
+	}
+	sort.Ints(allIndices)
+
+	selected := allIndices
+	if len(include) > 0 {
+		selected = make([]int, 0, len(include))
+		seen := make(map[int]struct{}, len(include))
+		for _, idx := range include {
+			if _, ok := indexed[idx]; !ok {
+				continue
+			}
+			if _, dup := seen[idx]; dup {
+				continue
+			}
+			seen[idx] = struct{}{}
+			selected = append(selected, idx)
+		}
+		sort.Ints(selected)
+	}
+	if len(exclude) > 0 {
+		skip := make(map[int]struct{}, len(exclude))
+		for _, idx := range exclude {
+			skip[idx] = struct{}{}
+		}
+		filtered := selected[:0]
+		for _, idx := range selected {
+			if _, ok := skip[idx]; ok {
+				continue
+			}
+			filtered = append(filtered, idx)
+		}
+		selected = filtered
+	}
+	if len(selected) == 0 {
+		return nil, fmt.Errorf("no NVIDIA GPUs selected")
+	}
+	return selected, nil
 }
 
 // ── SSE helpers ───────────────────────────────────────────────────────────────
@@ -110,6 +403,11 @@ func streamCmdJob(j *jobState, cmd *exec.Cmd) error {
 
 	scanDone := make(chan error, 1)
 	go func() {
+		defer func() {
+			if rec := recover(); rec != nil {
+				scanDone <- fmt.Errorf("stream scanner panic: %v", rec)
+			}
+		}()
 		scanner := bufio.NewScanner(pr)
 		scanner.Buffer(make([]byte, 0, 64*1024), 1024*1024)
 		for scanner.Scan() {
@@ -184,12 +482,13 @@ func (h *handler) handleAPISATRun(target string) http.HandlerFunc {
 			return
 		}
 
-		var body struct {
-			Duration           int      `json:"duration"`
-			DiagLevel          int      `json:"diag_level"`
-			GPUIndices         []int    `json:"gpu_indices"`
-			ExcludeGPUIndices  []int    `json:"exclude_gpu_indices"`
-			Loader             string   `json:"loader"`
+			var body struct {
+				Duration           int      `json:"duration"`
+				StressMode         bool     `json:"stress_mode"`
+				GPUIndices         []int    `json:"gpu_indices"`
+				ExcludeGPUIndices  []int    `json:"exclude_gpu_indices"`
+				StaggerGPUStart    bool     `json:"stagger_gpu_start"`
+				Loader             string   `json:"loader"`
 			Profile            string   `json:"profile"`
 			DisplayName        string   `json:"display_name"`
 			PlatformComponents []string `json:"platform_components"`
@@ -202,31 +501,85 @@ func (h *handler) handleAPISATRun(target string) http.HandlerFunc {
 		}
 
 		name := taskDisplayName(target, body.Profile, body.Loader)
-		t := &Task{
-			ID:        newJobID("sat-" + target),
-			Name:      name,
-			Target:    target,
-			Status:    TaskPending,
-			CreatedAt: time.Now(),
-			params: taskParams{
+		if strings.TrimSpace(body.DisplayName) != "" {
+			name = body.DisplayName
+		}
+			params := taskParams{
 				Duration:           body.Duration,
-				DiagLevel:          body.DiagLevel,
+				StressMode:         body.StressMode,
 				GPUIndices:         body.GPUIndices,
 				ExcludeGPUIndices:  body.ExcludeGPUIndices,
+				StaggerGPUStart:    body.StaggerGPUStart,
 				Loader:             body.Loader,
-				BurnProfile:        body.Profile,
-				DisplayName:        body.DisplayName,
-				PlatformComponents: body.PlatformComponents,
-			},
+			BurnProfile:        body.Profile,
+			DisplayName:        body.DisplayName,
+			PlatformComponents: body.PlatformComponents,
 		}
-		if strings.TrimSpace(body.DisplayName) != "" {
-			t.Name = body.DisplayName
+		tasks, err := buildNvidiaTaskSet(target, 0, time.Now(), params, name, h.opts.App, "sat-"+target)
+		if err != nil {
+			writeError(w, http.StatusBadRequest, err.Error())
+			return
 		}
-		globalQueue.enqueue(t)
-		writeJSON(w, map[string]string{"task_id": t.ID, "job_id": t.ID})
+		for _, t := range tasks {
+			globalQueue.enqueue(t)
+		}
+		writeTaskRunResponse(w, tasks)
 	}
 }
 
+func (h *handler) handleAPIBenchmarkNvidiaRun(w http.ResponseWriter, r *http.Request) {
+	if h.opts.App == nil {
+		writeError(w, http.StatusServiceUnavailable, "app not configured")
+		return
+	}
+
+	var body struct {
+		Profile           string `json:"profile"`
+		SizeMB            int    `json:"size_mb"`
+		GPUIndices        []int  `json:"gpu_indices"`
+		ExcludeGPUIndices []int  `json:"exclude_gpu_indices"`
+		RunNCCL           *bool  `json:"run_nccl"`
+		ParallelGPUs      *bool  `json:"parallel_gpus"`
+		DisplayName       string `json:"display_name"`
+	}
+	if r.Body != nil {
+		if err := json.NewDecoder(r.Body).Decode(&body); err != nil && !errors.Is(err, io.EOF) {
+			writeError(w, http.StatusBadRequest, "invalid request body")
+			return
+		}
+	}
+
+	runNCCL := true
+	if body.RunNCCL != nil {
+		runNCCL = *body.RunNCCL
+	}
+	parallelGPUs := false
+	if body.ParallelGPUs != nil {
+		parallelGPUs = *body.ParallelGPUs
+	}
+	name := taskDisplayName("nvidia-benchmark", "", "")
+	if strings.TrimSpace(body.DisplayName) != "" {
+		name = body.DisplayName
+	}
+	tasks, err := buildNvidiaTaskSet("nvidia-benchmark", 15, time.Now(), taskParams{
+		GPUIndices:        body.GPUIndices,
+		ExcludeGPUIndices: body.ExcludeGPUIndices,
+		SizeMB:            body.SizeMB,
+		BenchmarkProfile:  body.Profile,
+		RunNCCL:           runNCCL,
+		ParallelGPUs:      parallelGPUs,
+		DisplayName:       body.DisplayName,
+	}, name, h.opts.App, "benchmark-nvidia")
+	if err != nil {
+		writeError(w, http.StatusBadRequest, err.Error())
+		return
+	}
+	for _, t := range tasks {
+		globalQueue.enqueue(t)
+	}
+	writeTaskRunResponse(w, tasks)
+}
+
 func (h *handler) handleAPISATStream(w http.ResponseWriter, r *http.Request) {
 	id := r.URL.Query().Get("job_id")
 	if id == "" {
@@ -330,11 +683,13 @@ func (h *handler) handleAPIServicesAction(w http.ResponseWriter, r *http.Request
 		return
 	}
 	result, err := h.opts.App.ServiceActionResult(req.Name, action)
+	status := "ok"
 	if err != nil {
-		writeError(w, http.StatusInternalServerError, err.Error())
-		return
+		status = "error"
 	}
-	writeJSON(w, map[string]string{"status": "ok", "output": result.Body})
+	// Always return 200 with output so the frontend can display the actual
+	// systemctl error message instead of a generic "exit status 1".
+	writeJSON(w, map[string]string{"status": status, "output": result.Body})
 }
 
 // ── Network ───────────────────────────────────────────────────────────────────
@@ -486,6 +841,58 @@ func (h *handler) handleAPIExportUSBBundle(w http.ResponseWriter, r *http.Reques
 
 // ── GPU presence ──────────────────────────────────────────────────────────────
 
+func (h *handler) handleAPIGNVIDIAGPUs(w http.ResponseWriter, _ *http.Request) {
+	if h.opts.App == nil {
+		writeError(w, http.StatusServiceUnavailable, "app not configured")
+		return
+	}
+	gpus, err := h.opts.App.ListNvidiaGPUs()
+	if err != nil {
+		writeError(w, http.StatusInternalServerError, err.Error())
+		return
+	}
+	if gpus == nil {
+		gpus = []platform.NvidiaGPU{}
+	}
+	writeJSON(w, gpus)
+}
+
+func (h *handler) handleAPIGNVIDIAGPUStatuses(w http.ResponseWriter, _ *http.Request) {
+	if h.opts.App == nil {
+		writeError(w, http.StatusServiceUnavailable, "app not configured")
+		return
+	}
+	gpus, err := apiListNvidiaGPUStatuses(h.opts.App)
+	if err != nil {
+		writeError(w, http.StatusInternalServerError, err.Error())
+		return
+	}
+	if gpus == nil {
+		gpus = []platform.NvidiaGPUStatus{}
+	}
+	writeJSON(w, gpus)
+}
+
+func (h *handler) handleAPIGNVIDIAReset(w http.ResponseWriter, r *http.Request) {
+	if h.opts.App == nil {
+		writeError(w, http.StatusServiceUnavailable, "app not configured")
+		return
+	}
+	var req struct {
+		Index int `json:"index"`
+	}
+	if err := json.NewDecoder(r.Body).Decode(&req); err != nil {
+		writeError(w, http.StatusBadRequest, "invalid request body")
+		return
+	}
+	result, err := h.opts.App.ResetNvidiaGPU(req.Index)
+	status := "ok"
+	if err != nil {
+		status = "error"
+	}
+	writeJSON(w, map[string]string{"status": status, "output": result.Body})
+}
+
 func (h *handler) handleAPIGPUPresence(w http.ResponseWriter, r *http.Request) {
 	if h.opts.App == nil {
 		writeError(w, http.StatusServiceUnavailable, "app not configured")
@@ -511,14 +918,33 @@ func (h *handler) handleAPIGPUTools(w http.ResponseWriter, _ *http.Request) {
 	_, amdErr := os.Stat("/dev/kfd")
 	nvidiaUp := nvidiaErr == nil
 	amdUp := amdErr == nil
+	_, dcgmErr := exec.LookPath("dcgmi")
+	_, ncclStressErr := exec.LookPath("bee-nccl-gpu-stress")
+	_, johnErr := exec.LookPath("bee-john-gpu-stress")
+	_, beeBurnErr := exec.LookPath("bee-gpu-burn")
+	_, nvBandwidthErr := exec.LookPath("nvbandwidth")
+	profErr := lookPathAny("dcgmproftester", "dcgmproftester13", "dcgmproftester12", "dcgmproftester11")
 	writeJSON(w, []toolEntry{
-		{ID: "bee-gpu-burn", Available: nvidiaUp, Vendor: "nvidia"},
-		{ID: "john", Available: nvidiaUp, Vendor: "nvidia"},
-		{ID: "nccl", Available: nvidiaUp, Vendor: "nvidia"},
+		{ID: "nvidia-compute", Available: nvidiaUp && profErr == nil, Vendor: "nvidia"},
+		{ID: "nvidia-targeted-power", Available: nvidiaUp && dcgmErr == nil, Vendor: "nvidia"},
+		{ID: "nvidia-pulse", Available: nvidiaUp && dcgmErr == nil, Vendor: "nvidia"},
+		{ID: "nvidia-interconnect", Available: nvidiaUp && ncclStressErr == nil, Vendor: "nvidia"},
+		{ID: "nvidia-bandwidth", Available: nvidiaUp && dcgmErr == nil && nvBandwidthErr == nil, Vendor: "nvidia"},
+		{ID: "bee-gpu-burn", Available: nvidiaUp && beeBurnErr == nil, Vendor: "nvidia"},
+		{ID: "john", Available: nvidiaUp && johnErr == nil, Vendor: "nvidia"},
 		{ID: "rvs", Available: amdUp, Vendor: "amd"},
 	})
 }
 
+func lookPathAny(names ...string) error {
+	for _, name := range names {
+		if _, err := exec.LookPath(name); err == nil {
+			return nil
+		}
+	}
+	return exec.ErrNotFound
+}
+
 // ── System ────────────────────────────────────────────────────────────────────
 
 func (h *handler) handleAPIRAMStatus(w http.ResponseWriter, r *http.Request) {
@@ -557,7 +983,7 @@ func (h *handler) handleAPIInstallToRAM(w http.ResponseWriter, r *http.Request)
 
 var standardTools = []string{
 	"dmidecode", "smartctl", "nvme", "lspci", "ipmitool",
-	"nvidia-smi", "memtester", "stress-ng", "nvtop",
+	"nvidia-smi", "dcgmi", "nv-hostengine", "memtester", "stress-ng", "nvtop",
 	"mstflint", "qrencode",
 }
 
@@ -952,107 +1378,3 @@ func (h *handler) rollbackPendingNetworkChange() error {
 	return nil
 }
 
-// ── Display / Screen Resolution ───────────────────────────────────────────────
-
-type displayMode struct {
-	Output  string `json:"output"`
-	Mode    string `json:"mode"`
-	Current bool   `json:"current"`
-}
-
-type displayInfo struct {
-	Output  string        `json:"output"`
-	Modes   []displayMode `json:"modes"`
-	Current string        `json:"current"`
-}
-
-var xrandrOutputRE = regexp.MustCompile(`^(\S+)\s+connected`)
-var xrandrModeRE = regexp.MustCompile(`^\s{3}(\d+x\d+)\s`)
-var xrandrCurrentRE = regexp.MustCompile(`\*`)
-
-func parseXrandrOutput(out string) []displayInfo {
-	var infos []displayInfo
-	var cur *displayInfo
-	for _, line := range strings.Split(out, "\n") {
-		if m := xrandrOutputRE.FindStringSubmatch(line); m != nil {
-			if cur != nil {
-				infos = append(infos, *cur)
-			}
-			cur = &displayInfo{Output: m[1]}
-			continue
-		}
-		if cur == nil {
-			continue
-		}
-		if m := xrandrModeRE.FindStringSubmatch(line); m != nil {
-			isCurrent := xrandrCurrentRE.MatchString(line)
-			mode := displayMode{Output: cur.Output, Mode: m[1], Current: isCurrent}
-			cur.Modes = append(cur.Modes, mode)
-			if isCurrent {
-				cur.Current = m[1]
-			}
-		}
-	}
-	if cur != nil {
-		infos = append(infos, *cur)
-	}
-	return infos
-}
-
-func xrandrCommand(args ...string) *exec.Cmd {
-	cmd := exec.Command("xrandr", args...)
-	env := append([]string{}, os.Environ()...)
-	hasDisplay := false
-	hasXAuthority := false
-	for _, kv := range env {
-		if strings.HasPrefix(kv, "DISPLAY=") && strings.TrimPrefix(kv, "DISPLAY=") != "" {
-			hasDisplay = true
-		}
-		if strings.HasPrefix(kv, "XAUTHORITY=") && strings.TrimPrefix(kv, "XAUTHORITY=") != "" {
-			hasXAuthority = true
-		}
-	}
-	if !hasDisplay {
-		env = append(env, "DISPLAY=:0")
-	}
-	if !hasXAuthority {
-		env = append(env, "XAUTHORITY=/home/bee/.Xauthority")
-	}
-	cmd.Env = env
-	return cmd
-}
-
-func (h *handler) handleAPIDisplayResolutions(w http.ResponseWriter, _ *http.Request) {
-	out, err := xrandrCommand().Output()
-	if err != nil {
-		writeError(w, http.StatusInternalServerError, "xrandr: "+err.Error())
-		return
-	}
-	writeJSON(w, parseXrandrOutput(string(out)))
-}
-
-func (h *handler) handleAPIDisplaySet(w http.ResponseWriter, r *http.Request) {
-	var req struct {
-		Output string `json:"output"`
-		Mode   string `json:"mode"`
-	}
-	if err := json.NewDecoder(r.Body).Decode(&req); err != nil || req.Output == "" || req.Mode == "" {
-		writeError(w, http.StatusBadRequest, "output and mode are required")
-		return
-	}
-	// Validate mode looks like WxH to prevent injection
-	if !regexp.MustCompile(`^\d+x\d+$`).MatchString(req.Mode) {
-		writeError(w, http.StatusBadRequest, "invalid mode format")
-		return
-	}
-	// Validate output name (no special chars)
-	if !regexp.MustCompile(`^[A-Za-z0-9_\-]+$`).MatchString(req.Output) {
-		writeError(w, http.StatusBadRequest, "invalid output name")
-		return
-	}
-	if out, err := xrandrCommand("--output", req.Output, "--mode", req.Mode).CombinedOutput(); err != nil {
-		writeError(w, http.StatusInternalServerError, "xrandr: "+strings.TrimSpace(string(out)))
-		return
-	}
-	writeJSON(w, map[string]string{"status": "ok", "output": req.Output, "mode": req.Mode})
-}

audit/internal/webui/api_test.go

@@ -1,6 +1,7 @@
 package webui
 
 import (
+	"encoding/json"
 	"net/http/httptest"
 	"strings"
 	"testing"
@@ -9,30 +10,6 @@ import (
 	"bee/audit/internal/platform"
 )
 
-func TestXrandrCommandAddsDefaultX11Env(t *testing.T) {
-	t.Setenv("DISPLAY", "")
-	t.Setenv("XAUTHORITY", "")
-
-	cmd := xrandrCommand("--query")
-
-	var hasDisplay bool
-	var hasXAuthority bool
-	for _, kv := range cmd.Env {
-		if kv == "DISPLAY=:0" {
-			hasDisplay = true
-		}
-		if kv == "XAUTHORITY=/home/bee/.Xauthority" {
-			hasXAuthority = true
-		}
-	}
-	if !hasDisplay {
-		t.Fatalf("DISPLAY not injected: %v", cmd.Env)
-	}
-	if !hasXAuthority {
-		t.Fatalf("XAUTHORITY not injected: %v", cmd.Env)
-	}
-}
-
 func TestHandleAPISATRunDecodesBodyWithoutContentLength(t *testing.T) {
 	globalQueue.mu.Lock()
 	originalTasks := globalQueue.tasks
@@ -64,6 +41,141 @@ func TestHandleAPISATRunDecodesBodyWithoutContentLength(t *testing.T) {
 	}
 }
 
+func TestHandleAPIBenchmarkNvidiaRunQueuesSelectedGPUs(t *testing.T) {
+	globalQueue.mu.Lock()
+	originalTasks := globalQueue.tasks
+	globalQueue.tasks = nil
+	globalQueue.mu.Unlock()
+	t.Cleanup(func() {
+		globalQueue.mu.Lock()
+		globalQueue.tasks = originalTasks
+		globalQueue.mu.Unlock()
+	})
+	prevList := apiListNvidiaGPUs
+	apiListNvidiaGPUs = func(_ *app.App) ([]platform.NvidiaGPU, error) {
+		return []platform.NvidiaGPU{
+			{Index: 1, Name: "NVIDIA H100 PCIe"},
+			{Index: 3, Name: "NVIDIA H100 PCIe"},
+		}, nil
+	}
+	t.Cleanup(func() { apiListNvidiaGPUs = prevList })
+
+	h := &handler{opts: HandlerOptions{App: &app.App{}}}
+	req := httptest.NewRequest("POST", "/api/benchmark/nvidia/run", strings.NewReader(`{"profile":"standard","gpu_indices":[1,3],"run_nccl":false}`))
+	rec := httptest.NewRecorder()
+
+	h.handleAPIBenchmarkNvidiaRun(rec, req)
+
+	if rec.Code != 200 {
+		t.Fatalf("status=%d body=%s", rec.Code, rec.Body.String())
+	}
+	globalQueue.mu.Lock()
+	defer globalQueue.mu.Unlock()
+	if len(globalQueue.tasks) != 1 {
+		t.Fatalf("tasks=%d want 1", len(globalQueue.tasks))
+	}
+	task := globalQueue.tasks[0]
+	if task.Target != "nvidia-benchmark" {
+		t.Fatalf("target=%q want nvidia-benchmark", task.Target)
+	}
+	if got := task.params.GPUIndices; len(got) != 2 || got[0] != 1 || got[1] != 3 {
+		t.Fatalf("gpu indices=%v want [1 3]", got)
+	}
+	if task.params.RunNCCL {
+		t.Fatal("RunNCCL should reflect explicit false from request")
+	}
+}
+
+func TestHandleAPIBenchmarkNvidiaRunSplitsMixedGPUModels(t *testing.T) {
+	globalQueue.mu.Lock()
+	originalTasks := globalQueue.tasks
+	globalQueue.tasks = nil
+	globalQueue.mu.Unlock()
+	t.Cleanup(func() {
+		globalQueue.mu.Lock()
+		globalQueue.tasks = originalTasks
+		globalQueue.mu.Unlock()
+	})
+	prevList := apiListNvidiaGPUs
+	apiListNvidiaGPUs = func(_ *app.App) ([]platform.NvidiaGPU, error) {
+		return []platform.NvidiaGPU{
+			{Index: 0, Name: "NVIDIA H100 PCIe"},
+			{Index: 1, Name: "NVIDIA H100 PCIe"},
+			{Index: 2, Name: "NVIDIA H200 NVL"},
+		}, nil
+	}
+	t.Cleanup(func() { apiListNvidiaGPUs = prevList })
+
+	h := &handler{opts: HandlerOptions{App: &app.App{}}}
+	req := httptest.NewRequest("POST", "/api/benchmark/nvidia/run", strings.NewReader(`{"profile":"standard","gpu_indices":[0,1,2],"run_nccl":false}`))
+	rec := httptest.NewRecorder()
+
+	h.handleAPIBenchmarkNvidiaRun(rec, req)
+
+	if rec.Code != 200 {
+		t.Fatalf("status=%d body=%s", rec.Code, rec.Body.String())
+	}
+	var resp taskRunResponse
+	if err := json.Unmarshal(rec.Body.Bytes(), &resp); err != nil {
+		t.Fatalf("decode response: %v", err)
+	}
+	if len(resp.TaskIDs) != 2 {
+		t.Fatalf("task_ids=%v want 2 items", resp.TaskIDs)
+	}
+	globalQueue.mu.Lock()
+	defer globalQueue.mu.Unlock()
+	if len(globalQueue.tasks) != 2 {
+		t.Fatalf("tasks=%d want 2", len(globalQueue.tasks))
+	}
+	if got := globalQueue.tasks[0].params.GPUIndices; len(got) != 2 || got[0] != 0 || got[1] != 1 {
+		t.Fatalf("task[0] gpu indices=%v want [0 1]", got)
+	}
+	if got := globalQueue.tasks[1].params.GPUIndices; len(got) != 1 || got[0] != 2 {
+		t.Fatalf("task[1] gpu indices=%v want [2]", got)
+	}
+}
+
+func TestHandleAPISATRunSplitsMixedNvidiaTaskSet(t *testing.T) {
+	globalQueue.mu.Lock()
+	originalTasks := globalQueue.tasks
+	globalQueue.tasks = nil
+	globalQueue.mu.Unlock()
+	t.Cleanup(func() {
+		globalQueue.mu.Lock()
+		globalQueue.tasks = originalTasks
+		globalQueue.mu.Unlock()
+	})
+	prevList := apiListNvidiaGPUs
+	apiListNvidiaGPUs = func(_ *app.App) ([]platform.NvidiaGPU, error) {
+		return []platform.NvidiaGPU{
+			{Index: 0, Name: "NVIDIA H100 PCIe"},
+			{Index: 1, Name: "NVIDIA H100 PCIe"},
+			{Index: 2, Name: "NVIDIA H200 NVL"},
+		}, nil
+	}
+	t.Cleanup(func() { apiListNvidiaGPUs = prevList })
+
+	h := &handler{opts: HandlerOptions{App: &app.App{}}}
+	req := httptest.NewRequest("POST", "/api/sat/nvidia-targeted-power/run", strings.NewReader(`{"profile":"acceptance","gpu_indices":[0,1,2]}`))
+	rec := httptest.NewRecorder()
+
+	h.handleAPISATRun("nvidia-targeted-power").ServeHTTP(rec, req)
+
+	if rec.Code != 200 {
+		t.Fatalf("status=%d body=%s", rec.Code, rec.Body.String())
+	}
+	globalQueue.mu.Lock()
+	defer globalQueue.mu.Unlock()
+	if len(globalQueue.tasks) != 2 {
+		t.Fatalf("tasks=%d want 2", len(globalQueue.tasks))
+	}
+	if got := globalQueue.tasks[0].params.GPUIndices; len(got) != 2 || got[0] != 0 || got[1] != 1 {
+		t.Fatalf("task[0] gpu indices=%v want [0 1]", got)
+	}
+	if got := globalQueue.tasks[1].params.GPUIndices; len(got) != 1 || got[0] != 2 {
+		t.Fatalf("task[1] gpu indices=%v want [2]", got)
+	}
+}
 
 func TestPushFanRingsTracksByNameAndCarriesForwardMissingSamples(t *testing.T) {
 	h := &handler{}

audit/internal/webui/charts_svg.go

@@ -0,0 +1,773 @@
+package webui
+
+import (
+	"fmt"
+	"math"
+	"sort"
+	"strconv"
+	"strings"
+	"sync"
+	"time"
+
+	"bee/audit/internal/platform"
+)
+
+type chartTimelineSegment struct {
+	Start  time.Time
+	End    time.Time
+	Active bool
+}
+
+type chartScale struct {
+	Min   float64
+	Max   float64
+	Ticks []float64
+}
+
+type chartLayout struct {
+	Width      int
+	Height     int
+	PlotLeft   int
+	PlotRight  int
+	PlotTop    int
+	PlotBottom int
+}
+
+type metricChartSeries struct {
+	Name      string
+	AxisTitle string
+	Color     string
+	Values    []float64
+}
+
+var metricChartPalette = []string{
+	"#5794f2",
+	"#73bf69",
+	"#f2cc0c",
+	"#ff9830",
+	"#f2495c",
+	"#b877d9",
+	"#56d2f7",
+	"#8ab8ff",
+	"#9adf8f",
+	"#ffbe5c",
+}
+
+var gpuLabelCache struct {
+	mu       sync.Mutex
+	loadedAt time.Time
+	byIndex  map[int]string
+}
+
+func renderMetricChartSVG(title string, labels []string, times []time.Time, datasets [][]float64, names []string, yMin, yMax *float64, canvasHeight int, timeline []chartTimelineSegment) ([]byte, error) {
+	pointCount := len(labels)
+	if len(times) > pointCount {
+		pointCount = len(times)
+	}
+	if pointCount == 0 {
+		pointCount = 1
+		labels = []string{""}
+		times = []time.Time{time.Time{}}
+	}
+	if len(labels) < pointCount {
+		padded := make([]string, pointCount)
+		copy(padded, labels)
+		labels = padded
+	}
+	if len(times) < pointCount {
+		times = synthesizeChartTimes(times, pointCount)
+	}
+	for i := range datasets {
+		if len(datasets[i]) == 0 {
+			datasets[i] = make([]float64, pointCount)
+		}
+	}
+
+	statsLabel := chartStatsLabel(datasets)
+
+	legendItems := []metricChartSeries{}
+	for i, name := range names {
+		color := metricChartPalette[i%len(metricChartPalette)]
+		values := make([]float64, pointCount)
+		if i < len(datasets) {
+			copy(values, coalesceDataset(datasets[i], pointCount))
+		}
+		legendItems = append(legendItems, metricChartSeries{
+			Name:   name,
+			Color:  color,
+			Values: values,
+		})
+	}
+
+	scale := singleAxisChartScale(datasets, yMin, yMax)
+	layout := singleAxisChartLayout(canvasHeight, len(legendItems))
+	start, end := chartTimeBounds(times)
+
+	var b strings.Builder
+	writeSVGOpen(&b, layout.Width, layout.Height)
+	writeChartFrame(&b, title, statsLabel, layout.Width, layout.Height)
+	writeTimelineIdleSpans(&b, layout, start, end, timeline)
+	writeVerticalGrid(&b, layout, times, pointCount, 8)
+	writeHorizontalGrid(&b, layout, scale)
+	writeTimelineBoundaries(&b, layout, start, end, timeline)
+	writePlotBorder(&b, layout)
+	writeSingleAxisY(&b, layout, scale)
+	writeXAxisLabels(&b, layout, times, labels, start, end, 8)
+	for _, item := range legendItems {
+		writeSeriesPolyline(&b, layout, times, start, end, item.Values, scale, item.Color)
+	}
+	writeLegend(&b, layout, legendItems)
+	writeSVGClose(&b)
+	return []byte(b.String()), nil
+}
+
+func renderGPUOverviewChartSVG(idx int, samples []platform.LiveMetricSample, timeline []chartTimelineSegment) ([]byte, bool, error) {
+	temp := gpuDatasetByIndex(samples, idx, func(g platform.GPUMetricRow) float64 { return g.TempC })
+	power := gpuDatasetByIndex(samples, idx, func(g platform.GPUMetricRow) float64 { return g.PowerW })
+	coreClock := gpuDatasetByIndex(samples, idx, func(g platform.GPUMetricRow) float64 { return g.ClockMHz })
+	if temp == nil && power == nil && coreClock == nil {
+		return nil, false, nil
+	}
+	labels := sampleTimeLabels(samples)
+	times := sampleTimes(samples)
+	svg, err := drawGPUOverviewChartSVG(
+		gpuDisplayLabel(idx)+" Overview",
+		labels,
+		times,
+		[]metricChartSeries{
+			{Name: "Temp C", Values: coalesceDataset(temp, len(labels)), Color: "#f05a5a", AxisTitle: "Temp C"},
+			{Name: "Power W", Values: coalesceDataset(power, len(labels)), Color: "#ffb357", AxisTitle: "Power W"},
+			{Name: "Core Clock MHz", Values: coalesceDataset(coreClock, len(labels)), Color: "#73bf69", AxisTitle: "Core MHz"},
+		},
+		timeline,
+	)
+	if err != nil {
+		return nil, false, err
+	}
+	return svg, true, nil
+}
+
+func drawGPUOverviewChartSVG(title string, labels []string, times []time.Time, series []metricChartSeries, timeline []chartTimelineSegment) ([]byte, error) {
+	if len(series) != 3 {
+		return nil, fmt.Errorf("gpu overview requires 3 series, got %d", len(series))
+	}
+	const (
+		width      = 1400
+		height     = 840
+		plotLeft   = 180
+		plotRight  = 1220
+		plotTop    = 96
+		plotBottom = 660
+	)
+	const (
+		leftOuterAxis  = 72
+		leftInnerAxis  = 132
+		rightInnerAxis = 1268
+	)
+	layout := chartLayout{
+		Width:      width,
+		Height:     height,
+		PlotLeft:   plotLeft,
+		PlotRight:  plotRight,
+		PlotTop:    plotTop,
+		PlotBottom: plotBottom,
+	}
+	axisX := []int{leftOuterAxis, leftInnerAxis, rightInnerAxis}
+	pointCount := len(labels)
+	if len(times) > pointCount {
+		pointCount = len(times)
+	}
+	if pointCount == 0 {
+		pointCount = 1
+		labels = []string{""}
+		times = []time.Time{time.Time{}}
+	}
+	if len(labels) < pointCount {
+		padded := make([]string, pointCount)
+		copy(padded, labels)
+		labels = padded
+	}
+	if len(times) < pointCount {
+		times = synthesizeChartTimes(times, pointCount)
+	}
+	for i := range series {
+		if len(series[i].Values) == 0 {
+			series[i].Values = make([]float64, pointCount)
+		}
+	}
+
+	scales := make([]chartScale, len(series))
+	for i := range series {
+		min, max := chartSeriesBounds(series[i].Values)
+		ticks := chartNiceTicks(min, max, 8)
+		scales[i] = chartScale{
+			Min:   ticks[0],
+			Max:   ticks[len(ticks)-1],
+			Ticks: ticks,
+		}
+	}
+	start, end := chartTimeBounds(times)
+
+	var b strings.Builder
+	writeSVGOpen(&b, width, height)
+	writeChartFrame(&b, title, "", width, height)
+	writeTimelineIdleSpans(&b, layout, start, end, timeline)
+	writeVerticalGrid(&b, layout, times, pointCount, 8)
+	writeHorizontalGrid(&b, layout, scales[0])
+	writeTimelineBoundaries(&b, layout, start, end, timeline)
+	writePlotBorder(&b, layout)
+
+	for i, axisLineX := range axisX {
+		fmt.Fprintf(&b, `<line x1="%d" y1="%d" x2="%d" y2="%d" stroke="%s" stroke-width="1"/>`+"\n",
+			axisLineX, layout.PlotTop, axisLineX, layout.PlotBottom, series[i].Color)
+		fmt.Fprintf(&b, `<text x="%d" y="%d" text-anchor="middle" font-family="sans-serif" font-size="11" font-weight="700" fill="%s">%s</text>`+"\n",
+			axisLineX, 64, series[i].Color, sanitizeChartText(series[i].AxisTitle))
+		for _, tick := range scales[i].Ticks {
+			y := chartYForValue(valueClamp(tick, scales[i]), scales[i], layout.PlotTop, layout.PlotBottom)
+			label := sanitizeChartText(chartYAxisNumber(tick))
+			if i < 2 {
+				fmt.Fprintf(&b, `<line x1="%d" y1="%.1f" x2="%d" y2="%.1f" stroke="%s" stroke-width="1"/>`+"\n",
+					axisLineX, y, axisLineX+6, y, series[i].Color)
+				fmt.Fprintf(&b, `<text x="%d" y="%.1f" text-anchor="end" dy="4" font-family="sans-serif" font-size="10" fill="%s">%s</text>`+"\n",
+					axisLineX-8, y, series[i].Color, label)
+				continue
+			}
+			fmt.Fprintf(&b, `<line x1="%d" y1="%.1f" x2="%d" y2="%.1f" stroke="%s" stroke-width="1"/>`+"\n",
+				axisLineX, y, axisLineX-6, y, series[i].Color)
+			fmt.Fprintf(&b, `<text x="%d" y="%.1f" text-anchor="start" dy="4" font-family="sans-serif" font-size="10" fill="%s">%s</text>`+"\n",
+				axisLineX+8, y, series[i].Color, label)
+		}
+	}
+
+	writeXAxisLabels(&b, layout, times, labels, start, end, 8)
+	for i := range series {
+		writeSeriesPolyline(&b, layout, times, start, end, series[i].Values, scales[i], series[i].Color)
+	}
+	writeLegend(&b, layout, series)
+	writeSVGClose(&b)
+	return []byte(b.String()), nil
+}
+
+func metricsTimelineSegments(samples []platform.LiveMetricSample, now time.Time) []chartTimelineSegment {
+	if len(samples) == 0 {
+		return nil
+	}
+	times := sampleTimes(samples)
+	start, end := chartTimeBounds(times)
+	if start.IsZero() || end.IsZero() {
+		return nil
+	}
+	return chartTimelineSegmentsForRange(start, end, now, snapshotTaskHistory())
+}
+
+func snapshotTaskHistory() []Task {
+	globalQueue.mu.Lock()
+	defer globalQueue.mu.Unlock()
+	out := make([]Task, len(globalQueue.tasks))
+	for i, t := range globalQueue.tasks {
+		out[i] = *t
+	}
+	return out
+}
+
+func chartTimelineSegmentsForRange(start, end, now time.Time, tasks []Task) []chartTimelineSegment {
+	if start.IsZero() || end.IsZero() {
+		return nil
+	}
+	if end.Before(start) {
+		start, end = end, start
+	}
+	type interval struct {
+		start time.Time
+		end   time.Time
+	}
+	active := make([]interval, 0, len(tasks))
+	for _, task := range tasks {
+		if task.StartedAt == nil {
+			continue
+		}
+		intervalStart := task.StartedAt.UTC()
+		intervalEnd := now.UTC()
+		if task.DoneAt != nil {
+			intervalEnd = task.DoneAt.UTC()
+		}
+		if !intervalEnd.After(intervalStart) {
+			continue
+		}
+		if intervalEnd.Before(start) || intervalStart.After(end) {
+			continue
+		}
+		if intervalStart.Before(start) {
+			intervalStart = start
+		}
+		if intervalEnd.After(end) {
+			intervalEnd = end
+		}
+		active = append(active, interval{start: intervalStart, end: intervalEnd})
+	}
+	sort.Slice(active, func(i, j int) bool {
+		if active[i].start.Equal(active[j].start) {
+			return active[i].end.Before(active[j].end)
+		}
+		return active[i].start.Before(active[j].start)
+	})
+	merged := make([]interval, 0, len(active))
+	for _, span := range active {
+		if len(merged) == 0 {
+			merged = append(merged, span)
+			continue
+		}
+		last := &merged[len(merged)-1]
+		if !span.start.After(last.end) {
+			if span.end.After(last.end) {
+				last.end = span.end
+			}
+			continue
+		}
+		merged = append(merged, span)
+	}
+
+	segments := make([]chartTimelineSegment, 0, len(merged)*2+1)
+	cursor := start
+	for _, span := range merged {
+		if span.start.After(cursor) {
+			segments = append(segments, chartTimelineSegment{Start: cursor, End: span.start, Active: false})
+		}
+		segments = append(segments, chartTimelineSegment{Start: span.start, End: span.end, Active: true})
+		cursor = span.end
+	}
+	if cursor.Before(end) {
+		segments = append(segments, chartTimelineSegment{Start: cursor, End: end, Active: false})
+	}
+	if len(segments) == 0 {
+		segments = append(segments, chartTimelineSegment{Start: start, End: end, Active: false})
+	}
+	return segments
+}
+
+func sampleTimes(samples []platform.LiveMetricSample) []time.Time {
+	times := make([]time.Time, 0, len(samples))
+	for _, sample := range samples {
+		times = append(times, sample.Timestamp)
+	}
+	return times
+}
+
+func singleAxisChartScale(datasets [][]float64, yMin, yMax *float64) chartScale {
+	min, max := 0.0, 1.0
+	if yMin != nil && yMax != nil {
+		min, max = *yMin, *yMax
+	} else {
+		min, max = chartSeriesBounds(flattenDatasets(datasets))
+		if yMin != nil {
+			min = *yMin
+		}
+		if yMax != nil {
+			max = *yMax
+		}
+	}
+	ticks := chartNiceTicks(min, max, 8)
+	return chartScale{Min: ticks[0], Max: ticks[len(ticks)-1], Ticks: ticks}
+}
+
+func flattenDatasets(datasets [][]float64) []float64 {
+	total := 0
+	for _, ds := range datasets {
+		total += len(ds)
+	}
+	out := make([]float64, 0, total)
+	for _, ds := range datasets {
+		out = append(out, ds...)
+	}
+	return out
+}
+
+func singleAxisChartLayout(canvasHeight int, seriesCount int) chartLayout {
+	legendRows := 0
+	if chartLegendVisible(seriesCount) && seriesCount > 0 {
+		cols := 4
+		if seriesCount < cols {
+			cols = seriesCount
+		}
+		legendRows = (seriesCount + cols - 1) / cols
+	}
+	legendHeight := 0
+	if legendRows > 0 {
+		legendHeight = legendRows*24 + 24
+	}
+	return chartLayout{
+		Width:      1400,
+		Height:     canvasHeight,
+		PlotLeft:   96,
+		PlotRight:  1352,
+		PlotTop:    72,
+		PlotBottom: canvasHeight - 60 - legendHeight,
+	}
+}
+
+func chartTimeBounds(times []time.Time) (time.Time, time.Time) {
+	if len(times) == 0 {
+		return time.Time{}, time.Time{}
+	}
+	start := times[0].UTC()
+	end := start
+	for _, ts := range times[1:] {
+		t := ts.UTC()
+		if t.Before(start) {
+			start = t
+		}
+		if t.After(end) {
+			end = t
+		}
+	}
+	return start, end
+}
+
+func synthesizeChartTimes(times []time.Time, count int) []time.Time {
+	if count <= 0 {
+		return nil
+	}
+	if len(times) == count {
+		return times
+	}
+	if len(times) == 1 {
+		out := make([]time.Time, count)
+		for i := range out {
+			out[i] = times[0].Add(time.Duration(i) * time.Minute)
+		}
+		return out
+	}
+	base := time.Now().UTC().Add(-time.Duration(count-1) * time.Minute)
+	out := make([]time.Time, count)
+	for i := range out {
+		out[i] = base.Add(time.Duration(i) * time.Minute)
+	}
+	return out
+}
+
+func writeSVGOpen(b *strings.Builder, width, height int) {
+	fmt.Fprintf(b, `<svg xmlns="http://www.w3.org/2000/svg" width="%d" height="%d" viewBox="0 0 %d %d">`+"\n", width, height, width, height)
+}
+
+func writeSVGClose(b *strings.Builder) {
+	b.WriteString("</svg>\n")
+}
+
+func writeChartFrame(b *strings.Builder, title, subtitle string, width, height int) {
+	fmt.Fprintf(b, `<rect width="%d" height="%d" rx="10" ry="10" fill="#ffffff" stroke="#d7e0ea"/>`+"\n", width, height)
+	fmt.Fprintf(b, `<text x="%d" y="30" text-anchor="middle" font-family="sans-serif" font-size="16" font-weight="700" fill="#1f2937">%s</text>`+"\n",
+		width/2, sanitizeChartText(title))
+	if strings.TrimSpace(subtitle) != "" {
+		fmt.Fprintf(b, `<text x="%d" y="50" text-anchor="middle" font-family="sans-serif" font-size="12" font-weight="600" fill="#64748b">%s</text>`+"\n",
+			width/2, sanitizeChartText(subtitle))
+	}
+}
+
+func writePlotBorder(b *strings.Builder, layout chartLayout) {
+	fmt.Fprintf(b, `<rect x="%d" y="%d" width="%d" height="%d" fill="none" stroke="#cbd5e1" stroke-width="1"/>`+"\n",
+		layout.PlotLeft, layout.PlotTop, layout.PlotRight-layout.PlotLeft, layout.PlotBottom-layout.PlotTop)
+}
+
+func writeHorizontalGrid(b *strings.Builder, layout chartLayout, scale chartScale) {
+	b.WriteString(`<g stroke="#e2e8f0" stroke-width="1">` + "\n")
+	for _, tick := range scale.Ticks {
+		y := chartYForValue(tick, scale, layout.PlotTop, layout.PlotBottom)
+		fmt.Fprintf(b, `<line x1="%d" y1="%.1f" x2="%d" y2="%.1f"/>`+"\n",
+			layout.PlotLeft, y, layout.PlotRight, y)
+	}
+	b.WriteString(`</g>` + "\n")
+}
+
+func writeVerticalGrid(b *strings.Builder, layout chartLayout, times []time.Time, pointCount, target int) {
+	if pointCount <= 0 {
+		return
+	}
+	start, end := chartTimeBounds(times)
+	b.WriteString(`<g stroke="#edf2f7" stroke-width="1">` + "\n")
+	for _, idx := range gpuChartLabelIndices(pointCount, target) {
+		ts := chartPointTime(times, idx)
+		x := chartXForTime(ts, start, end, layout.PlotLeft, layout.PlotRight)
+		fmt.Fprintf(b, `<line x1="%.1f" y1="%d" x2="%.1f" y2="%d"/>`+"\n",
+			x, layout.PlotTop, x, layout.PlotBottom)
+	}
+	b.WriteString(`</g>` + "\n")
+}
+
+func writeSingleAxisY(b *strings.Builder, layout chartLayout, scale chartScale) {
+	fmt.Fprintf(b, `<line x1="%d" y1="%d" x2="%d" y2="%d" stroke="#64748b" stroke-width="1"/>`+"\n",
+		layout.PlotLeft, layout.PlotTop, layout.PlotLeft, layout.PlotBottom)
+	for _, tick := range scale.Ticks {
+		y := chartYForValue(tick, scale, layout.PlotTop, layout.PlotBottom)
+		fmt.Fprintf(b, `<line x1="%d" y1="%.1f" x2="%d" y2="%.1f" stroke="#64748b" stroke-width="1"/>`+"\n",
+			layout.PlotLeft, y, layout.PlotLeft-6, y)
+		fmt.Fprintf(b, `<text x="%d" y="%.1f" text-anchor="end" dy="4" font-family="sans-serif" font-size="10" fill="#475569">%s</text>`+"\n",
+			layout.PlotLeft-10, y, sanitizeChartText(chartYAxisNumber(tick)))
+	}
+}
+
+func writeXAxisLabels(b *strings.Builder, layout chartLayout, times []time.Time, labels []string, start, end time.Time, target int) {
+	pointCount := len(labels)
+	if len(times) > pointCount {
+		pointCount = len(times)
+	}
+	b.WriteString(`<g font-family="sans-serif" font-size="11" fill="#64748b" text-anchor="middle">` + "\n")
+	for _, idx := range gpuChartLabelIndices(pointCount, target) {
+		x := chartXForTime(chartPointTime(times, idx), start, end, layout.PlotLeft, layout.PlotRight)
+		label := ""
+		if idx < len(labels) {
+			label = labels[idx]
+		}
+		fmt.Fprintf(b, `<text x="%.1f" y="%d">%s</text>`+"\n", x, layout.PlotBottom+28, sanitizeChartText(label))
+	}
+	b.WriteString(`</g>` + "\n")
+	fmt.Fprintf(b, `<text x="%d" y="%d" text-anchor="middle" font-family="sans-serif" font-size="12" fill="#64748b">Time</text>`+"\n",
+		(layout.PlotLeft+layout.PlotRight)/2, layout.PlotBottom+48)
+}
+
+func writeSeriesPolyline(b *strings.Builder, layout chartLayout, times []time.Time, start, end time.Time, values []float64, scale chartScale, color string) {
+	if len(values) == 0 {
+		return
+	}
+	var points strings.Builder
+	for idx, value := range values {
+		if idx > 0 {
+			points.WriteByte(' ')
+		}
+		x := chartXForTime(chartPointTime(times, idx), start, end, layout.PlotLeft, layout.PlotRight)
+		y := chartYForValue(value, scale, layout.PlotTop, layout.PlotBottom)
+		points.WriteString(strconv.FormatFloat(x, 'f', 1, 64))
+		points.WriteByte(',')
+		points.WriteString(strconv.FormatFloat(y, 'f', 1, 64))
+	}
+	fmt.Fprintf(b, `<polyline points="%s" fill="none" stroke="%s" stroke-width="2.2" stroke-linejoin="round" stroke-linecap="round"/>`+"\n",
+		points.String(), color)
+	if len(values) == 1 {
+		x := chartXForTime(chartPointTime(times, 0), start, end, layout.PlotLeft, layout.PlotRight)
+		y := chartYForValue(values[0], scale, layout.PlotTop, layout.PlotBottom)
+		fmt.Fprintf(b, `<circle cx="%.1f" cy="%.1f" r="3.5" fill="%s"/>`+"\n", x, y, color)
+		return
+	}
+	peakIdx := 0
+	peakValue := values[0]
+	for idx, value := range values[1:] {
+		if value >= peakValue {
+			peakIdx = idx + 1
+			peakValue = value
+		}
+	}
+	x := chartXForTime(chartPointTime(times, peakIdx), start, end, layout.PlotLeft, layout.PlotRight)
+	y := chartYForValue(peakValue, scale, layout.PlotTop, layout.PlotBottom)
+	fmt.Fprintf(b, `<circle cx="%.1f" cy="%.1f" r="4.2" fill="%s" stroke="#ffffff" stroke-width="1.6"/>`+"\n", x, y, color)
+	fmt.Fprintf(b, `<path d="M %.1f %.1f L %.1f %.1f L %.1f %.1f Z" fill="%s" opacity="0.9"/>`+"\n",
+		x, y-10, x-5, y-18, x+5, y-18, color)
+}
+
+func writeLegend(b *strings.Builder, layout chartLayout, series []metricChartSeries) {
+	if !chartLegendVisible(len(series)) || len(series) == 0 {
+		return
+	}
+	cols := 4
+	if len(series) < cols {
+		cols = len(series)
+	}
+	cellWidth := float64(layout.PlotRight-layout.PlotLeft) / float64(cols)
+	baseY := layout.PlotBottom + 74
+	for i, item := range series {
+		row := i / cols
+		col := i % cols
+		x := float64(layout.PlotLeft) + cellWidth*float64(col) + 8
+		y := float64(baseY + row*24)
+		fmt.Fprintf(b, `<line x1="%.1f" y1="%.1f" x2="%.1f" y2="%.1f" stroke="%s" stroke-width="3"/>`+"\n",
+			x, y, x+28, y, item.Color)
+		fmt.Fprintf(b, `<text x="%.1f" y="%.1f" font-family="sans-serif" font-size="12" fill="#1f2937">%s</text>`+"\n",
+			x+38, y+4, sanitizeChartText(item.Name))
+	}
+}
+
+func writeTimelineIdleSpans(b *strings.Builder, layout chartLayout, start, end time.Time, segments []chartTimelineSegment) {
+	if len(segments) == 0 {
+		return
+	}
+	b.WriteString(`<g data-role="timeline-overlay">` + "\n")
+	for _, segment := range segments {
+		if segment.Active || !segment.End.After(segment.Start) {
+			continue
+		}
+		x0 := chartXForTime(segment.Start, start, end, layout.PlotLeft, layout.PlotRight)
+		x1 := chartXForTime(segment.End, start, end, layout.PlotLeft, layout.PlotRight)
+		fmt.Fprintf(b, `<rect x="%.1f" y="%d" width="%.1f" height="%d" fill="#475569" opacity="0.10"/>`+"\n",
+			x0, layout.PlotTop, math.Max(1, x1-x0), layout.PlotBottom-layout.PlotTop)
+	}
+	b.WriteString(`</g>` + "\n")
+}
+
+func writeTimelineBoundaries(b *strings.Builder, layout chartLayout, start, end time.Time, segments []chartTimelineSegment) {
+	if len(segments) == 0 {
+		return
+	}
+	seen := map[int]bool{}
+	b.WriteString(`<g data-role="timeline-boundaries" stroke="#94a3b8" stroke-width="1.2">` + "\n")
+	for i, segment := range segments {
+		if i > 0 {
+			x := int(math.Round(chartXForTime(segment.Start, start, end, layout.PlotLeft, layout.PlotRight)))
+			if !seen[x] {
+				seen[x] = true
+				fmt.Fprintf(b, `<line x1="%d" y1="%d" x2="%d" y2="%d"/>`+"\n", x, layout.PlotTop, x, layout.PlotBottom)
+			}
+		}
+		if i < len(segments)-1 {
+			x := int(math.Round(chartXForTime(segment.End, start, end, layout.PlotLeft, layout.PlotRight)))
+			if !seen[x] {
+				seen[x] = true
+				fmt.Fprintf(b, `<line x1="%d" y1="%d" x2="%d" y2="%d"/>`+"\n", x, layout.PlotTop, x, layout.PlotBottom)
+			}
+		}
+	}
+	b.WriteString(`</g>` + "\n")
+}
+
+func chartXForTime(ts, start, end time.Time, left, right int) float64 {
+	if !end.After(start) {
+		return float64(left+right) / 2
+	}
+	if ts.Before(start) {
+		ts = start
+	}
+	if ts.After(end) {
+		ts = end
+	}
+	ratio := float64(ts.Sub(start)) / float64(end.Sub(start))
+	return float64(left) + ratio*float64(right-left)
+}
+
+func chartPointTime(times []time.Time, idx int) time.Time {
+	if idx >= 0 && idx < len(times) && !times[idx].IsZero() {
+		return times[idx].UTC()
+	}
+	if len(times) > 0 && !times[0].IsZero() {
+		return times[0].UTC().Add(time.Duration(idx) * time.Minute)
+	}
+	return time.Now().UTC().Add(time.Duration(idx) * time.Minute)
+}
+
+func chartYForValue(value float64, scale chartScale, plotTop, plotBottom int) float64 {
+	if scale.Max <= scale.Min {
+		return float64(plotTop+plotBottom) / 2
+	}
+	return float64(plotBottom) - (value-scale.Min)/(scale.Max-scale.Min)*float64(plotBottom-plotTop)
+}
+
+func chartSeriesBounds(values []float64) (float64, float64) {
+	if len(values) == 0 {
+		return 0, 1
+	}
+	min, max := values[0], values[0]
+	for _, value := range values[1:] {
+		if value < min {
+			min = value
+		}
+		if value > max {
+			max = value
+		}
+	}
+	if min == max {
+		if max == 0 {
+			return 0, 1
+		}
+		pad := math.Abs(max) * 0.1
+		if pad == 0 {
+			pad = 1
+		}
+		min -= pad
+		max += pad
+	}
+	if min > 0 {
+		pad := (max - min) * 0.2
+		if pad == 0 {
+			pad = max * 0.1
+		}
+		min -= pad
+		if min < 0 {
+			min = 0
+		}
+		max += pad
+	}
+	return min, max
+}
+
+func chartNiceTicks(min, max float64, target int) []float64 {
+	if min == max {
+		max = min + 1
+	}
+	span := max - min
+	step := math.Pow(10, math.Floor(math.Log10(span/float64(target))))
+	for _, factor := range []float64{1, 2, 5, 10} {
+		if span/(factor*step) <= float64(target)*1.5 {
+			step = factor * step
+			break
+		}
+	}
+	low := math.Floor(min/step) * step
+	high := math.Ceil(max/step) * step
+	var ticks []float64
+	for value := low; value <= high+step*0.001; value += step {
+		ticks = append(ticks, math.Round(value*1e9)/1e9)
+	}
+	return ticks
+}
+
+func valueClamp(value float64, scale chartScale) float64 {
+	if value < scale.Min {
+		return scale.Min
+	}
+	if value > scale.Max {
+		return scale.Max
+	}
+	return value
+}
+
+func chartStatsLabel(datasets [][]float64) string {
+	mn, avg, mx := globalStats(datasets)
+	if mx <= 0 && avg <= 0 && mn <= 0 {
+		return ""
+	}
+	return fmt.Sprintf("min %s   avg %s   max %s",
+		chartLegendNumber(mn),
+		chartLegendNumber(avg),
+		chartLegendNumber(mx),
+	)
+}
+
+func gpuDisplayLabel(idx int) string {
+	if name := gpuModelNameByIndex(idx); name != "" {
+		return fmt.Sprintf("GPU %d — %s", idx, name)
+	}
+	return fmt.Sprintf("GPU %d", idx)
+}
+
+func gpuModelNameByIndex(idx int) string {
+	now := time.Now()
+	gpuLabelCache.mu.Lock()
+	if now.Sub(gpuLabelCache.loadedAt) > 30*time.Second || gpuLabelCache.byIndex == nil {
+		gpuLabelCache.loadedAt = now
+		gpuLabelCache.byIndex = loadGPUModelNames()
+	}
+	name := strings.TrimSpace(gpuLabelCache.byIndex[idx])
+	gpuLabelCache.mu.Unlock()
+	return name
+}
+
+func loadGPUModelNames() map[int]string {
+	out := map[int]string{}
+	gpus, err := platform.New().ListNvidiaGPUs()
+	if err != nil {
+		return out
+	}
+	for _, gpu := range gpus {
+		name := strings.TrimSpace(gpu.Name)
+		if name != "" {
+			out[gpu.Index] = name
+		}
+	}
+	return out
+}

audit/internal/webui/jobs.go

@@ -9,13 +9,14 @@ import (
 
 // jobState holds the output lines and completion status of an async job.
 type jobState struct {
-	lines   []string
-	done    bool
-	err     string
-	mu      sync.Mutex
-	subs    []chan string
-	cancel  func() // optional cancel function; nil if job is not cancellable
-	logPath string
+	lines        []string
+	done         bool
+	err          string
+	mu           sync.Mutex
+	subs         []chan string
+	cancel       func() // optional cancel function; nil if job is not cancellable
+	logPath      string
+	serialPrefix string
 }
 
 // abort cancels the job if it has a cancel function and is not yet done.
@@ -36,6 +37,9 @@ func (j *jobState) append(line string) {
 	if j.logPath != "" {
 		appendJobLog(j.logPath, line)
 	}
+	if j.serialPrefix != "" {
+		taskSerialWriteLine(j.serialPrefix + line)
+	}
 	for _, ch := range j.subs {
 		select {
 		case ch <- line:
@@ -84,12 +88,12 @@ func (m *jobManager) create(id string) *jobState {
 	j := &jobState{}
 	m.jobs[id] = j
 	// Schedule cleanup after 30 minutes
-	go func() {
+	goRecoverOnce("job cleanup", func() {
 		time.Sleep(30 * time.Minute)
 		m.mu.Lock()
 		delete(m.jobs, id)
 		m.mu.Unlock()
-	}()
+	})
 	return j
 }
 
@@ -107,8 +111,11 @@ func (m *jobManager) get(id string) (*jobState, bool) {
 	return j, ok
 }
 
-func newTaskJobState(logPath string) *jobState {
+func newTaskJobState(logPath string, serialPrefix ...string) *jobState {
 	j := &jobState{logPath: logPath}
+	if len(serialPrefix) > 0 {
+		j.serialPrefix = serialPrefix[0]
+	}
 	if logPath == "" {
 		return j
 	}

audit/internal/webui/kmsg_watcher.go

@@ -0,0 +1,242 @@
+package webui
+
+import (
+	"bufio"
+	"io"
+	"log/slog"
+	"os"
+	"strings"
+	"sync"
+	"time"
+
+	"bee/audit/internal/app"
+	"bee/audit/internal/platform"
+)
+
+// kmsgWatcher reads /dev/kmsg and accumulates hardware error events.
+// It supports multiple concurrent SAT tasks: a shared event window is open
+// while any SAT task is running, and flushed when all tasks complete.
+type kmsgWatcher struct {
+	mu          sync.Mutex
+	activeCount int // number of in-flight SAT tasks
+	window      *kmsgWindow
+	statusDB    *app.ComponentStatusDB
+}
+
+type kmsgWindow struct {
+	targets   []string // SAT targets running concurrently
+	startedAt time.Time
+	seen      map[kmsgEventKey]bool
+	events    []kmsgEvent
+}
+
+type kmsgEventKey struct {
+	id       string // BDF or device name
+	category string
+}
+
+type kmsgEvent struct {
+	timestamp time.Time
+	raw       string
+	ids       []string // BDF addresses or device names extracted
+	category  string
+}
+
+func newKmsgWatcher(statusDB *app.ComponentStatusDB) *kmsgWatcher {
+	return &kmsgWatcher{statusDB: statusDB}
+}
+
+// start launches the background kmsg reading goroutine.
+func (w *kmsgWatcher) start() {
+	goRecoverLoop("kmsg watcher", 5*time.Second, w.run)
+}
+
+func (w *kmsgWatcher) run() {
+	for {
+		f, err := os.Open("/dev/kmsg")
+		if err != nil {
+			slog.Warn("kmsg watcher unavailable", "err", err)
+			time.Sleep(30 * time.Second)
+			continue
+		}
+		// Best-effort seek to end so we only capture events from now forward.
+		_, _ = f.Seek(0, io.SeekEnd)
+
+		scanner := bufio.NewScanner(f)
+		scanner.Buffer(make([]byte, 64*1024), 64*1024)
+		for scanner.Scan() {
+			line := scanner.Text()
+			evt, ok := parseKmsgLine(line)
+			if !ok {
+				continue
+			}
+			w.mu.Lock()
+			if w.window != nil {
+				w.recordEvent(evt)
+			}
+			w.mu.Unlock()
+		}
+		if err := scanner.Err(); err != nil {
+			slog.Warn("kmsg watcher stopped", "err", err)
+		}
+		_ = f.Close()
+		time.Sleep(2 * time.Second)
+	}
+}
+
+// recordEvent appends evt to the active window, deduplicating by (id, category).
+// Must be called with w.mu held.
+func (w *kmsgWatcher) recordEvent(evt kmsgEvent) {
+	if len(evt.ids) == 0 {
+		key := kmsgEventKey{id: "", category: evt.category}
+		if !w.window.seen[key] {
+			w.window.seen[key] = true
+			w.window.events = append(w.window.events, evt)
+		}
+		return
+	}
+	for _, id := range evt.ids {
+		key := kmsgEventKey{id: id, category: evt.category}
+		if !w.window.seen[key] {
+			w.window.seen[key] = true
+			w.window.events = append(w.window.events, evt)
+		}
+	}
+}
+
+// NotifyTaskStarted increments the active task counter and opens a shared event window
+// if this is the first task starting.
+func (w *kmsgWatcher) NotifyTaskStarted(taskID, target string) {
+	w.mu.Lock()
+	defer w.mu.Unlock()
+	if w.activeCount == 0 {
+		w.window = &kmsgWindow{
+			startedAt: time.Now(),
+			seen:      make(map[kmsgEventKey]bool),
+		}
+	}
+	w.activeCount++
+	if w.window != nil {
+		w.window.targets = append(w.window.targets, target)
+	}
+}
+
+// NotifyTaskFinished decrements the active task counter. When all tasks finish,
+// it flushes the accumulated events to the status DB.
+func (w *kmsgWatcher) NotifyTaskFinished(taskID string) {
+	w.mu.Lock()
+	w.activeCount--
+	var window *kmsgWindow
+	if w.activeCount <= 0 {
+		w.activeCount = 0
+		window = w.window
+		w.window = nil
+	}
+	w.mu.Unlock()
+
+	if window == nil || len(window.events) == 0 {
+		return
+	}
+	goRecoverOnce("kmsg watcher flush", func() { w.flushWindow(window) })
+}
+
+func (w *kmsgWatcher) flushWindow(window *kmsgWindow) {
+	if w.statusDB == nil {
+		return
+	}
+	source := "watchdog:kmsg"
+	// Collect unique component keys from events.
+	seen := map[string]string{} // componentKey → first raw line
+	for _, evt := range window.events {
+		if len(evt.ids) == 0 {
+			// MCE or un-identified error.
+			key := "cpu:all"
+			if evt.category == "memory" {
+				key = "memory:all"
+			}
+			if _, exists := seen[key]; !exists {
+				seen[key] = evt.raw
+			}
+			continue
+		}
+		for _, id := range evt.ids {
+			var key string
+			switch evt.category {
+			case "gpu", "pcie":
+				key = "pcie:" + normalizeBDF(id)
+			case "storage":
+				key = "storage:" + id
+			default:
+				key = "pcie:" + normalizeBDF(id)
+			}
+			if _, exists := seen[key]; !exists {
+				seen[key] = evt.raw
+			}
+		}
+	}
+	for key, detail := range seen {
+		detail = "kernel error during SAT (" + strings.Join(window.targets, ",") + "): " + truncate(detail, 120)
+		w.statusDB.Record(key, source, "Warning", detail)
+	}
+}
+
+// parseKmsgLine parses a single /dev/kmsg line and returns an event if it matches
+// any pattern in platform.HardwareErrorPatterns.
+// kmsg format: "<priority>,<sequence>,<timestamp_usec>,-;message text"
+func parseKmsgLine(raw string) (kmsgEvent, bool) {
+	msg := raw
+	if idx := strings.Index(raw, ";"); idx >= 0 {
+		msg = strings.TrimSpace(raw[idx+1:])
+	}
+	if msg == "" {
+		return kmsgEvent{}, false
+	}
+
+	for _, p := range platform.HardwareErrorPatterns {
+		m := p.Re.FindStringSubmatch(msg)
+		if m == nil {
+			continue
+		}
+		evt := kmsgEvent{
+			timestamp: time.Now(),
+			raw:       msg,
+			category:  p.Category,
+		}
+		if p.BDFGroup > 0 && p.BDFGroup < len(m) {
+			evt.ids = append(evt.ids, normalizeBDF(m[p.BDFGroup]))
+		}
+		if p.DevGroup > 0 && p.DevGroup < len(m) {
+			evt.ids = append(evt.ids, m[p.DevGroup])
+		}
+		return evt, true
+	}
+	return kmsgEvent{}, false
+}
+
+// normalizeBDF normalizes a PCIe BDF to the 4-part form "0000:c8:00.0".
+func normalizeBDF(bdf string) string {
+	bdf = strings.ToLower(strings.TrimSpace(bdf))
+	if strings.Count(bdf, ":") == 1 {
+		return "0000:" + bdf
+	}
+	return bdf
+}
+
+func truncate(s string, max int) string {
+	if len(s) <= max {
+		return s
+	}
+	return s[:max] + "..."
+}
+
+// isSATTarget returns true for task targets that run hardware acceptance tests.
+func isSATTarget(target string) bool {
+	switch target {
+	case "nvidia", "nvidia-targeted-stress", "nvidia-benchmark", "nvidia-compute", "nvidia-targeted-power", "nvidia-pulse",
+		"nvidia-interconnect", "nvidia-bandwidth", "nvidia-stress", "memory", "memory-stress", "storage",
+		"cpu", "sat-stress", "amd", "amd-mem", "amd-bandwidth", "amd-stress",
+		"platform-stress":
+		return true
+	}
+	return false
+}

audit/internal/webui/metricsdb.go

@@ -8,6 +8,7 @@ import (
 	"path/filepath"
 	"sort"
 	"strconv"
+	"strings"
 	"time"
 
 	"bee/audit/internal/platform"
@@ -21,6 +22,13 @@ type MetricsDB struct {
 	db *sql.DB
 }
 
+func (m *MetricsDB) Close() error {
+	if m == nil || m.db == nil {
+		return nil
+	}
+	return m.db.Close()
+}
+
 // openMetricsDB opens (or creates) the metrics database at the given path.
 func openMetricsDB(path string) (*MetricsDB, error) {
 	if err := os.MkdirAll(filepath.Dir(path), 0755); err != nil {
@@ -54,6 +62,8 @@ CREATE TABLE IF NOT EXISTS gpu_metrics (
   usage_pct     REAL,
   mem_usage_pct REAL,
   power_w       REAL,
+  clock_mhz     REAL,
+  mem_clock_mhz REAL,
   PRIMARY KEY (ts, gpu_index)
 );
 CREATE TABLE IF NOT EXISTS fan_metrics (
@@ -70,6 +80,38 @@ CREATE TABLE IF NOT EXISTS temp_metrics (
   PRIMARY KEY (ts, name)
 );
 `)
+	if err != nil {
+		return err
+	}
+	if err := ensureMetricsColumn(db, "gpu_metrics", "clock_mhz", "REAL"); err != nil {
+		return err
+	}
+	return ensureMetricsColumn(db, "gpu_metrics", "mem_clock_mhz", "REAL")
+}
+
+func ensureMetricsColumn(db *sql.DB, table, column, definition string) error {
+	rows, err := db.Query("PRAGMA table_info(" + table + ")")
+	if err != nil {
+		return err
+	}
+	defer rows.Close()
+
+	for rows.Next() {
+		var cid int
+		var name, ctype string
+		var notNull, pk int
+		var dflt sql.NullString
+		if err := rows.Scan(&cid, &name, &ctype, &notNull, &dflt, &pk); err != nil {
+			return err
+		}
+		if strings.EqualFold(name, column) {
+			return nil
+		}
+	}
+	if err := rows.Err(); err != nil {
+		return err
+	}
+	_, err = db.Exec("ALTER TABLE " + table + " ADD COLUMN " + column + " " + definition)
 	return err
 }
 
@@ -91,8 +133,8 @@ func (m *MetricsDB) Write(s platform.LiveMetricSample) error {
 	}
 	for _, g := range s.GPUs {
 		_, err = tx.Exec(
-			`INSERT OR REPLACE INTO gpu_metrics(ts,gpu_index,temp_c,usage_pct,mem_usage_pct,power_w) VALUES(?,?,?,?,?,?)`,
-			ts, g.GPUIndex, g.TempC, g.UsagePct, g.MemUsagePct, g.PowerW,
+			`INSERT OR REPLACE INTO gpu_metrics(ts,gpu_index,temp_c,usage_pct,mem_usage_pct,power_w,clock_mhz,mem_clock_mhz) VALUES(?,?,?,?,?,?,?,?)`,
+			ts, g.GPUIndex, g.TempC, g.UsagePct, g.MemUsagePct, g.PowerW, g.ClockMHz, g.MemClockMHz,
 		)
 		if err != nil {
 			return err
@@ -129,6 +171,23 @@ func (m *MetricsDB) LoadAll() ([]platform.LiveMetricSample, error) {
 	return m.loadSamples(`SELECT ts,cpu_load_pct,mem_load_pct,power_w FROM sys_metrics ORDER BY ts`, nil)
 }
 
+// LoadBetween returns samples in chronological order within the given time window.
+func (m *MetricsDB) LoadBetween(start, end time.Time) ([]platform.LiveMetricSample, error) {
+	if m == nil {
+		return nil, nil
+	}
+	if start.IsZero() || end.IsZero() {
+		return nil, nil
+	}
+	if end.Before(start) {
+		start, end = end, start
+	}
+	return m.loadSamples(
+		`SELECT ts,cpu_load_pct,mem_load_pct,power_w FROM sys_metrics WHERE ts>=? AND ts<=? ORDER BY ts`,
+		start.Unix(), end.Unix(),
+	)
+}
+
 // loadSamples reconstructs LiveMetricSample rows from the normalized tables.
 func (m *MetricsDB) loadSamples(query string, args ...any) ([]platform.LiveMetricSample, error) {
 	rows, err := m.db.Query(query, args...)
@@ -163,7 +222,7 @@ func (m *MetricsDB) loadSamples(query string, args ...any) ([]platform.LiveMetri
 	}
 	gpuData := map[gpuKey]platform.GPUMetricRow{}
 	gRows, err := m.db.Query(
-		`SELECT ts,gpu_index,temp_c,usage_pct,mem_usage_pct,power_w FROM gpu_metrics WHERE ts>=? AND ts<=? ORDER BY ts,gpu_index`,
+		`SELECT ts,gpu_index,temp_c,usage_pct,mem_usage_pct,power_w,IFNULL(clock_mhz,0),IFNULL(mem_clock_mhz,0) FROM gpu_metrics WHERE ts>=? AND ts<=? ORDER BY ts,gpu_index`,
 		minTS, maxTS,
 	)
 	if err == nil {
@@ -171,7 +230,7 @@ func (m *MetricsDB) loadSamples(query string, args ...any) ([]platform.LiveMetri
 		for gRows.Next() {
 			var ts int64
 			var g platform.GPUMetricRow
-			if err := gRows.Scan(&ts, &g.GPUIndex, &g.TempC, &g.UsagePct, &g.MemUsagePct, &g.PowerW); err == nil {
+			if err := gRows.Scan(&ts, &g.GPUIndex, &g.TempC, &g.UsagePct, &g.MemUsagePct, &g.PowerW, &g.ClockMHz, &g.MemClockMHz); err == nil {
 				gpuData[gpuKey{ts, g.GPUIndex}] = g
 			}
 		}
@@ -283,7 +342,8 @@ func (m *MetricsDB) loadSamples(query string, args ...any) ([]platform.LiveMetri
 func (m *MetricsDB) ExportCSV(w io.Writer) error {
 	rows, err := m.db.Query(`
 		SELECT s.ts, s.cpu_load_pct, s.mem_load_pct, s.power_w,
-		       g.gpu_index, g.temp_c, g.usage_pct, g.mem_usage_pct, g.power_w
+		       g.gpu_index, g.temp_c, g.usage_pct, g.mem_usage_pct, g.power_w,
+		       g.clock_mhz, g.mem_clock_mhz
 		FROM sys_metrics s
 		LEFT JOIN gpu_metrics g ON g.ts = s.ts
 		ORDER BY s.ts, g.gpu_index
@@ -294,13 +354,13 @@ func (m *MetricsDB) ExportCSV(w io.Writer) error {
 	defer rows.Close()
 
 	cw := csv.NewWriter(w)
-	_ = cw.Write([]string{"ts", "cpu_load_pct", "mem_load_pct", "sys_power_w", "gpu_index", "gpu_temp_c", "gpu_usage_pct", "gpu_mem_pct", "gpu_power_w"})
+	_ = cw.Write([]string{"ts", "cpu_load_pct", "mem_load_pct", "sys_power_w", "gpu_index", "gpu_temp_c", "gpu_usage_pct", "gpu_mem_pct", "gpu_power_w", "gpu_clock_mhz", "gpu_mem_clock_mhz"})
 	for rows.Next() {
 		var ts int64
 		var cpu, mem, pwr float64
 		var gpuIdx sql.NullInt64
-		var gpuTemp, gpuUse, gpuMem, gpuPow sql.NullFloat64
-		if err := rows.Scan(&ts, &cpu, &mem, &pwr, &gpuIdx, &gpuTemp, &gpuUse, &gpuMem, &gpuPow); err != nil {
+		var gpuTemp, gpuUse, gpuMem, gpuPow, gpuClock, gpuMemClock sql.NullFloat64
+		if err := rows.Scan(&ts, &cpu, &mem, &pwr, &gpuIdx, &gpuTemp, &gpuUse, &gpuMem, &gpuPow, &gpuClock, &gpuMemClock); err != nil {
 			continue
 		}
 		row := []string{
@@ -316,9 +376,11 @@ func (m *MetricsDB) ExportCSV(w io.Writer) error {
 				strconv.FormatFloat(gpuUse.Float64, 'f', 1, 64),
 				strconv.FormatFloat(gpuMem.Float64, 'f', 1, 64),
 				strconv.FormatFloat(gpuPow.Float64, 'f', 1, 64),
+				strconv.FormatFloat(gpuClock.Float64, 'f', 1, 64),
+				strconv.FormatFloat(gpuMemClock.Float64, 'f', 1, 64),
 			)
 		} else {
-			row = append(row, "", "", "", "", "")
+			row = append(row, "", "", "", "", "", "", "")
 		}
 		_ = cw.Write(row)
 	}
@@ -326,9 +388,6 @@ func (m *MetricsDB) ExportCSV(w io.Writer) error {
 	return cw.Error()
 }
 
-// Close closes the database.
-func (m *MetricsDB) Close() { _ = m.db.Close() }
-
 func nullFloat(v float64) sql.NullFloat64 {
 	return sql.NullFloat64{Float64: v, Valid: true}
 }

audit/internal/webui/metricsdb_test.go

@@ -1,11 +1,13 @@
 package webui
 
 import (
+	"database/sql"
 	"path/filepath"
 	"testing"
 	"time"
 
 	"bee/audit/internal/platform"
+	_ "modernc.org/sqlite"
 )
 
 func TestMetricsDBLoadSamplesKeepsChronologicalRangeForGPUs(t *testing.T) {
@@ -67,3 +69,106 @@ func TestMetricsDBLoadSamplesKeepsChronologicalRangeForGPUs(t *testing.T) {
 		}
 	}
 }
+
+func TestMetricsDBMigratesLegacyGPUSchema(t *testing.T) {
+	path := filepath.Join(t.TempDir(), "metrics.db")
+	raw, err := sql.Open("sqlite", path)
+	if err != nil {
+		t.Fatalf("sql.Open: %v", err)
+	}
+	_, err = raw.Exec(`
+CREATE TABLE gpu_metrics (
+  ts            INTEGER NOT NULL,
+  gpu_index     INTEGER NOT NULL,
+  temp_c        REAL,
+  usage_pct     REAL,
+  mem_usage_pct REAL,
+  power_w       REAL,
+  PRIMARY KEY (ts, gpu_index)
+);
+CREATE TABLE sys_metrics (
+  ts           INTEGER NOT NULL,
+  cpu_load_pct REAL,
+  mem_load_pct REAL,
+  power_w      REAL,
+  PRIMARY KEY (ts)
+);
+CREATE TABLE fan_metrics (
+  ts   INTEGER NOT NULL,
+  name TEXT NOT NULL,
+  rpm  REAL,
+  PRIMARY KEY (ts, name)
+);
+CREATE TABLE temp_metrics (
+  ts      INTEGER NOT NULL,
+  name    TEXT NOT NULL,
+  grp     TEXT NOT NULL,
+  celsius REAL,
+  PRIMARY KEY (ts, name)
+);
+`)
+	if err != nil {
+		t.Fatalf("create legacy schema: %v", err)
+	}
+	_ = raw.Close()
+
+	db, err := openMetricsDB(path)
+	if err != nil {
+		t.Fatalf("openMetricsDB: %v", err)
+	}
+	defer db.Close()
+
+	now := time.Unix(1_700_000_100, 0).UTC()
+	err = db.Write(platform.LiveMetricSample{
+		Timestamp: now,
+		GPUs: []platform.GPUMetricRow{
+			{GPUIndex: 0, ClockMHz: 1410, MemClockMHz: 2600},
+		},
+	})
+	if err != nil {
+		t.Fatalf("Write: %v", err)
+	}
+
+	samples, err := db.LoadAll()
+	if err != nil {
+		t.Fatalf("LoadAll: %v", err)
+	}
+	if len(samples) != 1 || len(samples[0].GPUs) != 1 {
+		t.Fatalf("samples=%+v", samples)
+	}
+	if got := samples[0].GPUs[0].ClockMHz; got != 1410 {
+		t.Fatalf("ClockMHz=%v want 1410", got)
+	}
+	if got := samples[0].GPUs[0].MemClockMHz; got != 2600 {
+		t.Fatalf("MemClockMHz=%v want 2600", got)
+	}
+}
+
+func TestMetricsDBLoadBetweenFiltersWindow(t *testing.T) {
+	db, err := openMetricsDB(filepath.Join(t.TempDir(), "metrics.db"))
+	if err != nil {
+		t.Fatalf("openMetricsDB: %v", err)
+	}
+	defer db.Close()
+
+	base := time.Unix(1_700_000_000, 0).UTC()
+	for i := 0; i < 5; i++ {
+		if err := db.Write(platform.LiveMetricSample{
+			Timestamp:  base.Add(time.Duration(i) * time.Minute),
+			CPULoadPct: float64(i),
+		}); err != nil {
+			t.Fatalf("Write(%d): %v", i, err)
+		}
+	}
+
+	got, err := db.LoadBetween(base.Add(1*time.Minute), base.Add(3*time.Minute))
+	if err != nil {
+		t.Fatalf("LoadBetween: %v", err)
+	}
+	if len(got) != 3 {
+		t.Fatalf("LoadBetween len=%d want 3", len(got))
+	}
+	if !got[0].Timestamp.Equal(base.Add(1*time.Minute)) || !got[2].Timestamp.Equal(base.Add(3*time.Minute)) {
+		t.Fatalf("window=%v..%v", got[0].Timestamp, got[2].Timestamp)
+	}
+}

audit/internal/webui/serial_console.go

@@ -0,0 +1,41 @@
+package webui
+
+import (
+	"fmt"
+	"os"
+	"strings"
+	"time"
+)
+
+var taskSerialWriteLine = writeTaskSerialLine
+
+func writeTaskSerialLine(line string) {
+	line = strings.TrimSpace(line)
+	if line == "" {
+		return
+	}
+	payload := fmt.Sprintf("%s %s\n", time.Now().UTC().Format("2006-01-02 15:04:05Z"), line)
+	for _, path := range []string{"/dev/ttyS0", "/dev/ttyS1", "/dev/console"} {
+		f, err := os.OpenFile(path, os.O_WRONLY|os.O_APPEND, 0)
+		if err != nil {
+			continue
+		}
+		_, _ = f.WriteString(payload)
+		_ = f.Close()
+		return
+	}
+}
+
+func taskSerialPrefix(t *Task) string {
+	if t == nil {
+		return "[task] "
+	}
+	return fmt.Sprintf("[task %s %s] ", t.ID, t.Name)
+}
+
+func taskSerialEvent(t *Task, event string) {
+	if t == nil {
+		return
+	}
+	taskSerialWriteLine(fmt.Sprintf("%s%s", taskSerialPrefix(t), strings.TrimSpace(event)))
+}

audit/internal/webui/server.go

@@ -1,15 +1,19 @@
 package webui
 
 import (
+	"bufio"
 	"encoding/json"
 	"errors"
 	"fmt"
 	"html"
+	"io"
 	"log/slog"
 	"mime"
+	"net"
 	"net/http"
 	"os"
 	"path/filepath"
+	"runtime/debug"
 	"sort"
 	"strings"
 	"sync"
@@ -18,7 +22,6 @@ import (
 	"bee/audit/internal/app"
 	"bee/audit/internal/platform"
 	"bee/audit/internal/runtimeenv"
-	gocharts "github.com/go-analyze/charts"
 	"reanimator/chart/viewer"
 	"reanimator/chart/web"
 )
@@ -164,6 +167,8 @@ type handler struct {
 	// pending network change (rollback on timeout)
 	pendingNet   *pendingNetChange
 	pendingNetMu sync.Mutex
+	// kmsg hardware error watcher
+	kmsg *kmsgWatcher
 }
 
 // NewHandler creates the HTTP mux with all routes.
@@ -203,12 +208,24 @@ func NewHandler(opts HandlerOptions) http.Handler {
 	}
 	h.startMetricsCollector()
 
+	// Start kmsg hardware error watcher if the app (and its status DB) is available.
+	if opts.App != nil {
+		h.kmsg = newKmsgWatcher(opts.App.StatusDB)
+		h.kmsg.start()
+		globalQueue.kmsgWatcher = h.kmsg
+	}
+
 	globalQueue.startWorker(&opts)
 	mux := http.NewServeMux()
 
 	// ── Infrastructure ──────────────────────────────────────────────────────
 	mux.HandleFunc("GET /healthz", h.handleHealthz)
 	mux.HandleFunc("GET /api/ready", h.handleReady)
+	mux.HandleFunc("GET /loading", func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Cache-Control", "no-store")
+		w.Header().Set("Content-Type", "text/html; charset=utf-8")
+		_, _ = w.Write([]byte(loadingPageHTML))
+	})
 
 	// ── Existing read-only endpoints (preserved for compatibility) ──────────
 	mux.HandleFunc("GET /audit.json", h.handleAuditJSON)
@@ -225,6 +242,12 @@ func NewHandler(opts HandlerOptions) http.Handler {
 
 	// SAT
 	mux.HandleFunc("POST /api/sat/nvidia/run", h.handleAPISATRun("nvidia"))
+	mux.HandleFunc("POST /api/sat/nvidia-targeted-stress/run", h.handleAPISATRun("nvidia-targeted-stress"))
+	mux.HandleFunc("POST /api/sat/nvidia-compute/run", h.handleAPISATRun("nvidia-compute"))
+	mux.HandleFunc("POST /api/sat/nvidia-targeted-power/run", h.handleAPISATRun("nvidia-targeted-power"))
+	mux.HandleFunc("POST /api/sat/nvidia-pulse/run", h.handleAPISATRun("nvidia-pulse"))
+	mux.HandleFunc("POST /api/sat/nvidia-interconnect/run", h.handleAPISATRun("nvidia-interconnect"))
+	mux.HandleFunc("POST /api/sat/nvidia-bandwidth/run", h.handleAPISATRun("nvidia-bandwidth"))
 	mux.HandleFunc("POST /api/sat/nvidia-stress/run", h.handleAPISATRun("nvidia-stress"))
 	mux.HandleFunc("POST /api/sat/memory/run", h.handleAPISATRun("memory"))
 	mux.HandleFunc("POST /api/sat/storage/run", h.handleAPISATRun("storage"))
@@ -238,6 +261,7 @@ func NewHandler(opts HandlerOptions) http.Handler {
 	mux.HandleFunc("POST /api/sat/platform-stress/run", h.handleAPISATRun("platform-stress"))
 	mux.HandleFunc("GET /api/sat/stream", h.handleAPISATStream)
 	mux.HandleFunc("POST /api/sat/abort", h.handleAPISATAbort)
+	mux.HandleFunc("POST /api/benchmark/nvidia/run", h.handleAPIBenchmarkNvidiaRun)
 
 	// Tasks
 	mux.HandleFunc("GET /api/tasks", h.handleAPITasksList)
@@ -246,6 +270,9 @@ func NewHandler(opts HandlerOptions) http.Handler {
 	mux.HandleFunc("POST /api/tasks/{id}/cancel", h.handleAPITasksCancel)
 	mux.HandleFunc("POST /api/tasks/{id}/priority", h.handleAPITasksPriority)
 	mux.HandleFunc("GET /api/tasks/{id}/stream", h.handleAPITasksStream)
+	mux.HandleFunc("GET /api/tasks/{id}/charts", h.handleAPITaskChartsIndex)
+	mux.HandleFunc("GET /api/tasks/{id}/chart/", h.handleAPITaskChartSVG)
+	mux.HandleFunc("GET /tasks/{id}", h.handleTaskPage)
 
 	// Services
 	mux.HandleFunc("GET /api/services", h.handleAPIServicesList)
@@ -268,12 +295,11 @@ func NewHandler(opts HandlerOptions) http.Handler {
 	// Tools
 	mux.HandleFunc("GET /api/tools/check", h.handleAPIToolsCheck)
 
-	// Display
-	mux.HandleFunc("GET /api/display/resolutions", h.handleAPIDisplayResolutions)
-	mux.HandleFunc("POST /api/display/set", h.handleAPIDisplaySet)
-
 	// GPU presence / tools
 	mux.HandleFunc("GET /api/gpu/presence", h.handleAPIGPUPresence)
+	mux.HandleFunc("GET /api/gpu/nvidia", h.handleAPIGNVIDIAGPUs)
+	mux.HandleFunc("GET /api/gpu/nvidia-status", h.handleAPIGNVIDIAGPUStatuses)
+	mux.HandleFunc("POST /api/gpu/nvidia-reset", h.handleAPIGNVIDIAReset)
 	mux.HandleFunc("GET /api/gpu/tools", h.handleAPIGPUTools)
 
 	// System
@@ -300,11 +326,11 @@ func NewHandler(opts HandlerOptions) http.Handler {
 	mux.HandleFunc("GET /", h.handlePage)
 
 	h.mux = mux
-	return mux
+	return recoverMiddleware(mux)
 }
 
 func (h *handler) startMetricsCollector() {
-	go func() {
+	goRecoverLoop("metrics collector", 2*time.Second, func() {
 		ticker := time.NewTicker(metricsCollectInterval)
 		defer ticker.Stop()
 		for range ticker.C {
@@ -315,7 +341,7 @@ func (h *handler) startMetricsCollector() {
 			h.feedRings(sample)
 			h.setLatestMetric(sample)
 		}
-	}()
+	})
 }
 
 func (h *handler) setLatestMetric(sample platform.LiveMetricSample) {
@@ -336,7 +362,81 @@ func (h *handler) latestMetric() (platform.LiveMetricSample, bool) {
 
 // ListenAndServe starts the HTTP server.
 func ListenAndServe(addr string, opts HandlerOptions) error {
-	return http.ListenAndServe(addr, NewHandler(opts))
+	srv := &http.Server{
+		Addr:              addr,
+		Handler:           NewHandler(opts),
+		ReadHeaderTimeout: 5 * time.Second,
+		ReadTimeout:       30 * time.Second,
+		IdleTimeout:       2 * time.Minute,
+	}
+	return srv.ListenAndServe()
+}
+
+type trackingResponseWriter struct {
+	http.ResponseWriter
+	wroteHeader bool
+}
+
+func (w *trackingResponseWriter) WriteHeader(statusCode int) {
+	w.wroteHeader = true
+	w.ResponseWriter.WriteHeader(statusCode)
+}
+
+func (w *trackingResponseWriter) Write(p []byte) (int, error) {
+	w.wroteHeader = true
+	return w.ResponseWriter.Write(p)
+}
+
+func (w *trackingResponseWriter) Flush() {
+	w.wroteHeader = true
+	if f, ok := w.ResponseWriter.(http.Flusher); ok {
+		f.Flush()
+	}
+}
+
+func (w *trackingResponseWriter) Hijack() (net.Conn, *bufio.ReadWriter, error) {
+	h, ok := w.ResponseWriter.(http.Hijacker)
+	if !ok {
+		return nil, nil, fmt.Errorf("hijacking not supported")
+	}
+	return h.Hijack()
+}
+
+func (w *trackingResponseWriter) Push(target string, opts *http.PushOptions) error {
+	p, ok := w.ResponseWriter.(http.Pusher)
+	if !ok {
+		return http.ErrNotSupported
+	}
+	return p.Push(target, opts)
+}
+
+func (w *trackingResponseWriter) ReadFrom(r io.Reader) (int64, error) {
+	rf, ok := w.ResponseWriter.(io.ReaderFrom)
+	if !ok {
+		return io.Copy(w.ResponseWriter, r)
+	}
+	w.wroteHeader = true
+	return rf.ReadFrom(r)
+}
+
+func recoverMiddleware(next http.Handler) http.Handler {
+	return http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		tw := &trackingResponseWriter{ResponseWriter: w}
+		defer func() {
+			if rec := recover(); rec != nil {
+				slog.Error("http handler panic",
+					"method", r.Method,
+					"path", r.URL.Path,
+					"panic", fmt.Sprint(rec),
+					"stack", string(debug.Stack()),
+				)
+				if !tw.wroteHeader {
+					http.Error(tw, "internal server error", http.StatusInternalServerError)
+				}
+			}
+		}()
+		next.ServeHTTP(tw, r)
+	})
 }
 
 // ── Infrastructure handlers ──────────────────────────────────────────────────
@@ -466,13 +566,44 @@ func (h *handler) handleMetricsChartSVG(w http.ResponseWriter, r *http.Request)
 		http.Error(w, "metrics database not available", http.StatusServiceUnavailable)
 		return
 	}
-	datasets, names, labels, title, yMin, yMax, ok := h.chartDataFromDB(path)
+	samples, err := h.metricsDB.LoadAll()
+	if err != nil || len(samples) == 0 {
+		http.Error(w, "metrics history unavailable", http.StatusServiceUnavailable)
+		return
+	}
+	timeline := metricsTimelineSegments(samples, time.Now())
+	if idx, sub, ok := parseGPUChartPath(path); ok && sub == "overview" {
+		buf, ok, err := renderGPUOverviewChartSVG(idx, samples, timeline)
+		if err != nil {
+			http.Error(w, err.Error(), http.StatusInternalServerError)
+			return
+		}
+		if !ok {
+			http.Error(w, "metrics history unavailable", http.StatusServiceUnavailable)
+			return
+		}
+		w.Header().Set("Content-Type", "image/svg+xml")
+		w.Header().Set("Cache-Control", "no-store")
+		_, _ = w.Write(buf)
+		return
+	}
+	datasets, names, labels, title, yMin, yMax, ok := chartDataFromSamples(path, samples)
 	if !ok {
 		http.Error(w, "metrics history unavailable", http.StatusServiceUnavailable)
 		return
 	}
 
-	buf, err := renderChartSVG(title, datasets, names, labels, yMin, yMax)
+	buf, err := renderMetricChartSVG(
+		title,
+		labels,
+		sampleTimes(samples),
+		datasets,
+		names,
+		yMin,
+		yMax,
+		chartCanvasHeightForPath(path, len(names)),
+		timeline,
+	)
 	if err != nil {
 		http.Error(w, err.Error(), http.StatusInternalServerError)
 		return
@@ -482,14 +613,6 @@ func (h *handler) handleMetricsChartSVG(w http.ResponseWriter, r *http.Request)
 	_, _ = w.Write(buf)
 }
 
-func (h *handler) chartDataFromDB(path string) ([][]float64, []string, []string, string, *float64, *float64, bool) {
-	samples, err := h.metricsDB.LoadAll()
-	if err != nil || len(samples) == 0 {
-		return nil, nil, nil, "", nil, nil, false
-	}
-	return chartDataFromSamples(path, samples)
-}
-
 func chartDataFromSamples(path string, samples []platform.LiveMetricSample) ([][]float64, []string, []string, string, *float64, *float64, bool) {
 	var datasets [][]float64
 	var names []string
@@ -569,18 +692,24 @@ func chartDataFromSamples(path string, samples []platform.LiveMetricSample) ([][
 		yMin = floatPtr(0)
 		yMax = autoMax120(datasets...)
 
+	case path == "gpu-all-clock":
+		title = "GPU Core Clock"
+		datasets, names = gpuDatasets(samples, func(g platform.GPUMetricRow) float64 { return g.ClockMHz })
+		yMin, yMax = autoBounds120(datasets...)
+
+	case path == "gpu-all-memclock":
+		title = "GPU Memory Clock"
+		datasets, names = gpuDatasets(samples, func(g platform.GPUMetricRow) float64 { return g.MemClockMHz })
+		yMin, yMax = autoBounds120(datasets...)
+
 	case strings.HasPrefix(path, "gpu/"):
-		rest := strings.TrimPrefix(path, "gpu/")
-		sub := ""
-		if i := strings.LastIndex(rest, "-"); i > 0 {
-			sub = rest[i+1:]
-			rest = rest[:i]
+		idx, sub, ok := parseGPUChartPath(path)
+		if !ok {
+			return nil, nil, nil, "", nil, nil, false
 		}
-		idx := 0
-		fmt.Sscanf(rest, "%d", &idx)
 		switch sub {
 		case "load":
-			title = fmt.Sprintf("GPU %d Load", idx)
+			title = gpuDisplayLabel(idx) + " Load"
 			util := gpuDatasetByIndex(samples, idx, func(g platform.GPUMetricRow) float64 { return g.UsagePct })
 			mem := gpuDatasetByIndex(samples, idx, func(g platform.GPUMetricRow) float64 { return g.MemUsagePct })
 			if util == nil && mem == nil {
@@ -591,7 +720,7 @@ func chartDataFromSamples(path string, samples []platform.LiveMetricSample) ([][
 			yMin = floatPtr(0)
 			yMax = floatPtr(100)
 		case "temp":
-			title = fmt.Sprintf("GPU %d Temperature", idx)
+			title = gpuDisplayLabel(idx) + " Temperature"
 			temp := gpuDatasetByIndex(samples, idx, func(g platform.GPUMetricRow) float64 { return g.TempC })
 			if temp == nil {
 				return nil, nil, nil, "", nil, nil, false
@@ -600,8 +729,26 @@ func chartDataFromSamples(path string, samples []platform.LiveMetricSample) ([][
 			names = []string{"Temp °C"}
 			yMin = floatPtr(0)
 			yMax = autoMax120(temp)
+		case "clock":
+			title = gpuDisplayLabel(idx) + " Core Clock"
+			clock := gpuDatasetByIndex(samples, idx, func(g platform.GPUMetricRow) float64 { return g.ClockMHz })
+			if clock == nil {
+				return nil, nil, nil, "", nil, nil, false
+			}
+			datasets = [][]float64{clock}
+			names = []string{"Core Clock MHz"}
+			yMin, yMax = autoBounds120(clock)
+		case "memclock":
+			title = gpuDisplayLabel(idx) + " Memory Clock"
+			clock := gpuDatasetByIndex(samples, idx, func(g platform.GPUMetricRow) float64 { return g.MemClockMHz })
+			if clock == nil {
+				return nil, nil, nil, "", nil, nil, false
+			}
+			datasets = [][]float64{clock}
+			names = []string{"Memory Clock MHz"}
+			yMin, yMax = autoBounds120(clock)
 		default:
-			title = fmt.Sprintf("GPU %d Power", idx)
+			title = gpuDisplayLabel(idx) + " Power"
 			power := gpuDatasetByIndex(samples, idx, func(g platform.GPUMetricRow) float64 { return g.PowerW })
 			if power == nil {
 				return nil, nil, nil, "", nil, nil, false
@@ -618,6 +765,26 @@ func chartDataFromSamples(path string, samples []platform.LiveMetricSample) ([][
 	return datasets, names, labels, title, yMin, yMax, len(datasets) > 0
 }
 
+func parseGPUChartPath(path string) (idx int, sub string, ok bool) {
+	if !strings.HasPrefix(path, "gpu/") {
+		return 0, "", false
+	}
+	rest := strings.TrimPrefix(path, "gpu/")
+	if rest == "" {
+		return 0, "", false
+	}
+	sub = ""
+	if i := strings.LastIndex(rest, "-"); i > 0 {
+		sub = rest[i+1:]
+		rest = rest[:i]
+	}
+	n, err := fmt.Sscanf(rest, "%d", &idx)
+	if err != nil || n != 1 {
+		return 0, "", false
+	}
+	return idx, sub, true
+}
+
 func sampleTimeLabels(samples []platform.LiveMetricSample) []string {
 	labels := make([]string, len(samples))
 	if len(samples) == 0 {
@@ -710,7 +877,7 @@ func gpuDatasets(samples []platform.LiveMetricSample, pick func(platform.GPUMetr
 			continue
 		}
 		datasets = append(datasets, ds)
-		names = append(names, fmt.Sprintf("GPU %d", idx))
+		names = append(names, gpuDisplayLabel(idx))
 	}
 	return datasets, names
 }
@@ -843,64 +1010,37 @@ func autoBounds120(datasets ...[]float64) (*float64, *float64) {
 	return floatPtr(low), floatPtr(high)
 }
 
-// renderChartSVG renders a line chart SVG with a fixed Y-axis range.
-func renderChartSVG(title string, datasets [][]float64, names []string, labels []string, yMin, yMax *float64) ([]byte, error) {
-	n := len(labels)
-	if n == 0 {
-		n = 1
-		labels = []string{""}
+func gpuChartLabelIndices(total, target int) []int {
+	if total <= 0 {
+		return nil
 	}
-	for i := range datasets {
-		if len(datasets[i]) == 0 {
-			datasets[i] = make([]float64, n)
-		}
+	if total == 1 {
+		return []int{0}
 	}
-	// Append global min/avg/max to title.
-	mn, avg, mx := globalStats(datasets)
-	if mx > 0 {
-		title = fmt.Sprintf("%s    ↓%s  ~%s  ↑%s",
-			title,
-			chartLegendNumber(mn),
-			chartLegendNumber(avg),
-			chartLegendNumber(mx),
-		)
+	step := total / target
+	if step < 1 {
+		step = 1
 	}
-	title = sanitizeChartText(title)
-	names = sanitizeChartTexts(names)
-	sparse := sanitizeChartTexts(sparseLabels(labels, 6))
+	var indices []int
+	for i := 0; i < total; i += step {
+		indices = append(indices, i)
+	}
+	if indices[len(indices)-1] != total-1 {
+		indices = append(indices, total-1)
+	}
+	return indices
+}
 
-	opt := gocharts.NewLineChartOptionWithData(datasets)
-	opt.Title = gocharts.TitleOption{Text: title}
-	opt.XAxis.Labels = sparse
-	opt.Legend = gocharts.LegendOption{SeriesNames: names}
-	if chartLegendVisible(len(names)) {
-		opt.Legend.Offset = gocharts.OffsetStr{Top: gocharts.PositionBottom}
-		opt.Legend.OverlayChart = gocharts.Ptr(false)
-	} else {
-		opt.Legend.Show = gocharts.Ptr(false)
-	}
-	opt.Symbol = gocharts.SymbolNone
-	// Right padding: reserve space for the MarkLine label (library recommendation).
-	opt.Padding = gocharts.NewBox(20, 20, 80, 20)
-	if yMin != nil || yMax != nil {
-		opt.YAxis = []gocharts.YAxisOption{chartYAxisOption(yMin, yMax)}
+func chartCanvasHeightForPath(path string, seriesCount int) int {
+	height := chartCanvasHeight(seriesCount)
+	if isGPUChartPath(path) {
+		return height * 2
 	}
+	return height
+}
 
-	// Add a single peak mark line on the series that holds the global maximum.
-	peakIdx, _ := globalPeakSeries(datasets)
-	if peakIdx >= 0 && peakIdx < len(opt.SeriesList) {
-		opt.SeriesList[peakIdx].MarkLine = gocharts.NewMarkLine(gocharts.SeriesMarkTypeMax)
-	}
-
-	p := gocharts.NewPainter(gocharts.PainterOptions{
-		OutputFormat: gocharts.ChartOutputSVG,
-		Width:        1400,
-		Height:       chartCanvasHeight(len(names)),
-	}, gocharts.PainterThemeOption(gocharts.GetTheme("grafana")))
-	if err := p.LineChart(opt); err != nil {
-		return nil, err
-	}
-	return p.Bytes()
+func isGPUChartPath(path string) bool {
+	return strings.HasPrefix(path, "gpu-all-") || strings.HasPrefix(path, "gpu/")
 }
 
 func chartLegendVisible(seriesCount int) bool {
@@ -914,30 +1054,6 @@ func chartCanvasHeight(seriesCount int) int {
 	return 288
 }
 
-func chartYAxisOption(yMin, yMax *float64) gocharts.YAxisOption {
-	return gocharts.YAxisOption{
-		Min:            yMin,
-		Max:            yMax,
-		LabelCount:     11,
-		ValueFormatter: chartYAxisNumber,
-	}
-}
-
-// globalPeakSeries returns the index of the series containing the global maximum
-// value across all datasets, and that maximum value.
-func globalPeakSeries(datasets [][]float64) (idx int, peak float64) {
-	idx = -1
-	for i, ds := range datasets {
-		for _, v := range ds {
-			if v > peak {
-				peak = v
-				idx = i
-			}
-		}
-	}
-	return idx, peak
-}
-
 // globalStats returns min, average, and max across all values in all datasets.
 func globalStats(datasets [][]float64) (mn, avg, mx float64) {
 	var sum float64
@@ -977,21 +1093,6 @@ func sanitizeChartText(s string) string {
 	}, s))
 }
 
-func sanitizeChartTexts(in []string) []string {
-	out := make([]string, len(in))
-	for i, s := range in {
-		out[i] = sanitizeChartText(s)
-	}
-	return out
-}
-
-func safeIdx(s []float64, i int) float64 {
-	if i < len(s) {
-		return s[i]
-	}
-	return 0
-}
-
 func snapshotNamedRings(rings []*namedMetricsRing) ([][]float64, []string, []string) {
 	var datasets [][]float64
 	var names []string
@@ -1078,20 +1179,6 @@ func chartYAxisNumber(v float64) string {
 	return out
 }
 
-func sparseLabels(labels []string, n int) []string {
-	out := make([]string, len(labels))
-	step := len(labels) / n
-	if step < 1 {
-		step = 1
-	}
-	for i, l := range labels {
-		if i%step == 0 {
-			out[i] = l
-		}
-	}
-	return out
-}
-
 func (h *handler) handleAPIMetricsExportCSV(w http.ResponseWriter, r *http.Request) {
 	if h.metricsDB == nil {
 		http.Error(w, "metrics database not available", http.StatusServiceUnavailable)
@@ -1107,6 +1194,11 @@ func (h *handler) handleAPIMetricsExportCSV(w http.ResponseWriter, r *http.Reque
 
 func (h *handler) handleReady(w http.ResponseWriter, r *http.Request) {
 	w.Header().Set("Cache-Control", "no-store")
+	if strings.TrimSpace(h.opts.AuditPath) == "" {
+		w.WriteHeader(http.StatusOK)
+		_, _ = w.Write([]byte("ready"))
+		return
+	}
 	if _, err := os.Stat(h.opts.AuditPath); err != nil {
 		w.WriteHeader(http.StatusServiceUnavailable)
 		_, _ = w.Write([]byte("starting"))
@@ -1120,37 +1212,106 @@ const loadingPageHTML = `<!DOCTYPE html>
 <html lang="en">
 <head>
 <meta charset="UTF-8">
-<title>EASY-BEE</title>
+<title>EASY-BEE — Starting</title>
 <style>
 *{margin:0;padding:0;box-sizing:border-box}
 html,body{height:100%;background:#0f1117;display:flex;align-items:center;justify-content:center;font-family:'Courier New',monospace;color:#e2e8f0}
-.logo{font-size:13px;line-height:1.4;color:#f6c90e;margin-bottom:48px;white-space:pre}
-.spinner{width:48px;height:48px;border:4px solid #2d3748;border-top-color:#f6c90e;border-radius:50%;animation:spin .8s linear infinite;margin:0 auto 24px}
+.wrap{text-align:center;width:420px}
+.logo{font-size:11px;line-height:1.4;color:#f6c90e;margin-bottom:6px;white-space:pre;text-align:left}
+.subtitle{font-size:12px;color:#a0aec0;text-align:left;margin-bottom:24px;padding-left:2px}
+.spinner{width:36px;height:36px;border:3px solid #2d3748;border-top-color:#f6c90e;border-radius:50%;animation:spin .8s linear infinite;margin:0 auto 14px}
+.spinner.hidden{display:none}
 @keyframes spin{to{transform:rotate(360deg)}}
-.status{font-size:14px;color:#a0aec0;letter-spacing:.05em}
+.status{font-size:13px;color:#a0aec0;margin-bottom:20px;min-height:18px}
+table{width:100%;border-collapse:collapse;font-size:12px;margin-bottom:20px;display:none}
+td{padding:3px 6px;text-align:left}
+td:first-child{color:#718096;width:55%}
+.ok{color:#68d391}
+.run{color:#f6c90e}
+.fail{color:#fc8181}
+.dim{color:#4a5568}
+.btn{background:#1a202c;color:#a0aec0;border:1px solid #2d3748;padding:7px 18px;font-size:12px;cursor:pointer;font-family:inherit;display:none}
+.btn:hover{border-color:#718096;color:#e2e8f0}
 </style>
 </head>
 <body>
-<div style="text-align:center">
+<div class="wrap">
   <div class="logo">  ███████╗ █████╗ ███████╗██╗   ██╗      ██████╗ ███████╗███████╗
   ██╔════╝██╔══██╗██╔════╝╚██╗ ██╔╝      ██╔══██╗██╔════╝██╔════╝
   █████╗  ███████║███████╗ ╚████╔╝ █████╗██████╔╝█████╗  █████╗
   ██╔══╝  ██╔══██║╚════██║  ╚██╔╝  ╚════╝██╔══██╗██╔══╝  ██╔══╝
   ███████╗██║  ██║███████║   ██║         ██████╔╝███████╗███████╗
   ╚══════╝╚═╝  ╚═╝╚══════╝   ╚═╝         ╚═════╝ ╚══════╝╚══════╝</div>
-  <div class="spinner"></div>
-  <div class="status" id="s">Starting up...</div>
+  <div class="subtitle">Hardware Audit LiveCD</div>
+  <div class="spinner" id="spin"></div>
+  <div class="status" id="st">Connecting to bee-web...</div>
+  <table id="tbl"></table>
+  <button class="btn" id="btn" onclick="go()">Open app now</button>
 </div>
 <script>
-function probe(){
-  fetch('/api/ready',{cache:'no-store'})
-    .then(function(r){
-      if(r.ok){window.location.replace('/');}
-      else{setTimeout(probe,1000);}
+(function(){
+var gone = false;
+function go(){ if(!gone){gone=true;window.location.replace('/');} }
+
+function icon(s){
+  if(s==='active')   return '<span class="ok">&#9679; active</span>';
+  if(s==='failed')   return '<span class="fail">&#10005; failed</span>';
+  if(s==='activating'||s==='reloading') return '<span class="run">&#9675; starting</span>';
+  if(s==='inactive') return '<span class="dim">&#9675; inactive</span>';
+  return '<span class="dim">'+s+'</span>';
+}
+
+function allSettled(svcs){
+  for(var i=0;i<svcs.length;i++){
+    var s=svcs[i].state;
+    if(s!=='active'&&s!=='failed'&&s!=='inactive') return false;
+  }
+  return true;
+}
+
+var pollTimer=null;
+
+function pollServices(){
+  fetch('/api/services',{cache:'no-store'})
+    .then(function(r){return r.json();})
+    .then(function(svcs){
+      if(!svcs||!svcs.length) return;
+      var tbl=document.getElementById('tbl');
+      tbl.style.display='';
+      var html='';
+      for(var i=0;i<svcs.length;i++)
+        html+='<tr><td>'+svcs[i].name+'</td><td>'+icon(svcs[i].state)+'</td></tr>';
+      tbl.innerHTML=html;
+      if(allSettled(svcs)){
+        clearInterval(pollTimer);
+        document.getElementById('spin').className='spinner hidden';
+        document.getElementById('st').textContent='Ready \u2014 opening...';
+        setTimeout(go,800);
+      }
     })
-    .catch(function(){setTimeout(probe,1000);});
+    .catch(function(){});
+}
+
+function probe(){
+  fetch('/healthz',{cache:'no-store'})
+    .then(function(r){
+      if(r.ok){
+        document.getElementById('st').textContent='bee-web running \u2014 checking services...';
+        document.getElementById('btn').style.display='';
+        pollServices();
+        pollTimer=setInterval(pollServices,1500);
+      } else {
+        document.getElementById('st').textContent='bee-web starting (status '+r.status+')...';
+        setTimeout(probe,500);
+      }
+    })
+    .catch(function(){
+      document.getElementById('st').textContent='Waiting for bee-web to start...';
+      setTimeout(probe,500);
+    });
 }
 probe();
+})();
 </script>
 </body>
 </html>`

audit/internal/webui/server_test.go

@@ -1,6 +1,7 @@
 package webui
 
 import (
+	"encoding/json"
 	"net/http"
 	"net/http/httptest"
 	"os"
@@ -34,6 +35,49 @@ func TestChartLegendNumber(t *testing.T) {
 	}
 }
 
+func TestRecoverMiddlewareReturns500OnPanic(t *testing.T) {
+	handler := recoverMiddleware(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		panic("boom")
+	}))
+	rec := httptest.NewRecorder()
+	req := httptest.NewRequest(http.MethodGet, "/panic", nil)
+
+	handler.ServeHTTP(rec, req)
+
+	if rec.Code != http.StatusInternalServerError {
+		t.Fatalf("status=%d want %d", rec.Code, http.StatusInternalServerError)
+	}
+	if !strings.Contains(rec.Body.String(), "internal server error") {
+		t.Fatalf("body=%q", rec.Body.String())
+	}
+}
+
+func TestRecoverMiddlewarePreservesStreamingInterfaces(t *testing.T) {
+	handler := recoverMiddleware(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if !sseStart(w) {
+			return
+		}
+		if !sseWrite(w, "tick", "ok") {
+			t.Fatal("expected sse write to succeed")
+		}
+	}))
+	rec := httptest.NewRecorder()
+	req := httptest.NewRequest(http.MethodGet, "/stream", nil)
+
+	handler.ServeHTTP(rec, req)
+
+	if rec.Code != http.StatusOK {
+		t.Fatalf("status=%d body=%s", rec.Code, rec.Body.String())
+	}
+	if got := rec.Header().Get("Content-Type"); got != "text/event-stream" {
+		t.Fatalf("content-type=%q", got)
+	}
+	body := rec.Body.String()
+	if !strings.Contains(body, "event: tick\n") || !strings.Contains(body, "data: ok\n\n") {
+		t.Fatalf("body=%q", body)
+	}
+}
+
 func TestChartDataFromSamplesUsesFullHistory(t *testing.T) {
 	samples := []platform.LiveMetricSample{
 		{
@@ -136,6 +180,39 @@ func TestChartDataFromSamplesKeepsStableGPUSeriesOrder(t *testing.T) {
 	}
 }
 
+func TestChartDataFromSamplesIncludesGPUClockCharts(t *testing.T) {
+	samples := []platform.LiveMetricSample{
+		{
+			Timestamp: time.Now().Add(-2 * time.Minute),
+			GPUs: []platform.GPUMetricRow{
+				{GPUIndex: 0, ClockMHz: 1400},
+				{GPUIndex: 3, ClockMHz: 1500},
+			},
+		},
+		{
+			Timestamp: time.Now().Add(-1 * time.Minute),
+			GPUs: []platform.GPUMetricRow{
+				{GPUIndex: 0, ClockMHz: 1410},
+				{GPUIndex: 3, ClockMHz: 1510},
+			},
+		},
+	}
+
+	datasets, names, _, title, _, _, ok := chartDataFromSamples("gpu-all-clock", samples)
+	if !ok {
+		t.Fatal("gpu-all-clock returned ok=false")
+	}
+	if title != "GPU Core Clock" {
+		t.Fatalf("title=%q", title)
+	}
+	if len(names) != 2 || names[0] != "GPU 0" || names[1] != "GPU 3" {
+		t.Fatalf("names=%v", names)
+	}
+	if got := datasets[1][1]; got != 1510 {
+		t.Fatalf("GPU 3 core clock=%v want 1510", got)
+	}
+}
+
 func TestNormalizePowerSeriesHoldsLastPositive(t *testing.T) {
 	got := normalizePowerSeries([]float64{0, 480, 0, 0, 510, 0})
 	want := []float64{0, 480, 480, 480, 510, 510}
@@ -157,6 +234,21 @@ func TestRenderMetricsUsesBufferedChartRefresh(t *testing.T) {
 	if !strings.Contains(body, "el.dataset.loading === '1'") {
 		t.Fatalf("metrics page should avoid overlapping chart reloads: %s", body)
 	}
+	if !strings.Contains(body, `id="gpu-metrics-section" style="display:none`) {
+		t.Fatalf("metrics page should keep gpu charts in a hidden dedicated section until GPUs are detected: %s", body)
+	}
+	if !strings.Contains(body, `id="gpu-chart-toggle"`) {
+		t.Fatalf("metrics page should render GPU chart mode toggle: %s", body)
+	}
+	if !strings.Contains(body, `/api/metrics/chart/gpu-all-clock.svg`) {
+		t.Fatalf("metrics page should include GPU core clock chart: %s", body)
+	}
+	if strings.Contains(body, `/api/metrics/chart/gpu-all-memclock.svg`) {
+		t.Fatalf("metrics page should not include GPU memory clock chart: %s", body)
+	}
+	if !strings.Contains(body, `renderGPUOverviewCards(indices, names)`) {
+		t.Fatalf("metrics page should build per-GPU chart cards dynamically: %s", body)
+	}
 }
 
 func TestChartLegendVisible(t *testing.T) {
@@ -199,6 +291,124 @@ func TestChartCanvasHeight(t *testing.T) {
 	}
 }
 
+func TestChartTimelineSegmentsForRangeMergesActiveSpansAndIdleGaps(t *testing.T) {
+	start := time.Date(2026, 4, 5, 12, 0, 0, 0, time.UTC)
+	end := start.Add(10 * time.Minute)
+	taskWindow := func(offsetStart, offsetEnd time.Duration) Task {
+		s := start.Add(offsetStart)
+		e := start.Add(offsetEnd)
+		return Task{
+			Name:      "task",
+			Status:    TaskDone,
+			StartedAt: &s,
+			DoneAt:    &e,
+		}
+	}
+	segments := chartTimelineSegmentsForRange(start, end, end, []Task{
+		taskWindow(1*time.Minute, 3*time.Minute),
+		taskWindow(2*time.Minute, 5*time.Minute),
+		taskWindow(7*time.Minute, 8*time.Minute),
+	})
+	if len(segments) != 5 {
+		t.Fatalf("segments=%d want 5: %#v", len(segments), segments)
+	}
+	wantActive := []bool{false, true, false, true, false}
+	wantMinutes := [][2]int{{0, 1}, {1, 5}, {5, 7}, {7, 8}, {8, 10}}
+	for i, segment := range segments {
+		if segment.Active != wantActive[i] {
+			t.Fatalf("segment[%d].Active=%v want %v", i, segment.Active, wantActive[i])
+		}
+		if got := int(segment.Start.Sub(start).Minutes()); got != wantMinutes[i][0] {
+			t.Fatalf("segment[%d] start=%d want %d", i, got, wantMinutes[i][0])
+		}
+		if got := int(segment.End.Sub(start).Minutes()); got != wantMinutes[i][1] {
+			t.Fatalf("segment[%d] end=%d want %d", i, got, wantMinutes[i][1])
+		}
+	}
+}
+
+func TestRenderMetricChartSVGIncludesTimelineOverlay(t *testing.T) {
+	start := time.Date(2026, 4, 5, 12, 0, 0, 0, time.UTC)
+	labels := []string{"12:00", "12:01", "12:02"}
+	times := []time.Time{start, start.Add(time.Minute), start.Add(2 * time.Minute)}
+	svg, err := renderMetricChartSVG(
+		"System Power",
+		labels,
+		times,
+		[][]float64{{300, 320, 310}},
+		[]string{"Power W"},
+		floatPtr(0),
+		floatPtr(400),
+		360,
+		[]chartTimelineSegment{
+			{Start: start, End: start.Add(time.Minute), Active: false},
+			{Start: start.Add(time.Minute), End: start.Add(2 * time.Minute), Active: true},
+		},
+	)
+	if err != nil {
+		t.Fatal(err)
+	}
+	body := string(svg)
+	if !strings.Contains(body, `data-role="timeline-overlay"`) {
+		t.Fatalf("svg missing timeline overlay: %s", body)
+	}
+	if !strings.Contains(body, `opacity="0.10"`) {
+		t.Fatalf("svg missing idle overlay opacity: %s", body)
+	}
+	if !strings.Contains(body, `System Power`) {
+		t.Fatalf("svg missing chart title: %s", body)
+	}
+}
+
+func TestHandleMetricsChartSVGRendersCustomSVG(t *testing.T) {
+	dir := t.TempDir()
+	db, err := openMetricsDB(filepath.Join(dir, "metrics.db"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	t.Cleanup(func() { _ = db.db.Close() })
+
+	start := time.Date(2026, 4, 5, 12, 0, 0, 0, time.UTC)
+	for i, sample := range []platform.LiveMetricSample{
+		{Timestamp: start, PowerW: 300},
+		{Timestamp: start.Add(time.Minute), PowerW: 320},
+		{Timestamp: start.Add(2 * time.Minute), PowerW: 310},
+	} {
+		if err := db.Write(sample); err != nil {
+			t.Fatalf("write sample %d: %v", i, err)
+		}
+	}
+
+	globalQueue.mu.Lock()
+	prevTasks := globalQueue.tasks
+	s := start.Add(30 * time.Second)
+	e := start.Add(90 * time.Second)
+	globalQueue.tasks = []*Task{{Name: "Burn", Status: TaskDone, StartedAt: &s, DoneAt: &e}}
+	globalQueue.mu.Unlock()
+	t.Cleanup(func() {
+		globalQueue.mu.Lock()
+		globalQueue.tasks = prevTasks
+		globalQueue.mu.Unlock()
+	})
+
+	h := &handler{opts: HandlerOptions{ExportDir: dir}, metricsDB: db}
+
+	rec := httptest.NewRecorder()
+	req := httptest.NewRequest(http.MethodGet, "/api/metrics/chart/server-power.svg", nil)
+	h.handleMetricsChartSVG(rec, req)
+
+	if rec.Code != http.StatusOK {
+		t.Fatalf("status=%d body=%s", rec.Code, rec.Body.String())
+	}
+	body := rec.Body.String()
+	if !strings.Contains(body, `data-role="timeline-overlay"`) {
+		t.Fatalf("custom svg response missing timeline overlay: %s", body)
+	}
+	if !strings.Contains(body, `stroke-linecap="round"`) {
+		t.Fatalf("custom svg response missing custom polyline styling: %s", body)
+	}
+}
+
 func TestNormalizeFanSeriesHoldsLastPositive(t *testing.T) {
 	got := normalizeFanSeries([]float64{4200, 0, 0, 4300, 0})
 	want := []float64{4200, 4200, 4200, 4300, 4300}
@@ -212,21 +422,6 @@ func TestNormalizeFanSeriesHoldsLastPositive(t *testing.T) {
 	}
 }
 
-func TestChartYAxisOption(t *testing.T) {
-	min := floatPtr(0)
-	max := floatPtr(100)
-	opt := chartYAxisOption(min, max)
-	if opt.Min != min || opt.Max != max {
-		t.Fatalf("chartYAxisOption min/max mismatch: %#v", opt)
-	}
-	if opt.LabelCount != 11 {
-		t.Fatalf("chartYAxisOption labelCount=%d want 11", opt.LabelCount)
-	}
-	if got := opt.ValueFormatter(1000); got != "1к" {
-		t.Fatalf("chartYAxisOption formatter(1000)=%q want 1к", got)
-	}
-}
-
 func TestSnapshotFanRingsUsesTimelineLabels(t *testing.T) {
 	r1 := newMetricsRing(4)
 	r2 := newMetricsRing(4)
@@ -275,9 +470,10 @@ func TestRootRendersDashboard(t *testing.T) {
 	}
 
 	handler := NewHandler(HandlerOptions{
-		Title:     "Bee Hardware Audit",
-		AuditPath: path,
-		ExportDir: exportDir,
+		Title:      "Bee Hardware Audit",
+		BuildLabel: "1.2.3",
+		AuditPath:  path,
+		ExportDir:  exportDir,
 	})
 
 	first := httptest.NewRecorder()
@@ -292,6 +488,11 @@ func TestRootRendersDashboard(t *testing.T) {
 	if !strings.Contains(first.Body.String(), `/viewer`) {
 		t.Fatalf("first body missing viewer link: %s", first.Body.String())
 	}
+	versionIdx := strings.Index(first.Body.String(), `Version 1.2.3`)
+	navIdx := strings.Index(first.Body.String(), `href="/"`)
+	if versionIdx == -1 || navIdx == -1 || versionIdx > navIdx {
+		t.Fatalf("version should render near top of sidebar before nav links: %s", first.Body.String())
+	}
 	if got := first.Header().Get("Cache-Control"); got != "no-store" {
 		t.Fatalf("first cache-control=%q", got)
 	}
@@ -329,7 +530,7 @@ func TestRootShowsRunAuditButtonWhenSnapshotMissing(t *testing.T) {
 		t.Fatalf("status=%d", rec.Code)
 	}
 	body := rec.Body.String()
-	if !strings.Contains(body, `Run Audit`) {
+	if !strings.Contains(body, `onclick="auditModalRun()">Run audit</button>`) {
 		t.Fatalf("dashboard missing run audit button: %s", body)
 	}
 	if strings.Contains(body, `No audit data`) {
@@ -337,6 +538,18 @@ func TestRootShowsRunAuditButtonWhenSnapshotMissing(t *testing.T) {
 	}
 }
 
+func TestReadyIsOKWhenAuditPathIsUnset(t *testing.T) {
+	handler := NewHandler(HandlerOptions{})
+	rec := httptest.NewRecorder()
+	handler.ServeHTTP(rec, httptest.NewRequest(http.MethodGet, "/api/ready", nil))
+	if rec.Code != http.StatusOK {
+		t.Fatalf("status=%d body=%s", rec.Code, rec.Body.String())
+	}
+	if strings.TrimSpace(rec.Body.String()) != "ready" {
+		t.Fatalf("body=%q want ready", rec.Body.String())
+	}
+}
+
 func TestAuditPageRendersViewerFrameAndActions(t *testing.T) {
 	dir := t.TempDir()
 	path := filepath.Join(dir, "audit.json")
@@ -359,7 +572,7 @@ func TestAuditPageRendersViewerFrameAndActions(t *testing.T) {
 	}
 }
 
-func TestTasksPageRendersLogModalAndPaginationControls(t *testing.T) {
+func TestTasksPageRendersOpenLinksAndPaginationControls(t *testing.T) {
 	handler := NewHandler(HandlerOptions{})
 	rec := httptest.NewRecorder()
 	handler.ServeHTTP(rec, httptest.NewRequest(http.MethodGet, "/tasks", nil))
@@ -367,8 +580,8 @@ func TestTasksPageRendersLogModalAndPaginationControls(t *testing.T) {
 		t.Fatalf("status=%d", rec.Code)
 	}
 	body := rec.Body.String()
-	if !strings.Contains(body, `id="task-log-overlay"`) {
-		t.Fatalf("tasks page missing log modal overlay: %s", body)
+	if !strings.Contains(body, `Open a task to view its saved logs and charts.`) {
+		t.Fatalf("tasks page missing task report hint: %s", body)
 	}
 	if !strings.Contains(body, `_taskPageSize = 50`) {
 		t.Fatalf("tasks page missing pagination size config: %s", body)
@@ -378,7 +591,7 @@ func TestTasksPageRendersLogModalAndPaginationControls(t *testing.T) {
 	}
 }
 
-func TestToolsPageRendersRestartGPUDriversButton(t *testing.T) {
+func TestToolsPageRendersNvidiaSelfHealSection(t *testing.T) {
 	handler := NewHandler(HandlerOptions{})
 	rec := httptest.NewRecorder()
 	handler.ServeHTTP(rec, httptest.NewRequest(http.MethodGet, "/tools", nil))
@@ -386,15 +599,307 @@ func TestToolsPageRendersRestartGPUDriversButton(t *testing.T) {
 		t.Fatalf("status=%d", rec.Code)
 	}
 	body := rec.Body.String()
+	if !strings.Contains(body, `NVIDIA Self Heal`) {
+		t.Fatalf("tools page missing nvidia self heal section: %s", body)
+	}
 	if !strings.Contains(body, `Restart GPU Drivers`) {
 		t.Fatalf("tools page missing restart gpu drivers button: %s", body)
 	}
-	if !strings.Contains(body, `svcAction('bee-nvidia', 'restart')`) {
-		t.Fatalf("tools page missing bee-nvidia restart action: %s", body)
+	if !strings.Contains(body, `nvidiaRestartDrivers()`) {
+		t.Fatalf("tools page missing nvidiaRestartDrivers action: %s", body)
+	}
+	if !strings.Contains(body, `/api/gpu/nvidia-status`) {
+		t.Fatalf("tools page missing nvidia status api usage: %s", body)
+	}
+	if !strings.Contains(body, `nvidiaResetGPU(`) {
+		t.Fatalf("tools page missing nvidiaResetGPU action: %s", body)
 	}
 	if !strings.Contains(body, `id="boot-source-text"`) {
 		t.Fatalf("tools page missing boot source field: %s", body)
 	}
+	if !strings.Contains(body, `Export to USB`) {
+		t.Fatalf("tools page missing export to usb section: %s", body)
+	}
+	if !strings.Contains(body, `Support Bundle</button>`) {
+		t.Fatalf("tools page missing support bundle usb button: %s", body)
+	}
+}
+
+func TestBenchmarkPageRendersGPUSelectionControls(t *testing.T) {
+	handler := NewHandler(HandlerOptions{})
+	rec := httptest.NewRecorder()
+	handler.ServeHTTP(rec, httptest.NewRequest(http.MethodGet, "/benchmark", nil))
+	if rec.Code != http.StatusOK {
+		t.Fatalf("status=%d", rec.Code)
+	}
+	body := rec.Body.String()
+	for _, needle := range []string{
+		`href="/benchmark"`,
+		`id="benchmark-gpu-list"`,
+		`/api/gpu/nvidia`,
+		`/api/benchmark/nvidia/run`,
+		`benchmark-run-nccl`,
+	} {
+		if !strings.Contains(body, needle) {
+			t.Fatalf("benchmark page missing %q: %s", needle, body)
+		}
+	}
+}
+
+func TestBenchmarkPageRendersSavedResultsTable(t *testing.T) {
+	dir := t.TempDir()
+	exportDir := filepath.Join(dir, "export")
+	runDir := filepath.Join(exportDir, "bee-benchmark", "gpu-benchmark-20260406-120000")
+	if err := os.MkdirAll(runDir, 0755); err != nil {
+		t.Fatal(err)
+	}
+	result := platform.NvidiaBenchmarkResult{
+		GeneratedAt:      time.Date(2026, time.April, 6, 12, 0, 0, 0, time.UTC),
+		BenchmarkProfile: "standard",
+		OverallStatus:    "OK",
+		GPUs: []platform.BenchmarkGPUResult{
+			{
+				Index: 0,
+				Name:  "NVIDIA H100 PCIe",
+				Scores: platform.BenchmarkScorecard{
+					CompositeScore: 1176.25,
+				},
+			},
+			{
+				Index: 1,
+				Name:  "NVIDIA H100 PCIe",
+				Scores: platform.BenchmarkScorecard{
+					CompositeScore: 1168.50,
+				},
+			},
+		},
+	}
+	raw, err := json.Marshal(result)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(filepath.Join(runDir, "result.json"), raw, 0644); err != nil {
+		t.Fatal(err)
+	}
+
+	handler := NewHandler(HandlerOptions{ExportDir: exportDir})
+	rec := httptest.NewRecorder()
+	handler.ServeHTTP(rec, httptest.NewRequest(http.MethodGet, "/benchmark", nil))
+	if rec.Code != http.StatusOK {
+		t.Fatalf("status=%d", rec.Code)
+	}
+	body := rec.Body.String()
+	wantTime := result.GeneratedAt.Local().Format("2006-01-02 15:04:05")
+	for _, needle := range []string{
+		`Benchmark Results`,
+		`Composite score by saved benchmark run and GPU.`,
+		`GPU #0 — NVIDIA H100 PCIe`,
+		`GPU #1 — NVIDIA H100 PCIe`,
+		`#1`,
+		wantTime,
+		`1176.25`,
+		`1168.50`,
+	} {
+		if !strings.Contains(body, needle) {
+			t.Fatalf("benchmark page missing %q: %s", needle, body)
+		}
+	}
+}
+
+func TestValidatePageRendersNvidiaTargetedStressCard(t *testing.T) {
+	handler := NewHandler(HandlerOptions{})
+	rec := httptest.NewRecorder()
+	handler.ServeHTTP(rec, httptest.NewRequest(http.MethodGet, "/validate", nil))
+	if rec.Code != http.StatusOK {
+		t.Fatalf("status=%d", rec.Code)
+	}
+	body := rec.Body.String()
+	for _, needle := range []string{
+		`NVIDIA GPU Targeted Stress`,
+		`nvidia-targeted-stress`,
+		`controlled NVIDIA DCGM load`,
+		`<code>dcgmi diag targeted_stress</code>`,
+		`NVIDIA GPU Selection`,
+		`All NVIDIA validate tasks use only the GPUs selected here.`,
+		`Select All`,
+		`id="sat-gpu-list"`,
+	} {
+		if !strings.Contains(body, needle) {
+			t.Fatalf("validate page missing %q: %s", needle, body)
+		}
+	}
+}
+
+func TestBurnPageRendersGoalBasedNVIDIACards(t *testing.T) {
+	handler := NewHandler(HandlerOptions{})
+	rec := httptest.NewRecorder()
+	handler.ServeHTTP(rec, httptest.NewRequest(http.MethodGet, "/burn", nil))
+	if rec.Code != http.StatusOK {
+		t.Fatalf("status=%d", rec.Code)
+	}
+	body := rec.Body.String()
+	for _, needle := range []string{
+		`NVIDIA Max Compute Load`,
+		`dcgmproftester`,
+		`NCCL`,
+		`Validate → Stress mode`,
+		`id="burn-gpu-list"`,
+	} {
+		if !strings.Contains(body, needle) {
+			t.Fatalf("burn page missing %q: %s", needle, body)
+		}
+	}
+}
+
+func TestTaskDetailPageRendersSavedReport(t *testing.T) {
+	dir := t.TempDir()
+	exportDir := filepath.Join(dir, "export")
+	reportDir := filepath.Join(exportDir, "tasks", "task-1_cpu_sat_done")
+	if err := os.MkdirAll(reportDir, 0755); err != nil {
+		t.Fatal(err)
+	}
+	reportPath := filepath.Join(reportDir, "report.html")
+	if err := os.WriteFile(reportPath, []byte(`<div class="card"><div class="card-head">Task Report</div><div class="card-body">saved report</div></div>`), 0644); err != nil {
+		t.Fatal(err)
+	}
+
+	globalQueue.mu.Lock()
+	origTasks := globalQueue.tasks
+	globalQueue.tasks = []*Task{{
+		ID:             "task-1",
+		Name:           "CPU SAT",
+		Target:         "cpu",
+		Status:         TaskDone,
+		CreatedAt:      time.Now(),
+		ArtifactsDir:   reportDir,
+		ReportHTMLPath: reportPath,
+	}}
+	globalQueue.mu.Unlock()
+	t.Cleanup(func() {
+		globalQueue.mu.Lock()
+		globalQueue.tasks = origTasks
+		globalQueue.mu.Unlock()
+	})
+
+	handler := NewHandler(HandlerOptions{Title: "Bee Hardware Audit", ExportDir: exportDir})
+
+	rec := httptest.NewRecorder()
+	handler.ServeHTTP(rec, httptest.NewRequest(http.MethodGet, "/tasks/task-1", nil))
+	if rec.Code != http.StatusOK {
+		t.Fatalf("status=%d", rec.Code)
+	}
+	body := rec.Body.String()
+	if !strings.Contains(body, `saved report`) {
+		t.Fatalf("task detail page missing saved report: %s", body)
+	}
+	if !strings.Contains(body, `Back to Tasks`) {
+		t.Fatalf("task detail page missing back link: %s", body)
+	}
+}
+
+func TestTaskDetailPageRendersCancelForRunningTask(t *testing.T) {
+	globalQueue.mu.Lock()
+	origTasks := globalQueue.tasks
+	globalQueue.tasks = []*Task{{
+		ID:        "task-live-1",
+		Name:      "CPU SAT",
+		Target:    "cpu",
+		Status:    TaskRunning,
+		CreatedAt: time.Now(),
+	}}
+	globalQueue.mu.Unlock()
+	t.Cleanup(func() {
+		globalQueue.mu.Lock()
+		globalQueue.tasks = origTasks
+		globalQueue.mu.Unlock()
+	})
+
+	handler := NewHandler(HandlerOptions{Title: "Bee Hardware Audit"})
+	rec := httptest.NewRecorder()
+	handler.ServeHTTP(rec, httptest.NewRequest(http.MethodGet, "/tasks/task-live-1", nil))
+	if rec.Code != http.StatusOK {
+		t.Fatalf("status=%d", rec.Code)
+	}
+	body := rec.Body.String()
+	if !strings.Contains(body, `Cancel</button>`) {
+		t.Fatalf("task detail page missing cancel button: %s", body)
+	}
+	if !strings.Contains(body, `function cancelTaskDetail(id)`) {
+		t.Fatalf("task detail page missing cancel handler: %s", body)
+	}
+	if !strings.Contains(body, `/api/tasks/' + id + '/cancel`) {
+		t.Fatalf("task detail page missing cancel endpoint: %s", body)
+	}
+	if !strings.Contains(body, `id="task-live-charts"`) {
+		t.Fatalf("task detail page missing live charts container: %s", body)
+	}
+	if !strings.Contains(body, `/api/tasks/' + taskId + '/charts`) {
+		t.Fatalf("task detail page missing live charts index endpoint: %s", body)
+	}
+}
+
+func TestTaskChartSVGUsesTaskTimeWindow(t *testing.T) {
+	dir := t.TempDir()
+	metricsPath := filepath.Join(dir, "metrics.db")
+	prevMetricsPath := taskReportMetricsDBPath
+	taskReportMetricsDBPath = metricsPath
+	t.Cleanup(func() { taskReportMetricsDBPath = prevMetricsPath })
+
+	db, err := openMetricsDB(metricsPath)
+	if err != nil {
+		t.Fatalf("openMetricsDB: %v", err)
+	}
+	base := time.Now().UTC()
+	samples := []platform.LiveMetricSample{
+		{Timestamp: base.Add(-3 * time.Minute), PowerW: 100},
+		{Timestamp: base.Add(-2 * time.Minute), PowerW: 200},
+		{Timestamp: base.Add(-1 * time.Minute), PowerW: 300},
+	}
+	for _, sample := range samples {
+		if err := db.Write(sample); err != nil {
+			t.Fatalf("Write: %v", err)
+		}
+	}
+	_ = db.Close()
+
+	started := base.Add(-2*time.Minute - 5*time.Second)
+	done := base.Add(-1*time.Minute + 5*time.Second)
+	globalQueue.mu.Lock()
+	origTasks := globalQueue.tasks
+	globalQueue.tasks = []*Task{{
+		ID:        "task-chart-1",
+		Name:      "Power Window",
+		Target:    "cpu",
+		Status:    TaskDone,
+		CreatedAt: started.Add(-10 * time.Second),
+		StartedAt: &started,
+		DoneAt:    &done,
+	}}
+	globalQueue.mu.Unlock()
+	t.Cleanup(func() {
+		globalQueue.mu.Lock()
+		globalQueue.tasks = origTasks
+		globalQueue.mu.Unlock()
+	})
+
+	handler := NewHandler(HandlerOptions{Title: "Bee Hardware Audit"})
+	req := httptest.NewRequest(http.MethodGet, "/api/tasks/task-chart-1/chart/server-power.svg", nil)
+	req.SetPathValue("id", "task-chart-1")
+	rec := httptest.NewRecorder()
+	handler.ServeHTTP(rec, req)
+	if rec.Code != http.StatusOK {
+		t.Fatalf("status=%d body=%s", rec.Code, rec.Body.String())
+	}
+	body := rec.Body.String()
+	if !strings.Contains(body, "System Power") {
+		t.Fatalf("task chart missing expected title: %s", body)
+	}
+	if !strings.Contains(body, "min 200") {
+		t.Fatalf("task chart stats should start from in-window sample: %s", body)
+	}
+	if strings.Contains(body, "min 100") {
+		t.Fatalf("task chart should not include pre-task sample in stats: %s", body)
+	}
 }
 
 func TestViewerRendersLatestSnapshot(t *testing.T) {
@@ -518,3 +1023,98 @@ func TestRuntimeHealthEndpointReturnsJSON(t *testing.T) {
 		t.Fatalf("body=%q want %q", strings.TrimSpace(rec.Body.String()), body)
 	}
 }
+
+func TestDashboardRendersRuntimeHealthTable(t *testing.T) {
+	dir := t.TempDir()
+	path := filepath.Join(dir, "audit.json")
+	exportDir := filepath.Join(dir, "export")
+	if err := os.MkdirAll(exportDir, 0755); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(path, []byte(`{"collected_at":"2026-03-15T00:00:00Z","hardware":{"board":{"serial_number":"SERIAL-1"}}}`), 0644); err != nil {
+		t.Fatal(err)
+	}
+	health := `{
+  "status":"PARTIAL",
+  "checked_at":"2026-03-16T10:00:00Z",
+  "export_dir":"/tmp/export",
+  "driver_ready":true,
+  "cuda_ready":false,
+  "network_status":"PARTIAL",
+  "issues":[
+    {"code":"dhcp_partial","description":"At least one interface did not obtain IPv4 connectivity."},
+    {"code":"cuda_runtime_not_ready","description":"CUDA runtime is not ready for GPU SAT."}
+  ],
+  "tools":[
+    {"name":"dmidecode","ok":true},
+    {"name":"nvidia-smi","ok":false}
+  ],
+  "services":[
+    {"name":"bee-web","status":"active"},
+    {"name":"bee-nvidia","status":"inactive"}
+  ]
+}`
+	if err := os.WriteFile(filepath.Join(exportDir, "runtime-health.json"), []byte(health), 0644); err != nil {
+		t.Fatal(err)
+	}
+	componentStatus := `[
+  {
+    "component_key":"cpu:all",
+    "status":"Warning",
+    "error_summary":"cpu SAT: FAILED",
+    "history":[{"at":"2026-03-16T10:00:00Z","status":"Warning","source":"sat:cpu","detail":"cpu SAT: FAILED"}]
+  },
+  {
+    "component_key":"memory:all",
+    "status":"OK",
+    "history":[{"at":"2026-03-16T10:01:00Z","status":"OK","source":"sat:memory","detail":"memory SAT: OK"}]
+  },
+  {
+    "component_key":"storage:nvme0n1",
+    "status":"Critical",
+    "error_summary":"storage SAT: FAILED",
+    "history":[{"at":"2026-03-16T10:02:00Z","status":"Critical","source":"sat:storage","detail":"storage SAT: FAILED"}]
+  },
+  {
+    "component_key":"pcie:gpu:nvidia",
+    "status":"Warning",
+    "error_summary":"nvidia SAT: FAILED",
+    "history":[{"at":"2026-03-16T10:03:00Z","status":"Warning","source":"sat:nvidia","detail":"nvidia SAT: FAILED"}]
+  }
+]`
+	if err := os.WriteFile(filepath.Join(exportDir, "component-status.json"), []byte(componentStatus), 0644); err != nil {
+		t.Fatal(err)
+	}
+
+	handler := NewHandler(HandlerOptions{AuditPath: path, ExportDir: exportDir})
+	rec := httptest.NewRecorder()
+	handler.ServeHTTP(rec, httptest.NewRequest(http.MethodGet, "/", nil))
+	if rec.Code != http.StatusOK {
+		t.Fatalf("status=%d body=%s", rec.Code, rec.Body.String())
+	}
+	body := rec.Body.String()
+	for _, needle := range []string{
+		`Runtime Health`,
+		`<th>Check</th><th>Status</th><th>Source</th><th>Issue</th>`,
+		`Export Directory`,
+		`Network`,
+		`NVIDIA/AMD Driver`,
+		`CUDA / ROCm`,
+		`Required Utilities`,
+		`Bee Services`,
+		`<td>CPU</td>`,
+		`<td>Memory</td>`,
+		`<td>Storage</td>`,
+		`<td>GPU</td>`,
+		`CUDA runtime is not ready for GPU SAT.`,
+		`Missing: nvidia-smi`,
+		`bee-nvidia=inactive`,
+		`cpu SAT: FAILED`,
+		`storage SAT: FAILED`,
+		`sat:nvidia`,
+	} {
+		if !strings.Contains(body, needle) {
+			t.Fatalf("dashboard missing %q: %s", needle, body)
+		}
+	}
+}

audit/internal/webui/stability.go

@@ -0,0 +1,42 @@
+package webui
+
+import (
+	"fmt"
+	"log/slog"
+	"runtime/debug"
+	"time"
+)
+
+func goRecoverLoop(name string, restartDelay time.Duration, fn func()) {
+	go func() {
+		for {
+			if !runRecoverable(name, fn) {
+				return
+			}
+			if restartDelay > 0 {
+				time.Sleep(restartDelay)
+			}
+		}
+	}()
+}
+
+func goRecoverOnce(name string, fn func()) {
+	go func() {
+		_ = runRecoverable(name, fn)
+	}()
+}
+
+func runRecoverable(name string, fn func()) (panicked bool) {
+	defer func() {
+		if rec := recover(); rec != nil {
+			panicked = true
+			slog.Error("recovered panic",
+				"component", name,
+				"panic", fmt.Sprint(rec),
+				"stack", string(debug.Stack()),
+			)
+		}
+	}()
+	fn()
+	return false
+}

audit/internal/webui/task_page.go

@@ -0,0 +1,267 @@
+package webui
+
+import (
+	"encoding/json"
+	"fmt"
+	"html"
+	"net/http"
+	"os"
+	"strings"
+	"time"
+
+	"bee/audit/internal/platform"
+)
+
+func (h *handler) handleTaskPage(w http.ResponseWriter, r *http.Request) {
+	id := r.PathValue("id")
+	task, ok := globalQueue.findByID(id)
+	if !ok {
+		http.NotFound(w, r)
+		return
+	}
+	snapshot := *task
+	body := renderTaskDetailPage(h.opts, snapshot)
+	w.Header().Set("Cache-Control", "no-store")
+	w.Header().Set("Content-Type", "text/html; charset=utf-8")
+	_, _ = w.Write([]byte(body))
+}
+
+func (h *handler) handleAPITaskChartsIndex(w http.ResponseWriter, r *http.Request) {
+	task, samples, _, _, ok := h.taskSamplesForRequest(r)
+	if !ok {
+		http.NotFound(w, r)
+		return
+	}
+	type taskChartIndexEntry struct {
+		Title string `json:"title"`
+		File  string `json:"file"`
+	}
+	entries := make([]taskChartIndexEntry, 0)
+	for _, spec := range taskChartSpecsForSamples(samples) {
+		title, _, ok := renderTaskChartSVG(spec.Path, samples, taskTimelineForTask(task))
+		if !ok {
+			continue
+		}
+		entries = append(entries, taskChartIndexEntry{Title: title, File: spec.File})
+	}
+	w.Header().Set("Cache-Control", "no-store")
+	w.Header().Set("Content-Type", "application/json; charset=utf-8")
+	_ = json.NewEncoder(w).Encode(entries)
+}
+
+func (h *handler) handleAPITaskChartSVG(w http.ResponseWriter, r *http.Request) {
+	task, samples, _, _, ok := h.taskSamplesForRequest(r)
+	if !ok {
+		http.NotFound(w, r)
+		return
+	}
+	file := strings.TrimPrefix(r.URL.Path, "/api/tasks/"+task.ID+"/chart/")
+	path, ok := taskChartPathFromFile(file)
+	if !ok {
+		http.NotFound(w, r)
+		return
+	}
+	title, buf, hasData := renderTaskChartSVG(path, samples, taskTimelineForTask(task))
+	if !hasData || len(buf) == 0 || strings.TrimSpace(title) == "" {
+		http.Error(w, "metrics history unavailable", http.StatusServiceUnavailable)
+		return
+	}
+	w.Header().Set("Content-Type", "image/svg+xml")
+	w.Header().Set("Cache-Control", "no-store")
+	_, _ = w.Write(buf)
+}
+
+func renderTaskDetailPage(opts HandlerOptions, task Task) string {
+	title := task.Name
+	if strings.TrimSpace(title) == "" {
+		title = task.ID
+	}
+	var body strings.Builder
+	body.WriteString(`<div style="display:flex;align-items:center;gap:12px;margin-bottom:16px;flex-wrap:wrap">`)
+	body.WriteString(`<a class="btn btn-secondary btn-sm" href="/tasks">Back to Tasks</a>`)
+	if task.Status == TaskRunning || task.Status == TaskPending {
+		body.WriteString(`<button class="btn btn-danger btn-sm" onclick="cancelTaskDetail('` + html.EscapeString(task.ID) + `')">Cancel</button>`)
+	}
+	body.WriteString(`<span style="font-size:12px;color:var(--muted)">Artifacts are saved in the task folder under <code>./tasks</code>.</span>`)
+	body.WriteString(`</div>`)
+
+	if report := loadTaskReportFragment(task); report != "" {
+		body.WriteString(report)
+	} else {
+		body.WriteString(`<div class="card"><div class="card-head">Task Summary</div><div class="card-body">`)
+		body.WriteString(`<div style="font-size:18px;font-weight:700">` + html.EscapeString(title) + `</div>`)
+		body.WriteString(`<div style="margin-top:8px">` + renderTaskStatusBadge(task.Status) + `</div>`)
+		if strings.TrimSpace(task.ErrMsg) != "" {
+			body.WriteString(`<div style="margin-top:8px;color:var(--crit-fg)">` + html.EscapeString(task.ErrMsg) + `</div>`)
+		}
+		body.WriteString(`</div></div>`)
+	}
+
+	if task.Status == TaskRunning {
+		body.WriteString(`<div class="card"><div class="card-head">Live Charts</div><div class="card-body">`)
+		body.WriteString(`<div id="task-live-charts" style="display:flex;flex-direction:column;gap:16px;color:var(--muted);font-size:13px">Loading charts...</div>`)
+		body.WriteString(`</div></div>`)
+	}
+
+	if task.Status == TaskRunning || task.Status == TaskPending {
+		body.WriteString(`<div class="card"><div class="card-head">Live Logs</div><div class="card-body">`)
+		body.WriteString(`<div id="task-live-log" class="terminal" style="max-height:none;white-space:pre-wrap">Connecting...</div>`)
+		body.WriteString(`</div></div>`)
+		body.WriteString(`<script>
+function cancelTaskDetail(id) {
+  fetch('/api/tasks/' + id + '/cancel', {method:'POST'}).then(function(){
+    var term = document.getElementById('task-live-log');
+    if (term) {
+      term.textContent += '\nCancel requested.\n';
+      term.scrollTop = term.scrollHeight;
+    }
+  });
+}
+function renderTaskLiveCharts(taskId, charts) {
+  const host = document.getElementById('task-live-charts');
+  if (!host) return;
+  if (!Array.isArray(charts) || charts.length === 0) {
+    host.innerHTML = 'Waiting for metric samples...';
+    return;
+  }
+  const seen = {};
+  charts.forEach(function(chart) {
+    seen[chart.file] = true;
+    let img = host.querySelector('img[data-chart-file="' + chart.file + '"]');
+    if (img) {
+      const card = img.closest('.card');
+      if (card) {
+        const title = card.querySelector('.card-head');
+        if (title) title.textContent = chart.title;
+      }
+      return;
+    }
+    const card = document.createElement('div');
+    card.className = 'card';
+    card.style.margin = '0';
+    card.innerHTML = '<div class="card-head"></div><div class="card-body" style="padding:12px"></div>';
+    card.querySelector('.card-head').textContent = chart.title;
+    const body = card.querySelector('.card-body');
+    img = document.createElement('img');
+    img.setAttribute('data-task-chart', '1');
+    img.setAttribute('data-chart-file', chart.file);
+    img.setAttribute('data-base-src', '/api/tasks/' + taskId + '/chart/' + chart.file);
+    img.src = '/api/tasks/' + taskId + '/chart/' + chart.file + '?t=' + Date.now();
+    img.style.width = '100%';
+    img.style.display = 'block';
+    img.style.borderRadius = '6px';
+    img.alt = chart.title;
+    body.appendChild(img);
+    host.appendChild(card);
+  });
+  Array.from(host.querySelectorAll('img[data-task-chart="1"]')).forEach(function(img) {
+    const file = img.getAttribute('data-chart-file') || '';
+    if (seen[file]) return;
+    const card = img.closest('.card');
+    if (card) card.remove();
+  });
+}
+function loadTaskLiveCharts(taskId) {
+  fetch('/api/tasks/' + taskId + '/charts').then(function(r){ return r.json(); }).then(function(charts){
+    renderTaskLiveCharts(taskId, charts);
+  }).catch(function(){
+    const host = document.getElementById('task-live-charts');
+    if (host) host.innerHTML = 'Task charts are unavailable.';
+  });
+}
+function refreshTaskLiveCharts() {
+  document.querySelectorAll('img[data-task-chart="1"]').forEach(function(img){
+    const base = img.dataset.baseSrc;
+    if (!base) return;
+    img.src = base + '?t=' + Date.now();
+  });
+}
+var _taskDetailES = new EventSource('/api/tasks/` + html.EscapeString(task.ID) + `/stream');
+var _taskDetailTerm = document.getElementById('task-live-log');
+var _taskChartTimer = null;
+var _taskChartsFrozen = false;
+_taskDetailES.onopen = function(){ _taskDetailTerm.textContent = ''; };
+_taskDetailES.onmessage = function(e){ _taskDetailTerm.textContent += e.data + "\n"; _taskDetailTerm.scrollTop = _taskDetailTerm.scrollHeight; };
+_taskDetailES.addEventListener('done', function(e){
+  if (_taskChartTimer) clearInterval(_taskChartTimer);
+  _taskDetailES.close();
+  _taskDetailES = null;
+  _taskChartsFrozen = true;
+  _taskDetailTerm.textContent += (e.data ? '\nTask finished with error.\n' : '\nTask finished.\n');
+  _taskDetailTerm.scrollTop = _taskDetailTerm.scrollHeight;
+  refreshTaskLiveCharts();
+});
+_taskDetailES.onerror = function(){
+  if (_taskChartTimer) clearInterval(_taskChartTimer);
+  if (_taskDetailES) {
+    _taskDetailES.close();
+    _taskDetailES = null;
+  }
+};
+loadTaskLiveCharts('` + html.EscapeString(task.ID) + `');
+_taskChartTimer = setInterval(function(){
+  if (_taskChartsFrozen) return;
+  loadTaskLiveCharts('` + html.EscapeString(task.ID) + `');
+  refreshTaskLiveCharts();
+}, 2000);
+</script>`)
+	}
+
+	return layoutHead(opts.Title+" — "+title) +
+		layoutNav("tasks", opts.BuildLabel) +
+		`<div class="main"><div class="topbar"><h1>` + html.EscapeString(title) + `</h1></div><div class="content">` +
+		body.String() +
+		`</div></div></body></html>`
+}
+
+func loadTaskReportFragment(task Task) string {
+	if strings.TrimSpace(task.ReportHTMLPath) == "" {
+		return ""
+	}
+	data, err := os.ReadFile(task.ReportHTMLPath)
+	if err != nil || len(data) == 0 {
+		return ""
+	}
+	return string(data)
+}
+
+func taskArtifactDownloadLink(task Task, absPath string) string {
+	if strings.TrimSpace(absPath) == "" {
+		return ""
+	}
+	return fmt.Sprintf(`/export/file?path=%s`, absPath)
+}
+
+func (h *handler) taskSamplesForRequest(r *http.Request) (Task, []platform.LiveMetricSample, time.Time, time.Time, bool) {
+	id := r.PathValue("id")
+	taskPtr, ok := globalQueue.findByID(id)
+	if !ok {
+		return Task{}, nil, time.Time{}, time.Time{}, false
+	}
+	task := *taskPtr
+	start, end := taskTimeWindow(&task)
+	samples, err := loadTaskMetricSamples(start, end)
+	if err != nil {
+		return task, nil, start, end, true
+	}
+	return task, samples, start, end, true
+}
+
+func taskTimelineForTask(task Task) []chartTimelineSegment {
+	start, end := taskTimeWindow(&task)
+	return []chartTimelineSegment{{Start: start, End: end, Active: true}}
+}
+
+func taskChartPathFromFile(file string) (string, bool) {
+	file = strings.TrimSpace(file)
+	for _, spec := range taskDashboardChartSpecs {
+		if spec.File == file {
+			return spec.Path, true
+		}
+	}
+	if strings.HasPrefix(file, "gpu-") && strings.HasSuffix(file, "-overview.svg") {
+		id := strings.TrimSuffix(strings.TrimPrefix(file, "gpu-"), "-overview.svg")
+		return "gpu/" + id + "-overview", true
+	}
+	return "", false
+}

audit/internal/webui/task_report.go

@@ -0,0 +1,343 @@
+package webui
+
+import (
+	"encoding/json"
+	"fmt"
+	"html"
+	"os"
+	"path/filepath"
+	"sort"
+	"strings"
+	"time"
+
+	"bee/audit/internal/platform"
+)
+
+var taskReportMetricsDBPath = metricsDBPath
+
+type taskReport struct {
+	ID          string            `json:"id"`
+	Name        string            `json:"name"`
+	Target      string            `json:"target"`
+	Status      string            `json:"status"`
+	CreatedAt   time.Time         `json:"created_at"`
+	StartedAt   *time.Time        `json:"started_at,omitempty"`
+	DoneAt      *time.Time        `json:"done_at,omitempty"`
+	DurationSec int               `json:"duration_sec,omitempty"`
+	Error       string            `json:"error,omitempty"`
+	LogFile     string            `json:"log_file,omitempty"`
+	Charts      []taskReportChart `json:"charts,omitempty"`
+	GeneratedAt time.Time         `json:"generated_at"`
+}
+
+type taskReportChart struct {
+	Title string `json:"title"`
+	File  string `json:"file"`
+}
+
+type taskChartSpec struct {
+	Path string
+	File string
+}
+
+var taskDashboardChartSpecs = []taskChartSpec{
+	{Path: "server-load", File: "server-load.svg"},
+	{Path: "server-temp-cpu", File: "server-temp-cpu.svg"},
+	{Path: "server-temp-ambient", File: "server-temp-ambient.svg"},
+	{Path: "server-power", File: "server-power.svg"},
+	{Path: "server-fans", File: "server-fans.svg"},
+	{Path: "gpu-all-load", File: "gpu-all-load.svg"},
+	{Path: "gpu-all-memload", File: "gpu-all-memload.svg"},
+	{Path: "gpu-all-clock", File: "gpu-all-clock.svg"},
+	{Path: "gpu-all-power", File: "gpu-all-power.svg"},
+	{Path: "gpu-all-temp", File: "gpu-all-temp.svg"},
+}
+
+func taskChartSpecsForSamples(samples []platform.LiveMetricSample) []taskChartSpec {
+	specs := make([]taskChartSpec, 0, len(taskDashboardChartSpecs)+len(taskGPUIndices(samples)))
+	specs = append(specs, taskDashboardChartSpecs...)
+	for _, idx := range taskGPUIndices(samples) {
+		specs = append(specs, taskChartSpec{
+			Path: fmt.Sprintf("gpu/%d-overview", idx),
+			File: fmt.Sprintf("gpu-%d-overview.svg", idx),
+		})
+	}
+	return specs
+}
+
+func writeTaskReportArtifacts(t *Task) error {
+	if t == nil {
+		return nil
+	}
+	ensureTaskReportPaths(t)
+	if strings.TrimSpace(t.ArtifactsDir) == "" {
+		return nil
+	}
+	if err := os.MkdirAll(t.ArtifactsDir, 0755); err != nil {
+		return err
+	}
+
+	start, end := taskTimeWindow(t)
+	samples, _ := loadTaskMetricSamples(start, end)
+	charts, inlineCharts := writeTaskCharts(t.ArtifactsDir, start, end, samples)
+
+	logText := ""
+	if data, err := os.ReadFile(t.LogPath); err == nil {
+		logText = string(data)
+	}
+
+	report := taskReport{
+		ID:          t.ID,
+		Name:        t.Name,
+		Target:      t.Target,
+		Status:      t.Status,
+		CreatedAt:   t.CreatedAt,
+		StartedAt:   t.StartedAt,
+		DoneAt:      t.DoneAt,
+		DurationSec: taskElapsedSec(t, reportDoneTime(t)),
+		Error:       t.ErrMsg,
+		LogFile:     filepath.Base(t.LogPath),
+		Charts:      charts,
+		GeneratedAt: time.Now().UTC(),
+	}
+	if err := writeJSONFile(t.ReportJSONPath, report); err != nil {
+		return err
+	}
+	return os.WriteFile(t.ReportHTMLPath, []byte(renderTaskReportFragment(report, inlineCharts, logText)), 0644)
+}
+
+func reportDoneTime(t *Task) time.Time {
+	if t != nil && t.DoneAt != nil && !t.DoneAt.IsZero() {
+		return *t.DoneAt
+	}
+	return time.Now()
+}
+
+func taskTimeWindow(t *Task) (time.Time, time.Time) {
+	if t == nil {
+		now := time.Now().UTC()
+		return now, now
+	}
+	start := t.CreatedAt.UTC()
+	if t.StartedAt != nil && !t.StartedAt.IsZero() {
+		start = t.StartedAt.UTC()
+	}
+	end := time.Now().UTC()
+	if t.DoneAt != nil && !t.DoneAt.IsZero() {
+		end = t.DoneAt.UTC()
+	}
+	if end.Before(start) {
+		end = start
+	}
+	return start, end
+}
+
+func loadTaskMetricSamples(start, end time.Time) ([]platform.LiveMetricSample, error) {
+	db, err := openMetricsDB(taskReportMetricsDBPath)
+	if err != nil {
+		return nil, err
+	}
+	defer db.Close()
+	return db.LoadBetween(start, end)
+}
+
+func writeTaskCharts(dir string, start, end time.Time, samples []platform.LiveMetricSample) ([]taskReportChart, map[string]string) {
+	if len(samples) == 0 {
+		return nil, nil
+	}
+	timeline := []chartTimelineSegment{{Start: start, End: end, Active: true}}
+	var charts []taskReportChart
+	inline := make(map[string]string)
+	for _, spec := range taskChartSpecsForSamples(samples) {
+		title, svg, ok := renderTaskChartSVG(spec.Path, samples, timeline)
+		if !ok || len(svg) == 0 {
+			continue
+		}
+		path := filepath.Join(dir, spec.File)
+		if err := os.WriteFile(path, svg, 0644); err != nil {
+			continue
+		}
+		charts = append(charts, taskReportChart{Title: title, File: spec.File})
+		inline[spec.File] = string(svg)
+	}
+	return charts, inline
+}
+
+func renderTaskChartSVG(path string, samples []platform.LiveMetricSample, timeline []chartTimelineSegment) (string, []byte, bool) {
+	if idx, sub, ok := parseGPUChartPath(path); ok && sub == "overview" {
+		buf, hasData, err := renderGPUOverviewChartSVG(idx, samples, timeline)
+		if err != nil || !hasData {
+			return "", nil, false
+		}
+		return gpuDisplayLabel(idx) + " Overview", buf, true
+	}
+	datasets, names, labels, title, yMin, yMax, ok := chartDataFromSamples(path, samples)
+	if !ok {
+		return "", nil, false
+	}
+	buf, err := renderMetricChartSVG(
+		title,
+		labels,
+		sampleTimes(samples),
+		datasets,
+		names,
+		yMin,
+		yMax,
+		chartCanvasHeightForPath(path, len(names)),
+		timeline,
+	)
+	if err != nil {
+		return "", nil, false
+	}
+	return title, buf, true
+}
+
+func taskGPUIndices(samples []platform.LiveMetricSample) []int {
+	seen := map[int]bool{}
+	var out []int
+	for _, s := range samples {
+		for _, g := range s.GPUs {
+			if seen[g.GPUIndex] {
+				continue
+			}
+			seen[g.GPUIndex] = true
+			out = append(out, g.GPUIndex)
+		}
+	}
+	sort.Ints(out)
+	return out
+}
+
+func writeJSONFile(path string, v any) error {
+	data, err := json.MarshalIndent(v, "", "  ")
+	if err != nil {
+		return err
+	}
+	return os.WriteFile(path, data, 0644)
+}
+
+func renderTaskReportFragment(report taskReport, charts map[string]string, logText string) string {
+	var b strings.Builder
+	b.WriteString(`<div class="card"><div class="card-head">Task Report</div><div class="card-body">`)
+	b.WriteString(`<div class="grid2">`)
+	b.WriteString(`<div><div style="font-size:12px;color:var(--muted);margin-bottom:6px">Task</div><div style="font-size:16px;font-weight:700">` + html.EscapeString(report.Name) + `</div>`)
+	b.WriteString(`<div style="font-size:13px;color:var(--muted)">` + html.EscapeString(report.Target) + `</div></div>`)
+	b.WriteString(`<div><div style="font-size:12px;color:var(--muted);margin-bottom:6px">Status</div><div>` + renderTaskStatusBadge(report.Status) + `</div>`)
+	if strings.TrimSpace(report.Error) != "" {
+		b.WriteString(`<div style="margin-top:8px;font-size:13px;color:var(--crit-fg)">` + html.EscapeString(report.Error) + `</div>`)
+	}
+	b.WriteString(`</div></div>`)
+	b.WriteString(`<div style="margin-top:14px;font-size:13px;color:var(--muted)">`)
+	b.WriteString(`Started: ` + formatTaskTime(report.StartedAt, report.CreatedAt) + ` | Finished: ` + formatTaskTime(report.DoneAt, time.Time{}) + ` | Duration: ` + formatTaskDuration(report.DurationSec))
+	b.WriteString(`</div></div></div>`)
+	if benchmarkCard := renderTaskBenchmarkResultsCard(report.Target, logText); benchmarkCard != "" {
+		b.WriteString(benchmarkCard)
+	}
+
+	if len(report.Charts) > 0 {
+		for _, chart := range report.Charts {
+			b.WriteString(`<div class="card"><div class="card-head">` + html.EscapeString(chart.Title) + `</div><div class="card-body" style="padding:12px">`)
+			b.WriteString(charts[chart.File])
+			b.WriteString(`</div></div>`)
+		}
+	} else {
+		b.WriteString(`<div class="alert alert-info">No metric samples were captured during this task window.</div>`)
+	}
+
+	b.WriteString(`<div class="card"><div class="card-head">Logs</div><div class="card-body">`)
+	b.WriteString(`<div class="terminal" style="max-height:none;white-space:pre-wrap">` + html.EscapeString(strings.TrimSpace(logText)) + `</div>`)
+	b.WriteString(`</div></div>`)
+	return b.String()
+}
+
+func renderTaskBenchmarkResultsCard(target, logText string) string {
+	if strings.TrimSpace(target) != "nvidia-benchmark" {
+		return ""
+	}
+	resultPath := taskBenchmarkResultPath(logText)
+	if strings.TrimSpace(resultPath) == "" {
+		return ""
+	}
+	columns, runs := loadBenchmarkHistoryFromPaths([]string{resultPath})
+	if len(runs) == 0 {
+		return ""
+	}
+	return renderBenchmarkResultsCardFromRuns(
+		"Benchmark Results",
+		"Composite score for this benchmark task.",
+		"No benchmark results were saved for this task.",
+		columns,
+		runs,
+	)
+}
+
+func taskBenchmarkResultPath(logText string) string {
+	archivePath := taskArchivePathFromLog(logText)
+	if archivePath == "" {
+		return ""
+	}
+	runDir := strings.TrimSuffix(archivePath, ".tar.gz")
+	if runDir == archivePath {
+		return ""
+	}
+	return filepath.Join(runDir, "result.json")
+}
+
+func taskArchivePathFromLog(logText string) string {
+	lines := strings.Split(logText, "\n")
+	for i := len(lines) - 1; i >= 0; i-- {
+		line := strings.TrimSpace(lines[i])
+		if line == "" || !strings.HasPrefix(line, "Archive:") {
+			continue
+		}
+		path := strings.TrimSpace(strings.TrimPrefix(line, "Archive:"))
+		if strings.HasPrefix(path, "Archive written to ") {
+			path = strings.TrimSpace(strings.TrimPrefix(path, "Archive written to "))
+		}
+		if strings.HasSuffix(path, ".tar.gz") {
+			return path
+		}
+	}
+	return ""
+}
+
+func renderTaskStatusBadge(status string) string {
+	className := map[string]string{
+		TaskRunning:   "badge-ok",
+		TaskPending:   "badge-unknown",
+		TaskDone:      "badge-ok",
+		TaskFailed:    "badge-err",
+		TaskCancelled: "badge-unknown",
+	}[status]
+	if className == "" {
+		className = "badge-unknown"
+	}
+	label := strings.TrimSpace(status)
+	if label == "" {
+		label = "unknown"
+	}
+	return `<span class="badge ` + className + `">` + html.EscapeString(label) + `</span>`
+}
+
+func formatTaskTime(ts *time.Time, fallback time.Time) string {
+	if ts != nil && !ts.IsZero() {
+		return ts.Local().Format("2006-01-02 15:04:05")
+	}
+	if !fallback.IsZero() {
+		return fallback.Local().Format("2006-01-02 15:04:05")
+	}
+	return "n/a"
+}
+
+func formatTaskDuration(sec int) string {
+	if sec <= 0 {
+		return "n/a"
+	}
+	if sec < 60 {
+		return fmt.Sprintf("%ds", sec)
+	}
+	if sec < 3600 {
+		return fmt.Sprintf("%dm %02ds", sec/60, sec%60)
+	}
+	return fmt.Sprintf("%dh %02dm %02ds", sec/3600, (sec%3600)/60, sec%60)
+}

audit/internal/webui/tasks.go

@@ -4,10 +4,12 @@ import (
 	"context"
 	"encoding/json"
 	"fmt"
+	"log/slog"
 	"net/http"
 	"os"
 	"os/exec"
 	"path/filepath"
+	"runtime/debug"
 	"sort"
 	"strings"
 	"sync"
@@ -28,22 +30,29 @@ const (
 
 // taskNames maps target → human-readable name for validate (SAT) runs.
 var taskNames = map[string]string{
-	"nvidia":          "NVIDIA SAT",
-	"nvidia-stress":   "NVIDIA GPU Stress",
-	"memory":          "Memory SAT",
-	"storage":         "Storage SAT",
-	"cpu":             "CPU SAT",
-	"amd":             "AMD GPU SAT",
-	"amd-mem":         "AMD GPU MEM Integrity",
-	"amd-bandwidth":   "AMD GPU MEM Bandwidth",
-	"amd-stress":      "AMD GPU Burn-in",
-	"memory-stress":   "Memory Burn-in",
-	"sat-stress":      "SAT Stress (stressapptest)",
-	"platform-stress": "Platform Thermal Cycling",
-	"audit":           "Audit",
-	"support-bundle":  "Support Bundle",
-	"install":         "Install to Disk",
-	"install-to-ram":  "Install to RAM",
+	"nvidia":                 "NVIDIA SAT",
+	"nvidia-targeted-stress": "NVIDIA Targeted Stress Validate (dcgmi diag targeted_stress)",
+	"nvidia-benchmark":       "NVIDIA Benchmark",
+	"nvidia-compute":         "NVIDIA Max Compute Load (dcgmproftester)",
+	"nvidia-targeted-power":  "NVIDIA Targeted Power (dcgmi diag targeted_power)",
+	"nvidia-pulse":           "NVIDIA Pulse Test (dcgmi diag pulse_test)",
+	"nvidia-interconnect":    "NVIDIA Interconnect Test (NCCL all_reduce_perf)",
+	"nvidia-bandwidth":       "NVIDIA Bandwidth Test (NVBandwidth)",
+	"nvidia-stress":          "NVIDIA GPU Stress",
+	"memory":                 "Memory SAT",
+	"storage":                "Storage SAT",
+	"cpu":                    "CPU SAT",
+	"amd":                    "AMD GPU SAT",
+	"amd-mem":                "AMD GPU MEM Integrity",
+	"amd-bandwidth":          "AMD GPU MEM Bandwidth",
+	"amd-stress":             "AMD GPU Burn-in",
+	"memory-stress":          "Memory Burn-in",
+	"sat-stress":             "SAT Stress (stressapptest)",
+	"platform-stress":        "Platform Thermal Cycling",
+	"audit":                  "Audit",
+	"support-bundle":         "Support Bundle",
+	"install":                "Install to Disk",
+	"install-to-ram":         "Install to RAM",
 }
 
 // burnNames maps target → human-readable name when a burn profile is set.
@@ -83,17 +92,20 @@ func taskDisplayName(target, profile, loader string) string {
 
 // Task represents one unit of work in the queue.
 type Task struct {
-	ID         string     `json:"id"`
-	Name       string     `json:"name"`
-	Target     string     `json:"target"`
-	Priority   int        `json:"priority"`
-	Status     string     `json:"status"`
-	CreatedAt  time.Time  `json:"created_at"`
-	StartedAt  *time.Time `json:"started_at,omitempty"`
-	DoneAt     *time.Time `json:"done_at,omitempty"`
-	ElapsedSec int        `json:"elapsed_sec,omitempty"`
-	ErrMsg     string     `json:"error,omitempty"`
-	LogPath    string     `json:"log_path,omitempty"`
+	ID             string     `json:"id"`
+	Name           string     `json:"name"`
+	Target         string     `json:"target"`
+	Priority       int        `json:"priority"`
+	Status         string     `json:"status"`
+	CreatedAt      time.Time  `json:"created_at"`
+	StartedAt      *time.Time `json:"started_at,omitempty"`
+	DoneAt         *time.Time `json:"done_at,omitempty"`
+	ElapsedSec     int        `json:"elapsed_sec,omitempty"`
+	ErrMsg         string     `json:"error,omitempty"`
+	LogPath        string     `json:"log_path,omitempty"`
+	ArtifactsDir   string     `json:"artifacts_dir,omitempty"`
+	ReportJSONPath string     `json:"report_json_path,omitempty"`
+	ReportHTMLPath string     `json:"report_html_path,omitempty"`
 
 	// runtime fields (not serialised)
 	job    *jobState
@@ -103,83 +115,108 @@ type Task struct {
 // taskParams holds optional parameters parsed from the run request.
 type taskParams struct {
 	Duration           int      `json:"duration,omitempty"`
-	DiagLevel          int      `json:"diag_level,omitempty"`
+	StressMode         bool     `json:"stress_mode,omitempty"`
 	GPUIndices         []int    `json:"gpu_indices,omitempty"`
 	ExcludeGPUIndices  []int    `json:"exclude_gpu_indices,omitempty"`
+	StaggerGPUStart    bool     `json:"stagger_gpu_start,omitempty"`
+	SizeMB             int      `json:"size_mb,omitempty"`
+	Passes             int      `json:"passes,omitempty"`
 	Loader             string   `json:"loader,omitempty"`
 	BurnProfile        string   `json:"burn_profile,omitempty"`
+	BenchmarkProfile   string   `json:"benchmark_profile,omitempty"`
+	RunNCCL            bool     `json:"run_nccl,omitempty"`
+	ParallelGPUs       bool     `json:"parallel_gpus,omitempty"`
 	DisplayName        string   `json:"display_name,omitempty"`
 	Device             string   `json:"device,omitempty"` // for install
 	PlatformComponents []string `json:"platform_components,omitempty"`
 }
 
 type persistedTask struct {
-	ID        string     `json:"id"`
-	Name      string     `json:"name"`
-	Target    string     `json:"target"`
-	Priority  int        `json:"priority"`
-	Status    string     `json:"status"`
-	CreatedAt time.Time  `json:"created_at"`
-	StartedAt *time.Time `json:"started_at,omitempty"`
-	DoneAt    *time.Time `json:"done_at,omitempty"`
-	ErrMsg    string     `json:"error,omitempty"`
-	LogPath   string     `json:"log_path,omitempty"`
-	Params    taskParams `json:"params,omitempty"`
+	ID             string     `json:"id"`
+	Name           string     `json:"name"`
+	Target         string     `json:"target"`
+	Priority       int        `json:"priority"`
+	Status         string     `json:"status"`
+	CreatedAt      time.Time  `json:"created_at"`
+	StartedAt      *time.Time `json:"started_at,omitempty"`
+	DoneAt         *time.Time `json:"done_at,omitempty"`
+	ErrMsg         string     `json:"error,omitempty"`
+	LogPath        string     `json:"log_path,omitempty"`
+	ArtifactsDir   string     `json:"artifacts_dir,omitempty"`
+	ReportJSONPath string     `json:"report_json_path,omitempty"`
+	ReportHTMLPath string     `json:"report_html_path,omitempty"`
+	Params         taskParams `json:"params,omitempty"`
 }
 
 type burnPreset struct {
-	NvidiaDiag  int
 	DurationSec int
 }
 
 func resolveBurnPreset(profile string) burnPreset {
 	switch profile {
 	case "overnight":
-		return burnPreset{NvidiaDiag: 4, DurationSec: 8 * 60 * 60}
+		return burnPreset{DurationSec: 8 * 60 * 60}
 	case "acceptance":
-		return burnPreset{NvidiaDiag: 3, DurationSec: 60 * 60}
+		return burnPreset{DurationSec: 60 * 60}
 	default:
-		return burnPreset{NvidiaDiag: 1, DurationSec: 5 * 60}
+		return burnPreset{DurationSec: 5 * 60}
 	}
 }
 
+func boolToNvidiaStaggerSeconds(enabled bool, selected []int) int {
+	if enabled && len(selected) > 1 {
+		return 180
+	}
+	return 0
+}
+
 func resolvePlatformStressPreset(profile string) platform.PlatformStressOptions {
+	acceptanceCycles := []platform.PlatformStressCycle{
+		{LoadSec: 85, IdleSec: 5},
+		{LoadSec: 80, IdleSec: 10},
+		{LoadSec: 55, IdleSec: 5},
+		{LoadSec: 60, IdleSec: 0},
+		{LoadSec: 100, IdleSec: 10},
+		{LoadSec: 145, IdleSec: 15},
+		{LoadSec: 190, IdleSec: 20},
+		{LoadSec: 235, IdleSec: 25},
+		{LoadSec: 280, IdleSec: 30},
+		{LoadSec: 325, IdleSec: 35},
+		{LoadSec: 370, IdleSec: 40},
+		{LoadSec: 415, IdleSec: 45},
+		{LoadSec: 460, IdleSec: 50},
+		{LoadSec: 510, IdleSec: 0},
+	}
+
 	switch profile {
 	case "overnight":
-		return platform.PlatformStressOptions{Cycles: []platform.PlatformStressCycle{
-			{LoadSec: 600, IdleSec: 120},
-			{LoadSec: 600, IdleSec: 60},
-			{LoadSec: 600, IdleSec: 30},
-			{LoadSec: 600, IdleSec: 120},
-			{LoadSec: 600, IdleSec: 60},
-			{LoadSec: 600, IdleSec: 30},
-			{LoadSec: 600, IdleSec: 120},
-			{LoadSec: 600, IdleSec: 60},
-		}}
+		cycles := make([]platform.PlatformStressCycle, 0, len(acceptanceCycles)*8)
+		for range 8 {
+			cycles = append(cycles, acceptanceCycles...)
+		}
+		return platform.PlatformStressOptions{Cycles: cycles}
 	case "acceptance":
-		return platform.PlatformStressOptions{Cycles: []platform.PlatformStressCycle{
-			{LoadSec: 300, IdleSec: 60},
-			{LoadSec: 300, IdleSec: 30},
-			{LoadSec: 300, IdleSec: 60},
-			{LoadSec: 300, IdleSec: 30},
-		}}
+		return platform.PlatformStressOptions{Cycles: acceptanceCycles}
 	default: // smoke
 		return platform.PlatformStressOptions{Cycles: []platform.PlatformStressCycle{
-			{LoadSec: 90, IdleSec: 60},
-			{LoadSec: 90, IdleSec: 30},
+			{LoadSec: 85, IdleSec: 5},
+			{LoadSec: 80, IdleSec: 10},
+			{LoadSec: 55, IdleSec: 5},
+			{LoadSec: 60, IdleSec: 0},
 		}}
 	}
 }
 
 // taskQueue manages a priority-ordered list of tasks and runs them one at a time.
 type taskQueue struct {
-	mu        sync.Mutex
-	tasks     []*Task
-	trigger   chan struct{}
-	opts      *HandlerOptions // set by startWorker
-	statePath string
-	logsDir   string
-	started   bool
+	mu          sync.Mutex
+	tasks       []*Task
+	trigger     chan struct{}
+	opts        *HandlerOptions // set by startWorker
+	statePath   string
+	logsDir     string
+	started     bool
+	kmsgWatcher *kmsgWatcher
 }
 
 var globalQueue = &taskQueue{trigger: make(chan struct{}, 1)}
@@ -187,11 +224,11 @@ var globalQueue = &taskQueue{trigger: make(chan struct{}, 1)}
 const maxTaskHistory = 50
 
 var (
-	runMemoryAcceptancePackCtx = func(a *app.App, ctx context.Context, baseDir string, logFunc func(string)) (string, error) {
-		return a.RunMemoryAcceptancePackCtx(ctx, baseDir, logFunc)
+	runMemoryAcceptancePackCtx = func(a *app.App, ctx context.Context, baseDir string, sizeMB, passes int, logFunc func(string)) (string, error) {
+		return a.RunMemoryAcceptancePackCtx(ctx, baseDir, sizeMB, passes, logFunc)
 	}
-	runStorageAcceptancePackCtx = func(a *app.App, ctx context.Context, baseDir string, logFunc func(string)) (string, error) {
-		return a.RunStorageAcceptancePackCtx(ctx, baseDir, logFunc)
+	runStorageAcceptancePackCtx = func(a *app.App, ctx context.Context, baseDir string, extended bool, logFunc func(string)) (string, error) {
+		return a.RunStorageAcceptancePackCtx(ctx, baseDir, extended, logFunc)
 	}
 	runCPUAcceptancePackCtx = func(a *app.App, ctx context.Context, baseDir string, durationSec int, logFunc func(string)) (string, error) {
 		return a.RunCPUAcceptancePackCtx(ctx, baseDir, durationSec, logFunc)
@@ -231,6 +268,7 @@ func (q *taskQueue) enqueue(t *Task) {
 	q.prune()
 	q.persistLocked()
 	q.mu.Unlock()
+	taskSerialEvent(t, "queued")
 	select {
 	case q.trigger <- struct{}{}:
 	default:
@@ -376,7 +414,7 @@ func (q *taskQueue) startWorker(opts *HandlerOptions) {
 	if !q.started {
 		q.loadLocked()
 		q.started = true
-		go q.worker()
+		goRecoverLoop("task worker", 2*time.Second, q.worker)
 	}
 	hasPending := q.nextPending() != nil
 	q.mu.Unlock()
@@ -391,47 +429,101 @@ func (q *taskQueue) startWorker(opts *HandlerOptions) {
 func (q *taskQueue) worker() {
 	for {
 		<-q.trigger
-		setCPUGovernor("performance")
-		for {
-			q.mu.Lock()
-			t := q.nextPending()
-			if t == nil {
-				q.mu.Unlock()
-				break
-			}
-			now := time.Now()
-			t.Status = TaskRunning
-			t.StartedAt = &now
-			t.DoneAt = nil
-			t.ErrMsg = ""
-			j := newTaskJobState(t.LogPath)
-			ctx, cancel := context.WithCancel(context.Background())
-			j.cancel = cancel
-			t.job = j
-			q.persistLocked()
-			q.mu.Unlock()
+		func() {
+			setCPUGovernor("performance")
+			defer setCPUGovernor("powersave")
 
-			q.runTask(t, j, ctx)
-
-			q.mu.Lock()
-			now2 := time.Now()
-			t.DoneAt = &now2
-			if t.Status == TaskRunning { // not cancelled externally
-				if j.err != "" {
-					t.Status = TaskFailed
-					t.ErrMsg = j.err
-				} else {
-					t.Status = TaskDone
+			for {
+				q.mu.Lock()
+				t := q.nextPending()
+				if t == nil {
+					q.prune()
+					q.persistLocked()
+					q.mu.Unlock()
+					return
 				}
+				now := time.Now()
+				t.Status = TaskRunning
+				t.StartedAt = &now
+				t.DoneAt = nil
+				t.ErrMsg = ""
+				j := newTaskJobState(t.LogPath, taskSerialPrefix(t))
+				t.job = j
+				q.persistLocked()
+				q.mu.Unlock()
+
+				taskCtx, taskCancel := context.WithCancel(context.Background())
+				j.cancel = taskCancel
+				q.executeTask(t, j, taskCtx)
+				taskCancel()
+
+				q.mu.Lock()
+				q.prune()
+				q.persistLocked()
+				q.mu.Unlock()
 			}
-			q.prune()
-			q.persistLocked()
-			q.mu.Unlock()
-		}
-		setCPUGovernor("powersave")
+		}()
+
 	}
 }
 
+func (q *taskQueue) executeTask(t *Task, j *jobState, ctx context.Context) {
+	startedKmsgWatch := false
+	defer q.finalizeTaskRun(t, j)
+	defer func() {
+		if startedKmsgWatch && q.kmsgWatcher != nil {
+			q.kmsgWatcher.NotifyTaskFinished(t.ID)
+		}
+	}()
+	defer func() {
+		if rec := recover(); rec != nil {
+			msg := fmt.Sprintf("task panic: %v", rec)
+			slog.Error("task panic",
+				"task_id", t.ID,
+				"target", t.Target,
+				"panic", fmt.Sprint(rec),
+				"stack", string(debug.Stack()),
+			)
+			j.append("ERROR: " + msg)
+			j.finish(msg)
+		}
+	}()
+
+	if q.kmsgWatcher != nil && isSATTarget(t.Target) {
+		q.kmsgWatcher.NotifyTaskStarted(t.ID, t.Target)
+		startedKmsgWatch = true
+	}
+
+	q.runTask(t, j, ctx)
+}
+
+func (q *taskQueue) finalizeTaskRun(t *Task, j *jobState) {
+	q.mu.Lock()
+	now := time.Now()
+	t.DoneAt = &now
+	if t.Status == TaskRunning {
+		if j.err != "" {
+			t.Status = TaskFailed
+			t.ErrMsg = j.err
+		} else {
+			t.Status = TaskDone
+			t.ErrMsg = ""
+		}
+	}
+	q.finalizeTaskArtifactPathsLocked(t)
+	q.persistLocked()
+	q.mu.Unlock()
+
+	if err := writeTaskReportArtifacts(t); err != nil {
+		appendJobLog(t.LogPath, "WARN: task report generation failed: "+err.Error())
+	}
+	if t.ErrMsg != "" {
+		taskSerialEvent(t, "finished with status="+t.Status+" error="+t.ErrMsg)
+		return
+	}
+	taskSerialEvent(t, "finished with status="+t.Status)
+}
+
 // setCPUGovernor writes the given governor to all CPU scaling_governor sysfs files.
 // Silently ignores errors (e.g. when cpufreq is not available).
 func setCPUGovernor(governor string) {
@@ -469,9 +561,9 @@ func (q *taskQueue) runTask(t *Task, j *jobState, ctx context.Context) {
 			err = fmt.Errorf("app not configured")
 			break
 		}
-		diagLevel := t.params.DiagLevel
-		if t.params.BurnProfile != "" && diagLevel <= 0 {
-			diagLevel = resolveBurnPreset(t.params.BurnProfile).NvidiaDiag
+		diagLevel := 2
+		if t.params.StressMode {
+			diagLevel = 3
 		}
 		if len(t.params.GPUIndices) > 0 || diagLevel > 0 {
 			result, e := a.RunNvidiaAcceptancePackWithOptions(
@@ -485,7 +577,70 @@ func (q *taskQueue) runTask(t *Task, j *jobState, ctx context.Context) {
 		} else {
 			archive, err = a.RunNvidiaAcceptancePack("", j.append)
 		}
-	case "nvidia-stress":
+	case "nvidia-targeted-stress":
+		if a == nil {
+			err = fmt.Errorf("app not configured")
+			break
+		}
+		dur := t.params.Duration
+		if dur <= 0 {
+			dur = 300
+		}
+		archive, err = a.RunNvidiaTargetedStressValidatePack(ctx, "", dur, t.params.GPUIndices, j.append)
+	case "nvidia-benchmark":
+		if a == nil {
+			err = fmt.Errorf("app not configured")
+			break
+		}
+		archive, err = a.RunNvidiaBenchmarkCtx(ctx, "", platform.NvidiaBenchmarkOptions{
+			Profile:           t.params.BenchmarkProfile,
+			SizeMB:            t.params.SizeMB,
+			GPUIndices:        t.params.GPUIndices,
+			ExcludeGPUIndices: t.params.ExcludeGPUIndices,
+			RunNCCL:           t.params.RunNCCL,
+			ParallelGPUs:      t.params.ParallelGPUs,
+		}, j.append)
+		case "nvidia-compute":
+		if a == nil {
+			err = fmt.Errorf("app not configured")
+			break
+		}
+		dur := t.params.Duration
+		if t.params.BurnProfile != "" && dur <= 0 {
+			dur = resolveBurnPreset(t.params.BurnProfile).DurationSec
+		}
+			staggerSec := boolToNvidiaStaggerSeconds(t.params.StaggerGPUStart, t.params.GPUIndices)
+			if staggerSec > 0 {
+				j.append(fmt.Sprintf("NVIDIA staggered ramp-up enabled: %ds per GPU", staggerSec))
+			}
+			archive, err = a.RunNvidiaOfficialComputePack(ctx, "", dur, t.params.GPUIndices, staggerSec, j.append)
+	case "nvidia-targeted-power":
+		if a == nil {
+			err = fmt.Errorf("app not configured")
+			break
+		}
+		dur := t.params.Duration
+		if t.params.BurnProfile != "" && dur <= 0 {
+			dur = resolveBurnPreset(t.params.BurnProfile).DurationSec
+		}
+		archive, err = a.RunNvidiaTargetedPowerPack(ctx, "", dur, t.params.GPUIndices, j.append)
+	case "nvidia-pulse":
+		if a == nil {
+			err = fmt.Errorf("app not configured")
+			break
+		}
+		dur := t.params.Duration
+		if t.params.BurnProfile != "" && dur <= 0 {
+			dur = resolveBurnPreset(t.params.BurnProfile).DurationSec
+		}
+		archive, err = a.RunNvidiaPulseTestPack(ctx, "", dur, t.params.GPUIndices, j.append)
+	case "nvidia-bandwidth":
+		if a == nil {
+			err = fmt.Errorf("app not configured")
+			break
+		}
+		archive, err = a.RunNvidiaBandwidthPack(ctx, "", t.params.GPUIndices, j.append)
+	case "nvidia-interconnect":
 		if a == nil {
 			err = fmt.Errorf("app not configured")
 			break
@@ -495,23 +650,42 @@ func (q *taskQueue) runTask(t *Task, j *jobState, ctx context.Context) {
 			dur = resolveBurnPreset(t.params.BurnProfile).DurationSec
 		}
 		archive, err = runNvidiaStressPackCtx(a, ctx, "", platform.NvidiaStressOptions{
-			DurationSec:       dur,
-			Loader:            t.params.Loader,
-			GPUIndices:        t.params.GPUIndices,
-			ExcludeGPUIndices: t.params.ExcludeGPUIndices,
+			DurationSec: dur,
+			Loader:      platform.NvidiaStressLoaderNCCL,
+			GPUIndices:  t.params.GPUIndices,
 		}, j.append)
+	case "nvidia-stress":
+		if a == nil {
+			err = fmt.Errorf("app not configured")
+			break
+		}
+		dur := t.params.Duration
+		if t.params.BurnProfile != "" && dur <= 0 {
+			dur = resolveBurnPreset(t.params.BurnProfile).DurationSec
+		}
+			archive, err = runNvidiaStressPackCtx(a, ctx, "", platform.NvidiaStressOptions{
+				DurationSec:       dur,
+				Loader:            t.params.Loader,
+				GPUIndices:        t.params.GPUIndices,
+				ExcludeGPUIndices: t.params.ExcludeGPUIndices,
+				StaggerSeconds:    boolToNvidiaStaggerSeconds(t.params.StaggerGPUStart, t.params.GPUIndices),
+			}, j.append)
 	case "memory":
 		if a == nil {
 			err = fmt.Errorf("app not configured")
 			break
 		}
-		archive, err = runMemoryAcceptancePackCtx(a, ctx, "", j.append)
+		sizeMB, passes := 256, 1
+		if t.params.StressMode {
+			sizeMB, passes = 1024, 3
+		}
+		archive, err = runMemoryAcceptancePackCtx(a, ctx, "", sizeMB, passes, j.append)
 	case "storage":
 		if a == nil {
 			err = fmt.Errorf("app not configured")
 			break
 		}
-		archive, err = runStorageAcceptancePackCtx(a, ctx, "", j.append)
+		archive, err = runStorageAcceptancePackCtx(a, ctx, "", t.params.StressMode, j.append)
 	case "cpu":
 		if a == nil {
 			err = fmt.Errorf("app not configured")
@@ -522,7 +696,11 @@ func (q *taskQueue) runTask(t *Task, j *jobState, ctx context.Context) {
 			dur = resolveBurnPreset(t.params.BurnProfile).DurationSec
 		}
 		if dur <= 0 {
-			dur = 60
+			if t.params.StressMode {
+				dur = 1800
+			} else {
+				dur = 60
+			}
 		}
 		j.append(fmt.Sprintf("CPU stress duration: %ds", dur))
 		archive, err = runCPUAcceptancePackCtx(a, ctx, "", dur, j.append)
@@ -618,6 +796,19 @@ func (q *taskQueue) runTask(t *Task, j *jobState, ctx context.Context) {
 		return
 	}
 
+	// If the SAT archive was produced, check overall_status and write to component DB.
+	if archive != "" {
+		archivePath := app.ExtractArchivePath(archive)
+		if err == nil {
+			if app.ReadSATOverallStatus(archivePath) == "FAILED" {
+				err = fmt.Errorf("SAT overall_status=FAILED (see summary.txt)")
+			}
+		}
+		if db := q.statusDB(); db != nil {
+			app.ApplySATResultToDB(db, t.Target, archivePath)
+		}
+	}
+
 	if err != nil {
 		if ctx.Err() != nil {
 			j.append("Aborted.")
@@ -634,6 +825,13 @@ func (q *taskQueue) runTask(t *Task, j *jobState, ctx context.Context) {
 	j.finish("")
 }
 
+func (q *taskQueue) statusDB() *app.ComponentStatusDB {
+	if q.opts == nil || q.opts.App == nil {
+		return nil
+	}
+	return q.opts.App.StatusDB
+}
+
 func splitLines(s string) []string {
 	var out []string
 	for _, l := range splitNL(s) {
@@ -679,6 +877,7 @@ func (h *handler) handleAPITasksCancel(w http.ResponseWriter, r *http.Request) {
 		now := time.Now()
 		t.DoneAt = &now
 		globalQueue.persistLocked()
+		taskSerialEvent(t, "finished with status="+t.Status)
 		writeJSON(w, map[string]string{"status": "cancelled"})
 	case TaskRunning:
 		if t.job != nil {
@@ -688,6 +887,7 @@ func (h *handler) handleAPITasksCancel(w http.ResponseWriter, r *http.Request) {
 		now := time.Now()
 		t.DoneAt = &now
 		globalQueue.persistLocked()
+		taskSerialEvent(t, "finished with status="+t.Status)
 		writeJSON(w, map[string]string{"status": "cancelled"})
 	default:
 		writeError(w, http.StatusConflict, "task is not running or pending")
@@ -728,6 +928,7 @@ func (h *handler) handleAPITasksCancelAll(w http.ResponseWriter, _ *http.Request
 		case TaskPending:
 			t.Status = TaskCancelled
 			t.DoneAt = &now
+			taskSerialEvent(t, "finished with status="+t.Status)
 			n++
 		case TaskRunning:
 			if t.job != nil {
@@ -735,6 +936,7 @@ func (h *handler) handleAPITasksCancelAll(w http.ResponseWriter, _ *http.Request
 			}
 			t.Status = TaskCancelled
 			t.DoneAt = &now
+			taskSerialEvent(t, "finished with status="+t.Status)
 			n++
 		}
 	}
@@ -753,6 +955,7 @@ func (h *handler) handleAPITasksKillWorkers(w http.ResponseWriter, _ *http.Reque
 		case TaskPending:
 			t.Status = TaskCancelled
 			t.DoneAt = &now
+			taskSerialEvent(t, "finished with status="+t.Status)
 			cancelled++
 		case TaskRunning:
 			if t.job != nil {
@@ -760,6 +963,7 @@ func (h *handler) handleAPITasksKillWorkers(w http.ResponseWriter, _ *http.Reque
 			}
 			t.Status = TaskCancelled
 			t.DoneAt = &now
+			taskSerialEvent(t, "finished with status="+t.Status)
 			cancelled++
 		}
 	}
@@ -823,10 +1027,10 @@ func (h *handler) handleAPITasksStream(w http.ResponseWriter, r *http.Request) {
 }
 
 func (q *taskQueue) assignTaskLogPathLocked(t *Task) {
-	if t.LogPath != "" || q.logsDir == "" || t.ID == "" {
+	if q.logsDir == "" || t.ID == "" {
 		return
 	}
-	t.LogPath = filepath.Join(q.logsDir, t.ID+".log")
+	q.ensureTaskArtifactPathsLocked(t)
 }
 
 func (q *taskQueue) loadLocked() {
@@ -843,17 +1047,20 @@ func (q *taskQueue) loadLocked() {
 	}
 	for _, pt := range persisted {
 		t := &Task{
-			ID:        pt.ID,
-			Name:      pt.Name,
-			Target:    pt.Target,
-			Priority:  pt.Priority,
-			Status:    pt.Status,
-			CreatedAt: pt.CreatedAt,
-			StartedAt: pt.StartedAt,
-			DoneAt:    pt.DoneAt,
-			ErrMsg:    pt.ErrMsg,
-			LogPath:   pt.LogPath,
-			params:    pt.Params,
+			ID:             pt.ID,
+			Name:           pt.Name,
+			Target:         pt.Target,
+			Priority:       pt.Priority,
+			Status:         pt.Status,
+			CreatedAt:      pt.CreatedAt,
+			StartedAt:      pt.StartedAt,
+			DoneAt:         pt.DoneAt,
+			ErrMsg:         pt.ErrMsg,
+			LogPath:        pt.LogPath,
+			ArtifactsDir:   pt.ArtifactsDir,
+			ReportJSONPath: pt.ReportJSONPath,
+			ReportHTMLPath: pt.ReportHTMLPath,
+			params:         pt.Params,
 		}
 		q.assignTaskLogPathLocked(t)
 		if t.Status == TaskRunning {
@@ -884,17 +1091,20 @@ func (q *taskQueue) persistLocked() {
 	state := make([]persistedTask, 0, len(q.tasks))
 	for _, t := range q.tasks {
 		state = append(state, persistedTask{
-			ID:        t.ID,
-			Name:      t.Name,
-			Target:    t.Target,
-			Priority:  t.Priority,
-			Status:    t.Status,
-			CreatedAt: t.CreatedAt,
-			StartedAt: t.StartedAt,
-			DoneAt:    t.DoneAt,
-			ErrMsg:    t.ErrMsg,
-			LogPath:   t.LogPath,
-			Params:    t.params,
+			ID:             t.ID,
+			Name:           t.Name,
+			Target:         t.Target,
+			Priority:       t.Priority,
+			Status:         t.Status,
+			CreatedAt:      t.CreatedAt,
+			StartedAt:      t.StartedAt,
+			DoneAt:         t.DoneAt,
+			ErrMsg:         t.ErrMsg,
+			LogPath:        t.LogPath,
+			ArtifactsDir:   t.ArtifactsDir,
+			ReportJSONPath: t.ReportJSONPath,
+			ReportHTMLPath: t.ReportHTMLPath,
+			Params:         t.params,
 		})
 	}
 	data, err := json.MarshalIndent(state, "", "  ")
@@ -925,3 +1135,113 @@ func taskElapsedSec(t *Task, now time.Time) int {
 	}
 	return int(end.Sub(start).Round(time.Second) / time.Second)
 }
+
+func taskFolderStatus(status string) string {
+	status = strings.TrimSpace(strings.ToLower(status))
+	switch status {
+	case TaskRunning, TaskDone, TaskFailed, TaskCancelled:
+		return status
+	default:
+		return TaskPending
+	}
+}
+
+func sanitizeTaskFolderPart(s string) string {
+	s = strings.TrimSpace(strings.ToLower(s))
+	if s == "" {
+		return "task"
+	}
+	var b strings.Builder
+	lastDash := false
+	for _, r := range s {
+		isAlnum := (r >= 'a' && r <= 'z') || (r >= '0' && r <= '9')
+		if isAlnum {
+			b.WriteRune(r)
+			lastDash = false
+			continue
+		}
+		if !lastDash {
+			b.WriteByte('-')
+			lastDash = true
+		}
+	}
+	out := strings.Trim(b.String(), "-")
+	if out == "" {
+		return "task"
+	}
+	return out
+}
+
+func taskArtifactsDir(root string, t *Task, status string) string {
+	if strings.TrimSpace(root) == "" || t == nil {
+		return ""
+	}
+	prefix := taskFolderNumberPrefix(t.ID)
+	return filepath.Join(root, fmt.Sprintf("%s_%s_%s", prefix, sanitizeTaskFolderPart(t.Name), taskFolderStatus(status)))
+}
+
+func taskFolderNumberPrefix(taskID string) string {
+	taskID = strings.TrimSpace(taskID)
+	if strings.HasPrefix(taskID, "TASK-") && len(taskID) >= len("TASK-000") {
+		num := strings.TrimSpace(strings.TrimPrefix(taskID, "TASK-"))
+		if len(num) == 3 {
+			allDigits := true
+			for _, r := range num {
+				if r < '0' || r > '9' {
+					allDigits = false
+					break
+				}
+			}
+			if allDigits {
+				return num
+			}
+		}
+	}
+	fallback := sanitizeTaskFolderPart(taskID)
+	if fallback == "" {
+		return "000"
+	}
+	return fallback
+}
+
+func ensureTaskReportPaths(t *Task) {
+	if t == nil || strings.TrimSpace(t.ArtifactsDir) == "" {
+		return
+	}
+	if t.LogPath == "" || filepath.Base(t.LogPath) == "task.log" {
+		t.LogPath = filepath.Join(t.ArtifactsDir, "task.log")
+	}
+	t.ReportJSONPath = filepath.Join(t.ArtifactsDir, "report.json")
+	t.ReportHTMLPath = filepath.Join(t.ArtifactsDir, "report.html")
+}
+
+func (q *taskQueue) ensureTaskArtifactPathsLocked(t *Task) {
+	if t == nil || strings.TrimSpace(q.logsDir) == "" || strings.TrimSpace(t.ID) == "" {
+		return
+	}
+	if strings.TrimSpace(t.ArtifactsDir) == "" {
+		t.ArtifactsDir = taskArtifactsDir(q.logsDir, t, t.Status)
+	}
+	if t.ArtifactsDir != "" {
+		_ = os.MkdirAll(t.ArtifactsDir, 0755)
+	}
+	ensureTaskReportPaths(t)
+}
+
+func (q *taskQueue) finalizeTaskArtifactPathsLocked(t *Task) {
+	if t == nil || strings.TrimSpace(q.logsDir) == "" || strings.TrimSpace(t.ID) == "" {
+		return
+	}
+	q.ensureTaskArtifactPathsLocked(t)
+	dstDir := taskArtifactsDir(q.logsDir, t, t.Status)
+	if dstDir == "" {
+		return
+	}
+	if t.ArtifactsDir != "" && t.ArtifactsDir != dstDir {
+		if _, err := os.Stat(dstDir); err != nil {
+			_ = os.Rename(t.ArtifactsDir, dstDir)
+		}
+		t.ArtifactsDir = dstDir
+	}
+	ensureTaskReportPaths(t)
+}

audit/internal/webui/tasks_test.go

@@ -2,6 +2,7 @@ package webui
 
 import (
 	"context"
+	"encoding/json"
 	"net/http"
 	"net/http/httptest"
 	"os"
@@ -12,6 +13,7 @@ import (
 	"time"
 
 	"bee/audit/internal/app"
+	"bee/audit/internal/platform"
 )
 
 func TestTaskQueuePersistsAndRecoversPendingTasks(t *testing.T) {
@@ -161,6 +163,40 @@ func TestTaskQueueSnapshotSortsNewestFirst(t *testing.T) {
 	}
 }
 
+func TestNewJobIDUsesTASKPrefixAndZeroPadding(t *testing.T) {
+	globalQueue.mu.Lock()
+	origTasks := globalQueue.tasks
+	globalQueue.tasks = nil
+	globalQueue.mu.Unlock()
+	origCounter := jobCounter.Load()
+	jobCounter.Store(0)
+	t.Cleanup(func() {
+		globalQueue.mu.Lock()
+		globalQueue.tasks = origTasks
+		globalQueue.mu.Unlock()
+		jobCounter.Store(origCounter)
+	})
+
+	if got := newJobID("ignored"); got != "TASK-000" {
+		t.Fatalf("id=%q want TASK-000", got)
+	}
+	if got := newJobID("ignored"); got != "TASK-001" {
+		t.Fatalf("id=%q want TASK-001", got)
+	}
+}
+
+func TestTaskArtifactsDirStartsWithTaskNumber(t *testing.T) {
+	root := t.TempDir()
+	task := &Task{
+		ID:   "TASK-007",
+		Name: "NVIDIA Benchmark",
+	}
+	got := filepath.Base(taskArtifactsDir(root, task, TaskDone))
+	if !strings.HasPrefix(got, "007_") {
+		t.Fatalf("artifacts dir=%q want prefix 007_", got)
+	}
+}
+
 func TestHandleAPITasksStreamReplaysPersistedLogWithoutLiveJob(t *testing.T) {
 	dir := t.TempDir()
 	logPath := filepath.Join(dir, "task.log")
@@ -248,15 +284,205 @@ func TestHandleAPITasksStreamPendingTaskStartsSSEImmediately(t *testing.T) {
 	t.Fatalf("stream did not emit queued status promptly, body=%q", rec.Body.String())
 }
 
+func TestFinalizeTaskRunCreatesReportFolderAndArtifacts(t *testing.T) {
+	dir := t.TempDir()
+	metricsPath := filepath.Join(dir, "metrics.db")
+	prevMetricsPath := taskReportMetricsDBPath
+	taskReportMetricsDBPath = metricsPath
+	t.Cleanup(func() { taskReportMetricsDBPath = prevMetricsPath })
+
+	db, err := openMetricsDB(metricsPath)
+	if err != nil {
+		t.Fatalf("openMetricsDB: %v", err)
+	}
+	base := time.Now().UTC().Add(-45 * time.Second)
+	if err := db.Write(platform.LiveMetricSample{
+		Timestamp:  base,
+		CPULoadPct: 42,
+		MemLoadPct: 35,
+		PowerW:     510,
+	}); err != nil {
+		t.Fatalf("Write: %v", err)
+	}
+	_ = db.Close()
+
+	q := &taskQueue{
+		statePath: filepath.Join(dir, "tasks-state.json"),
+		logsDir:   filepath.Join(dir, "tasks"),
+		trigger:   make(chan struct{}, 1),
+	}
+	if err := os.MkdirAll(q.logsDir, 0755); err != nil {
+		t.Fatal(err)
+	}
+
+	started := time.Now().UTC().Add(-90 * time.Second)
+	task := &Task{
+		ID:        "task-1",
+		Name:      "CPU SAT",
+		Target:    "cpu",
+		Status:    TaskRunning,
+		CreatedAt: started.Add(-10 * time.Second),
+		StartedAt: &started,
+	}
+	q.assignTaskLogPathLocked(task)
+	appendJobLog(task.LogPath, "line-1")
+
+	job := newTaskJobState(task.LogPath)
+	job.finish("")
+	q.finalizeTaskRun(task, job)
+
+	if task.Status != TaskDone {
+		t.Fatalf("status=%q want %q", task.Status, TaskDone)
+	}
+	if !strings.Contains(filepath.Base(task.ArtifactsDir), "_done") {
+		t.Fatalf("artifacts dir=%q", task.ArtifactsDir)
+	}
+	if _, err := os.Stat(task.ReportJSONPath); err != nil {
+		t.Fatalf("report json: %v", err)
+	}
+	if _, err := os.Stat(task.ReportHTMLPath); err != nil {
+		t.Fatalf("report html: %v", err)
+	}
+	var report taskReport
+	data, err := os.ReadFile(task.ReportJSONPath)
+	if err != nil {
+		t.Fatalf("ReadFile(report.json): %v", err)
+	}
+	if err := json.Unmarshal(data, &report); err != nil {
+		t.Fatalf("Unmarshal(report.json): %v", err)
+	}
+	if report.ID != task.ID || report.Status != TaskDone {
+		t.Fatalf("report=%+v", report)
+	}
+	if len(report.Charts) == 0 {
+		t.Fatalf("expected charts in report, got none")
+	}
+}
+
+func TestWriteTaskReportArtifactsIncludesBenchmarkResultsForTask(t *testing.T) {
+	dir := t.TempDir()
+	metricsPath := filepath.Join(dir, "metrics.db")
+	prevMetricsPath := taskReportMetricsDBPath
+	taskReportMetricsDBPath = metricsPath
+	t.Cleanup(func() { taskReportMetricsDBPath = prevMetricsPath })
+
+	benchmarkDir := filepath.Join(dir, "bee-benchmark", "gpu-benchmark-20260406-120000")
+	if err := os.MkdirAll(benchmarkDir, 0755); err != nil {
+		t.Fatal(err)
+	}
+	result := platform.NvidiaBenchmarkResult{
+		GeneratedAt:      time.Date(2026, time.April, 6, 12, 0, 0, 0, time.UTC),
+		BenchmarkProfile: "standard",
+		OverallStatus:    "OK",
+		GPUs: []platform.BenchmarkGPUResult{
+			{
+				Index: 0,
+				Name:  "NVIDIA H100 PCIe",
+				Scores: platform.BenchmarkScorecard{
+					CompositeScore: 1176.25,
+				},
+			},
+		},
+	}
+	raw, err := json.Marshal(result)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(filepath.Join(benchmarkDir, "result.json"), raw, 0644); err != nil {
+		t.Fatal(err)
+	}
+
+	artifactsDir := filepath.Join(dir, "tasks", "task-bench_done")
+	if err := os.MkdirAll(artifactsDir, 0755); err != nil {
+		t.Fatal(err)
+	}
+	task := &Task{
+		ID:           "task-bench",
+		Name:         "NVIDIA Benchmark",
+		Target:       "nvidia-benchmark",
+		Status:       TaskDone,
+		CreatedAt:    time.Now().UTC().Add(-time.Minute),
+		ArtifactsDir: artifactsDir,
+	}
+	ensureTaskReportPaths(task)
+	logText := "line-1\nArchive: " + filepath.Join(dir, "bee-benchmark", "gpu-benchmark-20260406-120000.tar.gz") + "\n"
+	if err := os.WriteFile(task.LogPath, []byte(logText), 0644); err != nil {
+		t.Fatal(err)
+	}
+
+	if err := writeTaskReportArtifacts(task); err != nil {
+		t.Fatalf("writeTaskReportArtifacts: %v", err)
+	}
+
+	body, err := os.ReadFile(task.ReportHTMLPath)
+	if err != nil {
+		t.Fatalf("ReadFile(report.html): %v", err)
+	}
+	html := string(body)
+	for _, needle := range []string{
+		`Benchmark Results`,
+		`Composite score for this benchmark task.`,
+		`GPU #0 — NVIDIA H100 PCIe`,
+		`1176.25`,
+	} {
+		if !strings.Contains(html, needle) {
+			t.Fatalf("report missing %q: %s", needle, html)
+		}
+	}
+}
+
+func TestTaskLifecycleMirrorsToSerialConsole(t *testing.T) {
+	var lines []string
+	prev := taskSerialWriteLine
+	taskSerialWriteLine = func(line string) { lines = append(lines, line) }
+	t.Cleanup(func() { taskSerialWriteLine = prev })
+
+	dir := t.TempDir()
+	q := &taskQueue{
+		statePath: filepath.Join(dir, "tasks-state.json"),
+		logsDir:   filepath.Join(dir, "tasks"),
+		trigger:   make(chan struct{}, 1),
+	}
+	task := &Task{
+		ID:        "task-serial-1",
+		Name:      "CPU SAT",
+		Target:    "cpu",
+		Status:    TaskPending,
+		CreatedAt: time.Now().UTC(),
+	}
+
+	q.enqueue(task)
+	started := time.Now().UTC()
+	task.Status = TaskRunning
+	task.StartedAt = &started
+	job := newTaskJobState(task.LogPath, taskSerialPrefix(task))
+	job.append("Starting CPU SAT...")
+	job.append("CPU stress duration: 60s")
+	job.finish("")
+	q.finalizeTaskRun(task, job)
+
+	joined := strings.Join(lines, "\n")
+	for _, needle := range []string{
+		"queued",
+		"Starting CPU SAT...",
+		"CPU stress duration: 60s",
+		"finished with status=done",
+	} {
+		if !strings.Contains(joined, needle) {
+			t.Fatalf("serial mirror missing %q in %q", needle, joined)
+		}
+	}
+}
+
 func TestResolveBurnPreset(t *testing.T) {
 	tests := []struct {
 		profile string
 		want    burnPreset
 	}{
-		{profile: "smoke", want: burnPreset{NvidiaDiag: 1, DurationSec: 5 * 60}},
-		{profile: "acceptance", want: burnPreset{NvidiaDiag: 3, DurationSec: 60 * 60}},
-		{profile: "overnight", want: burnPreset{NvidiaDiag: 4, DurationSec: 8 * 60 * 60}},
-		{profile: "", want: burnPreset{NvidiaDiag: 1, DurationSec: 5 * 60}},
+		{profile: "smoke", want: burnPreset{DurationSec: 5 * 60}},
+		{profile: "acceptance", want: burnPreset{DurationSec: 60 * 60}},
+		{profile: "overnight", want: burnPreset{DurationSec: 8 * 60 * 60}},
+		{profile: "", want: burnPreset{DurationSec: 5 * 60}},
 	}
 	for _, tc := range tests {
 		if got := resolveBurnPreset(tc.profile); got != tc.want {
@@ -467,3 +693,52 @@ func TestRunTaskInstallUsesSharedCommandStreaming(t *testing.T) {
 		t.Fatalf("unexpected error: %q", j.err)
 	}
 }
+
+func TestExecuteTaskMarksPanicsAsFailedAndClosesKmsgWindow(t *testing.T) {
+	dir := t.TempDir()
+	q := &taskQueue{
+		opts:        &HandlerOptions{App: &app.App{}},
+		statePath:   filepath.Join(dir, "tasks-state.json"),
+		logsDir:     filepath.Join(dir, "tasks"),
+		kmsgWatcher: newKmsgWatcher(nil),
+	}
+	tk := &Task{
+		ID:        "cpu-panic-1",
+		Name:      "CPU SAT",
+		Target:    "cpu",
+		Status:    TaskRunning,
+		CreatedAt: time.Now(),
+	}
+	j := &jobState{}
+
+	orig := runCPUAcceptancePackCtx
+	runCPUAcceptancePackCtx = func(_ *app.App, _ context.Context, _ string, _ int, _ func(string)) (string, error) {
+		panic("boom")
+	}
+	defer func() { runCPUAcceptancePackCtx = orig }()
+
+	q.executeTask(tk, j, context.Background())
+
+	if tk.Status != TaskFailed {
+		t.Fatalf("status=%q want %q", tk.Status, TaskFailed)
+	}
+	if tk.DoneAt == nil {
+		t.Fatal("expected done_at to be set")
+	}
+	if !strings.Contains(tk.ErrMsg, "task panic: boom") {
+		t.Fatalf("task error=%q", tk.ErrMsg)
+	}
+	if !strings.Contains(j.err, "task panic: boom") {
+		t.Fatalf("job error=%q", j.err)
+	}
+	q.kmsgWatcher.mu.Lock()
+	activeCount := q.kmsgWatcher.activeCount
+	window := q.kmsgWatcher.window
+	q.kmsgWatcher.mu.Unlock()
+	if activeCount != 0 {
+		t.Fatalf("activeCount=%d want 0", activeCount)
+	}
+	if window != nil {
+		t.Fatalf("expected kmsg window to be cleared, got %+v", window)
+	}
+}

audit/scripts/resolve-version.sh

@@ -0,0 +1,16 @@
+#!/bin/sh
+set -eu
+
+tag="$(git describe --tags --match 'v[0-9]*' --abbrev=7 --dirty 2>/dev/null || true)"
+
+case "${tag}" in
+	v*)
+		printf '%s\n' "${tag#v}"
+		;;
+	"")
+		printf 'dev\n'
+		;;
+	*)
+		printf '%s\n' "${tag}"
+		;;
+esac

bible-local/docs/benchmark-clock-calibration.md

@@ -0,0 +1,248 @@
+# Benchmark clock calibration research
+
+## Status
+In progress. Baseline data from production servers pending.
+
+## Background
+
+The benchmark locks GPU clocks to `MaxGraphicsClockMHz` (boost) via `nvidia-smi -lgc`
+before the steady-state phase. The metric `low_sm_clock_vs_target` fires when
+`avg_steady_clock < locked_target * 0.90`.
+
+Problem: boost clock is the theoretical maximum under ideal cooling. In practice,
+even a healthy GPU in a non-ideal server will sustain clocks well below boost.
+The 90% threshold has no empirical basis.
+
+## Key observations (2026-04-06)
+
+### H100 PCIe — new card, server not designed for it
+- avg clock 1384 MHz, P95 1560 MHz (unstable, proba boost 1755 MHz)
+- Thermal sustain: 0.0 (sw_thermal covers entire steady window)
+- Stability: 70.0 — clocks erratic, no equilibrium found
+- Degradation: power_capped, thermal_limited, low_sm_clock_vs_target, variance_too_high
+
+### H200 NVL — new card, server not designed for it
+- avg clock = P95 = 1635 MHz (perfectly stable)
+- Thermal sustain: 0.0 (sw_thermal + sw_power cover entire steady window)
+- Stability: 92.0 — found stable thermal equilibrium at 1635 MHz
+- Degradation: power_capped, thermal_limited
+- Compute: 989 TOPS — card is computing correctly for its frequency
+
+### Key insight
+The meaningful distinction is not *whether* the card throttles but *how stably*
+it throttles. H200 found a thermal equilibrium (avg == P95, Stability 92),
+H100 did not (avg << P95, Stability 70). Both are new cards; the H100's
+instability may reflect a more severe thermal mismatch or a card issue.
+
+`sw_power ≈ sw_thermal` pattern = server cooling constraint, card likely OK.
+`hw_thermal >> sw_thermal` pattern = card itself overheating, investigate.
+
+## Hypothesis for baseline
+
+After testing on servers designed for their GPUs (proper cooling):
+- Healthy GPU under sustained load will run at a stable fraction of boost
+- Expected: avg_steady ≈ 80–95% of boost depending on model and TDP class
+- Base clock (`clocks.base.gr`) may be a better reference than boost:
+  a healthy card under real workload should comfortably exceed base clock
+
+## Baseline: H100 PCIe HBM2e — designed server (2026-04-06, 10 samples)
+
+Source: external stress test tool, ~90s runs, designed server, adequate power.
+
+### Healthy fingerprint
+
+- **Power**: hits cap ~340–360W immediately, stays flat throughout — HEALTHY
+- **Clock**: starts ~1750 MHz, oscillates and declines to ~1540–1600 MHz by 90s
+  - Avg steady (visual): **~1580–1620 MHz**
+  - vs boost 1755 MHz: **~91–92%**
+  - Oscillation is NORMAL — this is the boost algorithm balancing under power cap
+  - Stable power + oscillating clocks = healthy power-cap behavior
+- **Temperature**: linear rise ~38°C → 75–80°C over 90s (no runaway)
+- **Consistency**: all 10 samples within ±20 MHz — very repeatable
+
+### Characteristic patten
+Flat power line + oscillating/declining clock line = GPU correctly managed by
+power cap algorithm. Do NOT flag this as instability.
+
+### Clock CV implication
+The healthy oscillation WILL produce moderate ClockCVPct (~5–10%).
+The current `variance_too_high` threshold (StabilityScore < 85) may fire on
+healthy HBM2e PCIe cards. Needs recalibration.
+
+---
+
+## Baseline: H100 HBM3 OEM SXM Custom (restored) — 2 confirmed samples
+
+Source: pytorch_training_loop stress test, 120s (90s stress + 30s cooldown).
+Confirmed GPU: NVIDIA H100 80GB HBM3, GH100 rev a1.
+
+### GPU clock reference (from nvidia-smi, idle):
+- base_clock_mhz: **1095**
+- boost_clock_mhz: **1755** (nvidia-smi `clocks.max.graphics` at idle)
+- achieved_max_clock_mhz: **1980** (actual burst max observed by tool)
+- Our benchmark locks to `clocks.max.graphics` = likely 1980 MHz for this chip
+
+### Observed under 700W sustained load (both samples nearly identical):
+- Power: ~700W flat — SXM slot, adequate power confirmed
+- Clock steady range: **~1380–1480 MHz**, avg **~1420–1460 MHz**
+- vs 1980 MHz (lock target): **72–74%** — severely below
+- vs 1755 MHz (nvidia-smi boost): **81–83%**
+- vs 1095 MHz (base): 130% — above base but far below expected for SXM
+- Clock/Watt: ~2.1 MHz/W vs HBM2e ~4.6 MHz/W — 2× worse efficiency
+- Temperature: 38°C → 79–80°C (same rate as HBM2e)
+- Oscillation: present, similar character to HBM2e but at much lower frequency
+
+### Diagnosis
+These restored cards are degraded. A healthy H100 SXM in a designed server
+(DGX H100, HGX H100) should sustain ~1800–1900 MHz at 700W (~91–96% of 1980).
+The 72–74% result is a clear signal of silicon or VRM degradation from the
+refurbishment process.
+
+### Clock pattern note
+Images 8/9 (previously marked as "HBM3 restored") are now confirmed identical
+to images 19/20. Both sample sets show same degraded pattern — same batch.
+
+---
+
+## Baseline matrix (filled where data available)
+
+| GPU model | Config | Avg clock steady | vs boost | Clock/Watt | Notes |
+|---|---|---|---|---|---|
+| H100 PCIe HBM2e | designed server | 1580–1620 MHz | 91–92% | ~4.6 MHz/W | 10 samples, healthy |
+| H100 SXM HBM3 restored | 700W full | 1420–1460 MHz | 72–74% of 1980 | ~2.1 MHz/W | 4 samples confirmed, degraded |
+| H100 SXM HBM3 healthy | designed | ~1800–1900 MHz est. | ~91–96% est. | ~2.7 MHz/W est. | need real baseline |
+| H200 NVL | designed | TBD | TBD | TBD | need baseline |
+
+---
+
+## H100 official spec (from NVIDIA datasheet)
+
+Source: NVIDIA H100 Tensor Core GPU Datasheet (image 23, 2026-04-06).
+All TOPS marked * are with structural sparsity enabled. Divide by 2 for dense.
+
+| Model | FP16 Tensor (dense) | TF32 (dense) | FP8 (dense) | TDP | Memory |
+|---|---|---|---|---|---|
+| H100 80GB PCIe | 756 TFLOPS | 378 TFLOPS | 1,513 TFLOPS | 350W | HBM2e |
+| H100 NVL 94GB PCIe | 990 TFLOPS | 495 TFLOPS | 1,980 TFLOPS | 400W | HBM3 |
+| H100 80GB SXM (BQQV) | 989 TFLOPS | 494 TFLOPS | — | 700W | HBM3 |
+| H100 94GB SXM (BUBB) | 989 TFLOPS | 494 TFLOPS | — | 700W | HBM2e |
+
+Notes:
+- SXM boards do NOT list FP8 peak in this table (field empty)
+- fp8_e5m2 is unsupported on H100 PCIe HBM2e — confirmed in our tests
+- Tensor Cores: PCIe = 456, SXM = 528 (16% more on SXM)
+
+## Observed efficiency (H100 80GB PCIe, throttled server)
+
+From the report in this session (power+thermal throttle throughout steady):
+
+| Precision | Measured | Spec (dense) | % of spec |
+|---|---|---|---|
+| fp16_tensor | 329 TOPS | 756 TFLOPS | 44% |
+| fp32_tf32 | 115 TOPS | 378 TFLOPS | 30% |
+| fp8_e4m3 | 505 TOPS | 1,513 TFLOPS | 33% |
+
+33–44% of spec is expected given sustained power+thermal throttle (avg clock
+1384 MHz vs boost 1755 MHz = 79%). The GPU is computing correctly for its
+actual frequency — the low TOPS comes from throttle, not silicon defect.
+
+## H200 official spec (from NVIDIA datasheet, image 24, 2026-04-06)
+
+Format: without sparsity / with sparsity.
+
+| Model | FP16 Tensor (dense) | TF32 (dense) | FP8 (dense) | TDP | Memory |
+|---|---|---|---|---|---|
+| H200 NVL PCIe | 836 TFLOPS | 418 TFLOPS | 1,570 TFLOPS | 600W | HBM3e 141GB |
+| H200 SXM | 990 TFLOPS | 495 TFLOPS | 1,979 TFLOPS | 700W | HBM3e 141GB |
+
+## Observed efficiency (H200 NVL PCIe, throttled non-designed server)
+
+Avg clock 1635 MHz (62% of boost ~2619 MHz). Entire steady in thermal throttle.
+
+| Precision | Measured | Spec (dense) | % of spec |
+|---|---|---|---|
+| fp16_tensor | 340 TOPS | 836 TFLOPS | 41% |
+| fp32_tf32 | 120 TOPS | 418 TFLOPS | 29% |
+| fp8_e4m3 | 529 TOPS | 1,570 TFLOPS | 34% |
+
+Comparable to H100 PCIe efficiency (33–44%) despite different architecture —
+both are throttle-limited. Confirms that % of spec is not a quality signal,
+it reflects the thermal environment. tops_per_sm_per_ghz is the right metric.
+
+## Real-world GEMM efficiency reference (2026-04-06, web research)
+
+Sources: SemiAnalysis MI300X vs H100 vs H200 training benchmark; cuBLAS optimization
+worklog (hamzaelshafie.bearblog.dev); Lambda AI H100 performance analysis.
+
+### What healthy systems actually achieve:
+- H100 SXM in designed server: **~720 TFLOPS FP16 = ~73% of spec**
+- cuBLAS large square GEMM (8192³): up to **~83% flop utilization**
+- H200 NVL PCIe: no public data, extrapolating ~73% → ~610 TFLOPS FP16
+
+### Our results vs expectation:
+| GPU | Our FP16 | Expected (73%) | Our % of spec | Gap |
+|---|---|---|---|---|
+| H100 PCIe HBM2e | 329 TOPS | ~552 TFLOPS | 44% | ~1.7× below |
+| H200 NVL PCIe | 340 TOPS | ~610 TFLOPS | 41% | ~1.8× below |
+
+Our results are roughly **half** of what a healthy system achieves even under throttle.
+This is NOT normal — 30-44% is not the industry baseline.
+
+### Likely causes of the gap (in order of probability):
+1. **Thermal throttle** — confirmed, sw_thermal covers entire steady window
+2. **Power limit below TDP** — GPU may be software-limited below 350W/600W.
+   Previous user may have set a lower limit via nvidia-smi -pl and it was not
+   reset. Our normalization sets clock locks but does NOT reset power limit.
+   Key check: `nvidia-smi -q | grep "Power Limit"` — default vs enforced.
+3. **Matrix size** — ruled out. bee-gpu-burn uses 4096×4096×4096 for fp16,
+   8192×8192×4096 for fp8. These are large enough for peak tensor utilization.
+
+### Power limit gap analysis (H100 PCIe):
+- Avg clock 1384 MHz = 79% of boost 1755 MHz
+- Expected TOPS at 79% clock: 756 × 0.79 ≈ 597 TFLOPS
+- Actually measured: 329 TOPS = 55% of that estimate
+- Remaining gap after accounting for clock throttle: ~45%
+- Most likely explanation: enforced power limit < 350W TDP, further reducing
+  sustainable clock beyond what sw_thermal alone would cause.
+
+### Action item:
+Add `power.limit` (enforced) AND `power.default_limit` to queryBenchmarkGPUInfo
+so result.json shows if the card was pre-configured with a non-default limit.
+If enforced < default × 0.95 → add finding "GPU power limit is below default TDP".
+
+### CPU/RAM impact on GPU FLOPS:
+None. Pure on-GPU GEMM is fully compute-bound once data is in VRAM.
+CPU core count and host RAM are irrelevant.
+
+## Compute efficiency metric (proposed, no hardcode)
+
+Instead of comparing TOPS to a hardcoded spec, compute:
+  tops_per_sm_per_ghz = measured_tops / (sm_count × avg_clock_ghz)
+
+This is model-agnostic. A GPU computing correctly at its actual frequency
+will show a consistent tops_per_sm_per_ghz regardless of throttle level.
+A GPU with degraded silicon will show low tops_per_sm_per_ghz even at
+normal clocks.
+
+SM count is queryable: nvidia-smi --query-gpu=attribute.multiprocessor_count
+(needs to be added to queryBenchmarkGPUInfo).
+
+Reference values to establish after baseline runs:
+- H100 PCIe fp16_tensor: TBD tops/SM/GHz
+- H100 SXM fp16_tensor: TBD tops/SM/GHz
+
+## Proposed threshold changes (pending more data)
+
+1. **`low_sm_clock_vs_target`**: raise threshold from 90% to 85% based on observed
+   91–92% on healthy HBM2e. Or remove entirely — sw_power/sw_thermal already
+   capture the root cause.
+
+2. **`variance_too_high`** (StabilityScore < 85): healthy HBM2e WILL oscillate
+   under power cap. Consider suppressing this flag when power is flat and usage
+   is 100% (oscillation is expected). Or lower threshold to 70.
+
+3. **New signal: MHz/Watt efficiency**: if base_graphics_clock_mhz is available,
+   ratio avg_clock / power_w could identify degraded silicon (HBM3 restored S1
+   would have been caught by this).
+
+Decision deferred until baseline on SXM designed servers collected.

iso/builder/auto/config

@@ -32,7 +32,7 @@ lb config noauto \
     --memtest memtest86+ \
     --iso-volume "EASY_BEE_${BEE_GPU_VENDOR_UPPER:-NVIDIA}" \
     --iso-application "EASY-BEE-${BEE_GPU_VENDOR_UPPER:-NVIDIA}" \
-    --bootappend-live "boot=live components video=1920x1080 console=tty0 console=ttyS0,115200n8 loglevel=3 username=bee user-fullname=Bee modprobe.blacklist=nouveau,snd_hda_intel,snd_hda_codec_realtek,snd_hda_codec_generic,soundcore" \
+    --bootappend-live "boot=live components video=1920x1080 console=tty0 console=ttyS0,115200n8 loglevel=3 systemd.show_status=1 username=bee user-fullname=Bee modprobe.blacklist=nouveau,snd_hda_intel,snd_hda_codec_realtek,snd_hda_codec_generic,soundcore" \
     --apt-recommends false \
     --chroot-squashfs-compression-type zstd \
     "${@}"

iso/builder/bee-gpu-stress.c

@@ -36,7 +36,6 @@ typedef void *CUstream;
 #define MAX_CUBLAS_PROFILES 5
 #define MIN_PROFILE_BUDGET_BYTES ((size_t)4u * 1024u * 1024u)
 #define MIN_STREAM_BUDGET_BYTES ((size_t)64u * 1024u * 1024u)
-#define STRESS_LAUNCH_DEPTH 8
 
 static const char *ptx_source =
     ".version 6.0\n"
@@ -344,7 +343,6 @@ static int run_ptx_fallback(struct cuda_api *api,
     unsigned long iterations = 0;
     int mp_count = 0;
     int stream_count = 1;
-    int launches_per_wave = 0;
 
     memset(report, 0, sizeof(*report));
     snprintf(report->backend, sizeof(report->backend), "driver-ptx");
@@ -419,44 +417,42 @@ static int run_ptx_fallback(struct cuda_api *api,
 
     unsigned int threads = 256;
 
-    double start = now_seconds();
-    double deadline = start + (double)seconds;
+    double deadline = now_seconds() + (double)seconds;
+    double next_sync = now_seconds() + 1.0;
     while (now_seconds() < deadline) {
-        launches_per_wave = 0;
-        for (int depth = 0; depth < STRESS_LAUNCH_DEPTH && now_seconds() < deadline; depth++) {
-            int launched_this_batch = 0;
-            for (int lane = 0; lane < stream_count; lane++) {
-                unsigned int blocks = (unsigned int)((words[lane] + threads - 1) / threads);
-                if (!check_rc(api,
-                              "cuLaunchKernel",
-                              api->cuLaunchKernel(kernel,
-                                                  blocks,
-                                                  1,
-                                                  1,
-                                                  threads,
-                                                  1,
-                                                  1,
-                                                  0,
-                                                  streams[lane],
-                                                  params[lane],
-                                                  NULL))) {
-                    goto fail;
-                }
-                launches_per_wave++;
-                launched_this_batch++;
-            }
-            if (launched_this_batch <= 0) {
-                break;
+        int launched = 0;
+        for (int lane = 0; lane < stream_count; lane++) {
+            unsigned int blocks = (unsigned int)((words[lane] + threads - 1) / threads);
+            if (!check_rc(api,
+                          "cuLaunchKernel",
+                          api->cuLaunchKernel(kernel,
+                                              blocks,
+                                              1,
+                                              1,
+                                              threads,
+                                              1,
+                                              1,
+                                              0,
+                                              streams[lane],
+                                              params[lane],
+                                              NULL))) {
+                goto fail;
             }
+            launched++;
+            iterations++;
         }
-        if (launches_per_wave <= 0) {
+        if (launched <= 0) {
             goto fail;
         }
-        if (!check_rc(api, "cuCtxSynchronize", api->cuCtxSynchronize())) {
-            goto fail;
+        double now = now_seconds();
+        if (now >= next_sync || now >= deadline) {
+            if (!check_rc(api, "cuCtxSynchronize", api->cuCtxSynchronize())) {
+                goto fail;
+            }
+            next_sync = now + 1.0;
         }
-        iterations += (unsigned long)launches_per_wave;
     }
+    api->cuCtxSynchronize();
 
     if (!check_rc(api, "cuMemcpyDtoH", api->cuMemcpyDtoH(sample, device_mem[0], sizeof(sample)))) {
         goto fail;
@@ -468,11 +464,10 @@ static int run_ptx_fallback(struct cuda_api *api,
     report->iterations = iterations;
     snprintf(report->details,
              sizeof(report->details),
-             "fallback_int32=OK requested_mb=%d actual_mb=%d streams=%d queue_depth=%d per_stream_mb=%zu iterations=%lu\n",
+             "fallback_int32=OK requested_mb=%d actual_mb=%d streams=%d per_stream_mb=%zu iterations=%lu\n",
              size_mb,
              report->buffer_mb,
              report->stream_count,
-             STRESS_LAUNCH_DEPTH,
              bytes_per_stream[0] / (1024u * 1024u),
              iterations);
 
@@ -606,6 +601,20 @@ struct prepared_profile {
 };
 
 static const struct profile_desc k_profiles[] = {
+    {
+        "fp64",
+        "fp64",
+        80,
+        1,
+        0,
+        0,
+        8,
+        CUDA_R_64F,
+        CUDA_R_64F,
+        CUDA_R_64F,
+        CUDA_R_64F,
+        CUBLAS_COMPUTE_64F,
+    },
     {
         "fp32_tf32",
         "fp32",
@@ -1126,7 +1135,6 @@ static int run_cublaslt_stress(struct cuda_api *cuda,
     int stream_count = 1;
     int profile_count = (int)(sizeof(k_profiles) / sizeof(k_profiles[0]));
     int prepared_count = 0;
-    int wave_launches = 0;
     size_t requested_budget = 0;
     size_t total_budget = 0;
     size_t per_profile_budget = 0;
@@ -1193,11 +1201,10 @@ static int run_cublaslt_stress(struct cuda_api *cuda,
     report->buffer_mb = (int)(total_budget / (1024u * 1024u));
     append_detail(report->details,
                   sizeof(report->details),
-                  "requested_mb=%d actual_mb=%d streams=%d queue_depth=%d mp_count=%d per_worker_mb=%zu\n",
+                  "requested_mb=%d actual_mb=%d streams=%d mp_count=%d per_worker_mb=%zu\n",
                   size_mb,
                   report->buffer_mb,
                   report->stream_count,
-                  STRESS_LAUNCH_DEPTH,
                   mp_count,
                   per_profile_budget / (1024u * 1024u));
 
@@ -1246,50 +1253,55 @@ static int run_cublaslt_stress(struct cuda_api *cuda,
         return 0;
     }
 
+    /* Keep the GPU queue continuously full by submitting kernels without
+     * synchronizing after every wave.  A sync barrier after each small batch
+     * creates CPU↔GPU ping-pong gaps that prevent full TDP utilisation,
+     * especially when individual kernels are short.  Instead we sync at most
+     * once per second (for error detection) and once at the very end. */
     double deadline = now_seconds() + (double)seconds;
+    double next_sync = now_seconds() + 1.0;
     while (now_seconds() < deadline) {
-        wave_launches = 0;
-        for (int depth = 0; depth < STRESS_LAUNCH_DEPTH && now_seconds() < deadline; depth++) {
-            int launched_this_batch = 0;
-            for (int i = 0; i < prepared_count; i++) {
-                if (!prepared[i].ready) {
-                    continue;
-                }
-                if (!run_cublas_profile(handle, &cublas, &prepared[i])) {
-                    append_detail(report->details,
-                                  sizeof(report->details),
-                                  "%s=FAILED runtime\n",
-                                  prepared[i].desc.name);
-                    for (int j = 0; j < prepared_count; j++) {
-                        destroy_profile(&cublas, cuda, &prepared[j]);
-                    }
-                    cublas.cublasLtDestroy(handle);
-                    destroy_streams(cuda, streams, stream_count);
-                    cuda->cuCtxDestroy(ctx);
-                    return 0;
-                }
-                prepared[i].iterations++;
-                report->iterations++;
-                wave_launches++;
-                launched_this_batch++;
+        int launched = 0;
+        for (int i = 0; i < prepared_count; i++) {
+            if (!prepared[i].ready) {
+                continue;
             }
-            if (launched_this_batch <= 0) {
-                break;
+            if (!run_cublas_profile(handle, &cublas, &prepared[i])) {
+                append_detail(report->details,
+                              sizeof(report->details),
+                              "%s=FAILED runtime\n",
+                              prepared[i].desc.name);
+                for (int j = 0; j < prepared_count; j++) {
+                    destroy_profile(&cublas, cuda, &prepared[j]);
+                }
+                cublas.cublasLtDestroy(handle);
+                destroy_streams(cuda, streams, stream_count);
+                cuda->cuCtxDestroy(ctx);
+                return 0;
             }
+            prepared[i].iterations++;
+            report->iterations++;
+            launched++;
         }
-        if (wave_launches <= 0) {
+        if (launched <= 0) {
             break;
         }
-        if (!check_rc(cuda, "cuCtxSynchronize", cuda->cuCtxSynchronize())) {
-            for (int i = 0; i < prepared_count; i++) {
-                destroy_profile(&cublas, cuda, &prepared[i]);
+        double now = now_seconds();
+        if (now >= next_sync || now >= deadline) {
+            if (!check_rc(cuda, "cuCtxSynchronize", cuda->cuCtxSynchronize())) {
+                for (int i = 0; i < prepared_count; i++) {
+                    destroy_profile(&cublas, cuda, &prepared[i]);
+                }
+                cublas.cublasLtDestroy(handle);
+                destroy_streams(cuda, streams, stream_count);
+                cuda->cuCtxDestroy(ctx);
+                return 0;
             }
-            cublas.cublasLtDestroy(handle);
-            destroy_streams(cuda, streams, stream_count);
-            cuda->cuCtxDestroy(ctx);
-            return 0;
+            next_sync = now + 1.0;
         }
     }
+    /* Final drain — ensure all queued work finishes before we read results. */
+    cuda->cuCtxSynchronize();
 
     for (int i = 0; i < prepared_count; i++) {
         if (!prepared[i].ready) {

iso/builder/build-in-container.sh

@@ -41,15 +41,15 @@ while [ $# -gt 0 ]; do
             ;;
         *)
             echo "unknown arg: $1" >&2
-            echo "usage: $0 [--cache-dir /path] [--rebuild-image] [--clean-build] [--authorized-keys /path/to/authorized_keys] [--variant nvidia|amd|all]" >&2
+            echo "usage: $0 [--cache-dir /path] [--rebuild-image] [--clean-build] [--authorized-keys /path/to/authorized_keys] [--variant nvidia|nvidia-legacy|amd|nogpu|all]" >&2
             exit 1
             ;;
     esac
 done
 
 case "$VARIANT" in
-    nvidia|amd|nogpu|all) ;;
-    *) echo "unknown variant: $VARIANT (expected nvidia, amd, nogpu, or all)" >&2; exit 1 ;;
+    nvidia|nvidia-legacy|amd|nogpu|all) ;;
+    *) echo "unknown variant: $VARIANT (expected nvidia, nvidia-legacy, amd, nogpu, or all)" >&2; exit 1 ;;
 esac
 
 if [ "$CLEAN_CACHE" = "1" ]; then
@@ -61,8 +61,13 @@ if [ "$CLEAN_CACHE" = "1" ]; then
            "${CACHE_DIR:?}/lb-packages"
     echo "=== cleaning live-build work dirs ==="
     rm -rf "${REPO_ROOT}/dist/live-build-work-nvidia"
+    rm -rf "${REPO_ROOT}/dist/live-build-work-nvidia-legacy"
     rm -rf "${REPO_ROOT}/dist/live-build-work-amd"
     rm -rf "${REPO_ROOT}/dist/live-build-work-nogpu"
+    rm -rf "${REPO_ROOT}/dist/overlay-stage-nvidia"
+    rm -rf "${REPO_ROOT}/dist/overlay-stage-nvidia-legacy"
+    rm -rf "${REPO_ROOT}/dist/overlay-stage-amd"
+    rm -rf "${REPO_ROOT}/dist/overlay-stage-nogpu"
     echo "=== caches cleared, proceeding with build ==="
 fi
 
@@ -180,6 +185,9 @@ case "$VARIANT" in
     nvidia)
         run_variant nvidia
         ;;
+    nvidia-legacy)
+        run_variant nvidia-legacy
+        ;;
     amd)
         run_variant amd
         ;;
@@ -188,6 +196,7 @@ case "$VARIANT" in
         ;;
     all)
         run_variant nvidia
+        run_variant nvidia-legacy
         run_variant amd
         run_variant nogpu
         ;;

iso/builder/build-nvidia-module.sh

@@ -1,8 +1,10 @@
 #!/bin/sh
-# build-nvidia-module.sh — compile NVIDIA proprietary driver modules for Debian 12
+# build-nvidia-module.sh — compile NVIDIA kernel modules for Debian 12
 #
 # Downloads the official NVIDIA .run installer, extracts kernel modules and
-# userspace tools (nvidia-smi, libnvidia-ml). Everything is proprietary NVIDIA.
+# userspace tools (nvidia-smi, libnvidia-ml). Supports both:
+#   - open         -> kernel-open/ sources from the .run installer
+#   - proprietary  -> traditional proprietary kernel sources from the .run installer
 #
 # Output is cached in DIST_DIR/nvidia-<version>-<kver>/ so subsequent builds
 # are instant unless NVIDIA_DRIVER_VERSION or kernel version changes.
@@ -17,10 +19,19 @@ set -e
 NVIDIA_VERSION="$1"
 DIST_DIR="$2"
 DEBIAN_KERNEL_ABI="$3"
+NVIDIA_FLAVOR="${4:-open}"
 
-[ -n "$NVIDIA_VERSION" ]    || { echo "usage: $0 <nvidia-version> <dist-dir> <debian-kernel-abi>"; exit 1; }
-[ -n "$DIST_DIR" ]          || { echo "usage: $0 <nvidia-version> <dist-dir> <debian-kernel-abi>"; exit 1; }
-[ -n "$DEBIAN_KERNEL_ABI" ] || { echo "usage: $0 <nvidia-version> <dist-dir> <debian-kernel-abi>"; exit 1; }
+[ -n "$NVIDIA_VERSION" ]    || { echo "usage: $0 <nvidia-version> <dist-dir> <debian-kernel-abi> [open|proprietary]"; exit 1; }
+[ -n "$DIST_DIR" ]          || { echo "usage: $0 <nvidia-version> <dist-dir> <debian-kernel-abi> [open|proprietary]"; exit 1; }
+[ -n "$DEBIAN_KERNEL_ABI" ] || { echo "usage: $0 <nvidia-version> <dist-dir> <debian-kernel-abi> [open|proprietary]"; exit 1; }
+
+case "$NVIDIA_FLAVOR" in
+    open|proprietary) ;;
+    *)
+        echo "unsupported NVIDIA flavor: $NVIDIA_FLAVOR (expected open or proprietary)" >&2
+        exit 1
+        ;;
+esac
 
 KVER="${DEBIAN_KERNEL_ABI}-amd64"
 # On Debian, kernel headers are split into two packages:
@@ -31,22 +42,13 @@ KVER="${DEBIAN_KERNEL_ABI}-amd64"
 KDIR_ARCH="/usr/src/linux-headers-${KVER}"
 KDIR_COMMON="/usr/src/linux-headers-${DEBIAN_KERNEL_ABI}-common"
 
-echo "=== NVIDIA ${NVIDIA_VERSION} (proprietary) for kernel ${KVER} ==="
+echo "=== NVIDIA ${NVIDIA_VERSION} (${NVIDIA_FLAVOR}) for kernel ${KVER} ==="
 
-if [ ! -d "$KDIR_ARCH" ] || [ ! -d "$KDIR_COMMON" ]; then
-    echo "=== installing linux-headers-${KVER} ==="
-    DEBIAN_FRONTEND=noninteractive apt-get install -y \
-        "linux-headers-${KVER}" \
-        gcc make perl
-fi
-echo "kernel headers (arch):   $KDIR_ARCH"
-echo "kernel headers (common): $KDIR_COMMON"
-
-CACHE_DIR="${DIST_DIR}/nvidia-${NVIDIA_VERSION}-${KVER}"
+CACHE_DIR="${DIST_DIR}/nvidia-${NVIDIA_FLAVOR}-${NVIDIA_VERSION}-${KVER}"
 CACHE_ROOT="${BEE_CACHE_DIR:-${DIST_DIR}/cache}"
 DOWNLOAD_CACHE_DIR="${CACHE_ROOT}/nvidia-downloads"
 EXTRACT_CACHE_DIR="${CACHE_ROOT}/nvidia-extract"
-CACHE_LAYOUT_VERSION="2"
+CACHE_LAYOUT_VERSION="3"
 CACHE_LAYOUT_MARKER="${CACHE_DIR}/.cache-layout-v${CACHE_LAYOUT_VERSION}"
 if [ -d "$CACHE_DIR/modules" ] && [ -f "$CACHE_DIR/bin/nvidia-smi" ] \
         && [ -f "$CACHE_LAYOUT_MARKER" ] \
@@ -57,6 +59,15 @@ if [ -d "$CACHE_DIR/modules" ] && [ -f "$CACHE_DIR/bin/nvidia-smi" ] \
     exit 0
 fi
 
+if [ ! -d "$KDIR_ARCH" ] || [ ! -d "$KDIR_COMMON" ]; then
+    echo "=== installing linux-headers-${KVER} ==="
+    DEBIAN_FRONTEND=noninteractive apt-get install -y \
+        "linux-headers-${KVER}" \
+        gcc make perl
+fi
+echo "kernel headers (arch):   $KDIR_ARCH"
+echo "kernel headers (common): $KDIR_COMMON"
+
 # Download official NVIDIA .run installer with sha256 verification
 BASE_URL="https://download.nvidia.com/XFree86/Linux-x86_64/${NVIDIA_VERSION}"
 mkdir -p "$DOWNLOAD_CACHE_DIR" "$EXTRACT_CACHE_DIR"
@@ -90,12 +101,18 @@ EXTRACT_DIR="${EXTRACT_CACHE_DIR}/nvidia-extract-${NVIDIA_VERSION}"
 rm -rf "$EXTRACT_DIR"
 "$RUN_FILE" --extract-only --target "$EXTRACT_DIR"
 
-# Find kernel source directory (proprietary: kernel/, open: kernel-open/)
+# Find kernel source directory for the selected flavor.
 KERNEL_SRC=""
-for d in "$EXTRACT_DIR/kernel" "$EXTRACT_DIR/kernel-modules-sources" "$EXTRACT_DIR/kernel-source"; do
-    [ -f "$d/Makefile" ] && KERNEL_SRC="$d" && break
-done
-[ -n "$KERNEL_SRC" ] || { echo "ERROR: kernel source dir not found in:"; ls "$EXTRACT_DIR/"; exit 1; }
+if [ "$NVIDIA_FLAVOR" = "open" ]; then
+    for d in "$EXTRACT_DIR/kernel-open" "$EXTRACT_DIR/kernel-open/"*; do
+        [ -f "$d/Makefile" ] && KERNEL_SRC="$d" && break
+    done
+else
+    for d in "$EXTRACT_DIR/kernel" "$EXTRACT_DIR/kernel-modules-sources" "$EXTRACT_DIR/kernel-source"; do
+        [ -f "$d/Makefile" ] && KERNEL_SRC="$d" && break
+    done
+fi
+[ -n "$KERNEL_SRC" ] || { echo "ERROR: kernel source dir not found for flavor ${NVIDIA_FLAVOR} in:"; ls "$EXTRACT_DIR/"; exit 1; }
 echo "kernel source: $KERNEL_SRC"
 
 # Build kernel modules

iso/builder/build.sh

@@ -15,26 +15,46 @@ DIST_DIR="${REPO_ROOT}/dist"
 VENDOR_DIR="${REPO_ROOT}/iso/vendor"
 CACHE_ROOT="${BEE_CACHE_DIR:-${DIST_DIR}/cache}"
 AUTH_KEYS=""
+BUILD_VARIANT="nvidia"
 BEE_GPU_VENDOR="nvidia"
+BEE_NVIDIA_MODULE_FLAVOR="open"
 
 # parse args
 while [ $# -gt 0 ]; do
     case "$1" in
         --authorized-keys) AUTH_KEYS="$2"; shift 2 ;;
-        --variant) BEE_GPU_VENDOR="$2"; shift 2 ;;
+        --variant) BUILD_VARIANT="$2"; shift 2 ;;
         *) echo "unknown arg: $1"; exit 1 ;;
     esac
 done
 
-case "$BEE_GPU_VENDOR" in
-    nvidia|amd|nogpu) ;;
-    *) echo "unknown variant: $BEE_GPU_VENDOR (expected nvidia, amd, or nogpu)" >&2; exit 1 ;;
+case "$BUILD_VARIANT" in
+    nvidia)
+        BEE_GPU_VENDOR="nvidia"
+        BEE_NVIDIA_MODULE_FLAVOR="open"
+        ;;
+    nvidia-legacy)
+        BEE_GPU_VENDOR="nvidia"
+        BEE_NVIDIA_MODULE_FLAVOR="proprietary"
+        ;;
+    amd)
+        BEE_GPU_VENDOR="amd"
+        BEE_NVIDIA_MODULE_FLAVOR=""
+        ;;
+    nogpu)
+        BEE_GPU_VENDOR="nogpu"
+        BEE_NVIDIA_MODULE_FLAVOR=""
+        ;;
+    *)
+        echo "unknown variant: $BUILD_VARIANT (expected nvidia, nvidia-legacy, amd, or nogpu)" >&2
+        exit 1
+        ;;
 esac
 
-BUILD_WORK_DIR="${DIST_DIR}/live-build-work-${BEE_GPU_VENDOR}"
-OVERLAY_STAGE_DIR="${DIST_DIR}/overlay-stage-${BEE_GPU_VENDOR}"
+BUILD_WORK_DIR="${DIST_DIR}/live-build-work-${BUILD_VARIANT}"
+OVERLAY_STAGE_DIR="${DIST_DIR}/overlay-stage-${BUILD_VARIANT}"
 
-export BEE_GPU_VENDOR
+export BEE_GPU_VENDOR BEE_NVIDIA_MODULE_FLAVOR BUILD_VARIANT
 
 . "${BUILDER_DIR}/VERSIONS"
 export PATH="$PATH:/usr/local/go/bin"
@@ -54,15 +74,8 @@ resolve_audit_version() {
         return 0
     fi
 
-    tag="$(git -C "${REPO_ROOT}" describe --tags --match 'audit/v*' --abbrev=7 --dirty 2>/dev/null || true)"
-    if [ -z "${tag}" ]; then
-        tag="$(git -C "${REPO_ROOT}" describe --tags --match 'v[0-9]*' --abbrev=7 --dirty 2>/dev/null || true)"
-    fi
+    tag="$(git -C "${REPO_ROOT}" describe --tags --match 'v[0-9]*' --abbrev=7 --dirty 2>/dev/null || true)"
     case "${tag}" in
-        audit/v*)
-            echo "${tag#audit/v}"
-            return 0
-            ;;
         v*)
             echo "${tag#v}"
             return 0
@@ -309,6 +322,12 @@ memtest_fail() {
     return 0
 }
 
+nvidia_runtime_fail() {
+    msg="$1"
+    echo "ERROR: ${msg}" >&2
+    exit 1
+}
+
 iso_memtest_present() {
     iso_path="$1"
     iso_files="$(mktemp)"
@@ -446,6 +465,44 @@ validate_iso_memtest() {
     echo "=== memtest validation OK ==="
 }
 
+validate_iso_nvidia_runtime() {
+    iso_path="$1"
+    [ "$BEE_GPU_VENDOR" = "nvidia" ] || return 0
+
+    echo "=== validating NVIDIA runtime in ISO ==="
+
+    [ -f "$iso_path" ] || nvidia_runtime_fail "ISO not found for NVIDIA runtime validation: $iso_path"
+    require_iso_reader "$iso_path" >/dev/null 2>&1 || nvidia_runtime_fail "ISO reader unavailable for NVIDIA runtime validation"
+    command -v unsquashfs >/dev/null 2>&1 || nvidia_runtime_fail "unsquashfs is required for NVIDIA runtime validation"
+
+    squashfs_tmp="$(mktemp)"
+    squashfs_list="$(mktemp)"
+    iso_read_member "$iso_path" live/filesystem.squashfs "$squashfs_tmp" || {
+        rm -f "$squashfs_tmp" "$squashfs_list"
+        nvidia_runtime_fail "failed to extract live/filesystem.squashfs from ISO"
+    }
+    unsquashfs -ll "$squashfs_tmp" > "$squashfs_list" 2>/dev/null || {
+        rm -f "$squashfs_tmp" "$squashfs_list"
+        nvidia_runtime_fail "failed to inspect filesystem.squashfs from ISO"
+    }
+
+    grep -Eq 'usr/bin/dcgmi$' "$squashfs_list" || {
+        rm -f "$squashfs_tmp" "$squashfs_list"
+        nvidia_runtime_fail "dcgmi missing from final NVIDIA ISO"
+    }
+    grep -Eq 'usr/bin/nv-hostengine$' "$squashfs_list" || {
+        rm -f "$squashfs_tmp" "$squashfs_list"
+        nvidia_runtime_fail "nv-hostengine missing from final NVIDIA ISO"
+    }
+    grep -Eq 'usr/bin/dcgmproftester([0-9]+)?$' "$squashfs_list" || {
+        rm -f "$squashfs_tmp" "$squashfs_list"
+        nvidia_runtime_fail "dcgmproftester missing from final NVIDIA ISO"
+    }
+
+    rm -f "$squashfs_tmp" "$squashfs_list"
+    echo "=== NVIDIA runtime validation OK ==="
+}
+
 append_memtest_grub_entry() {
     grub_cfg="$1"
     [ -f "$grub_cfg" ] || return 1
@@ -590,7 +647,7 @@ recover_iso_memtest() {
 
 AUDIT_VERSION_EFFECTIVE="$(resolve_audit_version)"
 ISO_VERSION_EFFECTIVE="$(resolve_iso_version)"
-ISO_BASENAME="easy-bee-${BEE_GPU_VENDOR}-v${ISO_VERSION_EFFECTIVE}-amd64"
+ISO_BASENAME="easy-bee-${BUILD_VARIANT}-v${ISO_VERSION_EFFECTIVE}-amd64"
 # Versioned output directory: dist/easy-bee-v4.1/ — all final artefacts live here.
 OUT_DIR="${DIST_DIR}/easy-bee-v${ISO_VERSION_EFFECTIVE}"
 mkdir -p "${OUT_DIR}"
@@ -764,7 +821,7 @@ if [ ! -d "/usr/src/linux-headers-${KVER}" ]; then
     apt-get install -y "linux-headers-${KVER}"
 fi
 
-echo "=== bee ISO build (variant: ${BEE_GPU_VENDOR}) ==="
+echo "=== bee ISO build (variant: ${BUILD_VARIANT}) ==="
 echo "Debian: ${DEBIAN_VERSION}, Kernel ABI: ${DEBIAN_KERNEL_ABI}, Go: ${GO_VERSION}"
 echo "Audit version: ${AUDIT_VERSION_EFFECTIVE}, ISO version: ${ISO_VERSION_EFFECTIVE}"
 echo ""
@@ -834,7 +891,7 @@ if [ "$BEE_GPU_VENDOR" = "nvidia" ]; then
     fi
 fi
 
-echo "=== preparing staged overlay (${BEE_GPU_VENDOR}) ==="
+echo "=== preparing staged overlay (${BUILD_VARIANT}) ==="
 mkdir -p "${BUILD_WORK_DIR}" "${OVERLAY_STAGE_DIR}"
 
 # Sync builder config into variant work dir, preserving lb cache.
@@ -860,6 +917,86 @@ elif [ -d "${LB_PKG_CACHE}" ] && [ "$(ls -A "${LB_PKG_CACHE}" 2>/dev/null)" ]; t
     rsync -a "${LB_PKG_CACHE}/" "${BUILD_WORK_DIR}/cache/packages.chroot/"
 fi
 
+if [ "$BEE_GPU_VENDOR" != "nvidia" ] || [ "$BEE_NVIDIA_MODULE_FLAVOR" != "proprietary" ]; then
+    cat > "${BUILD_WORK_DIR}/config/bootloaders/grub-pc/grub.cfg" <<'EOF'
+source /boot/grub/config.cfg
+
+echo ""
+echo "  ███████╗ █████╗ ███████╗██╗   ██╗      ██████╗ ███████╗███████╗"
+echo "  ██╔════╝██╔══██╗██╔════╝╚██╗ ██╔╝      ██╔══██╗██╔════╝██╔════╝"
+echo "  █████╗  ███████║███████╗ ╚████╔╝ █████╗██████╔╝█████╗  █████╗"
+echo "  ██╔══╝  ██╔══██║╚════██║  ╚██╔╝  ╚════╝██╔══██╗██╔══╝  ██╔══╝"
+echo "  ███████╗██║  ██║███████║   ██║         ██████╔╝███████╗███████╗"
+echo "  ╚══════╝╚═╝  ╚═╝╚══════╝   ╚═╝         ╚═════╝ ╚══════╝╚══════╝"
+echo "  Hardware Audit LiveCD"
+echo ""
+
+menuentry "EASY-BEE" {
+    linux   @KERNEL_LIVE@ @APPEND_LIVE@ nomodeset net.ifnames=0 biosdevname=0 mitigations=off transparent_hugepage=always numa_balancing=disable nowatchdog nosoftlockup
+    initrd  @INITRD_LIVE@
+}
+
+submenu "EASY-BEE (advanced options) -->" {
+    menuentry "EASY-BEE — KMS (no nomodeset)" {
+        linux   @KERNEL_LIVE@ @APPEND_LIVE@ net.ifnames=0 biosdevname=0 mitigations=off transparent_hugepage=always numa_balancing=disable nowatchdog nosoftlockup
+        initrd  @INITRD_LIVE@
+    }
+
+    menuentry "EASY-BEE — fail-safe" {
+        linux   @KERNEL_LIVE@ @APPEND_LIVE@ nomodeset noapic noapm nodma nomce nolapic nosmp vga=normal net.ifnames=0 biosdevname=0
+        initrd  @INITRD_LIVE@
+    }
+}
+
+if [ "${grub_platform}" = "efi" ]; then
+    menuentry "Memory Test (memtest86+)" {
+        chainloader /boot/memtest86+x64.efi
+    }
+else
+    menuentry "Memory Test (memtest86+)" {
+        linux16 /boot/memtest86+x64.bin
+    }
+fi
+
+if [ "${grub_platform}" = "efi" ]; then
+    menuentry "UEFI Firmware Settings" {
+        fwsetup
+    }
+fi
+EOF
+
+    cat > "${BUILD_WORK_DIR}/config/bootloaders/isolinux/live.cfg.in" <<'EOF'
+label live-@FLAVOUR@-normal
+    menu label ^EASY-BEE
+    menu default
+    linux @LINUX@
+    initrd @INITRD@
+    append @APPEND_LIVE@
+
+label live-@FLAVOUR@-kms
+    menu label EASY-BEE (^graphics/KMS)
+    linux @LINUX@
+    initrd @INITRD@
+    append @APPEND_LIVE@ bee.display=kms
+
+label live-@FLAVOUR@-toram
+    menu label EASY-BEE (^load to RAM)
+    linux @LINUX@
+    initrd @INITRD@
+    append @APPEND_LIVE@ toram
+
+label live-@FLAVOUR@-failsafe
+    menu label EASY-BEE (^fail-safe)
+    linux @LINUX@
+    initrd @INITRD@
+    append @APPEND_LIVE@ memtest noapic noapm nodma nomce nolapic nosmp vga=normal
+
+label memtest
+    menu label ^Memory Test (memtest86+)
+    linux /boot/memtest86+x64.bin
+EOF
+fi
+
 rsync -a "${OVERLAY_DIR}/" "${OVERLAY_STAGE_DIR}/"
 rm -f \
     "${OVERLAY_STAGE_DIR}/etc/bee-ssh-password-fallback" \
@@ -944,10 +1081,10 @@ done
 # --- NVIDIA kernel modules and userspace libs ---
 if [ "$BEE_GPU_VENDOR" = "nvidia" ]; then
     run_step "build NVIDIA ${NVIDIA_DRIVER_VERSION} modules" "40-nvidia-module" \
-        sh "${BUILDER_DIR}/build-nvidia-module.sh" "${NVIDIA_DRIVER_VERSION}" "${DIST_DIR}" "${DEBIAN_KERNEL_ABI}"
+        sh "${BUILDER_DIR}/build-nvidia-module.sh" "${NVIDIA_DRIVER_VERSION}" "${DIST_DIR}" "${DEBIAN_KERNEL_ABI}" "${BEE_NVIDIA_MODULE_FLAVOR}"
 
     KVER="${DEBIAN_KERNEL_ABI}-amd64"
-    NVIDIA_CACHE="${DIST_DIR}/nvidia-${NVIDIA_DRIVER_VERSION}-${KVER}"
+    NVIDIA_CACHE="${DIST_DIR}/nvidia-${BEE_NVIDIA_MODULE_FLAVOR}-${NVIDIA_DRIVER_VERSION}-${KVER}"
 
     # Inject .ko files into overlay at /usr/local/lib/nvidia/
     OVERLAY_KMOD_DIR="${OVERLAY_STAGE_DIR}/usr/local/lib/nvidia"
@@ -1018,13 +1155,14 @@ GIT_COMMIT="$(git -C "${REPO_ROOT}" rev-parse --short HEAD 2>/dev/null || echo u
 
 if [ "$BEE_GPU_VENDOR" = "nvidia" ]; then
     GPU_VERSION_LINE="NVIDIA_DRIVER_VERSION=${NVIDIA_DRIVER_VERSION}
+NVIDIA_KERNEL_MODULES_FLAVOR=${BEE_NVIDIA_MODULE_FLAVOR}
 NCCL_VERSION=${NCCL_VERSION}
 NCCL_CUDA_VERSION=${NCCL_CUDA_VERSION}
 CUBLAS_VERSION=${CUBLAS_VERSION}
 CUDA_USERSPACE_VERSION=${CUDA_USERSPACE_VERSION}
 NCCL_TESTS_VERSION=${NCCL_TESTS_VERSION}
 JOHN_JUMBO_COMMIT=${JOHN_JUMBO_COMMIT}"
-    GPU_BUILD_INFO="nvidia:${NVIDIA_DRIVER_VERSION}"
+    GPU_BUILD_INFO="nvidia-${BEE_NVIDIA_MODULE_FLAVOR}:${NVIDIA_DRIVER_VERSION}"
 elif [ "$BEE_GPU_VENDOR" = "amd" ]; then
     GPU_VERSION_LINE="ROCM_VERSION=${ROCM_VERSION}"
     GPU_BUILD_INFO="rocm:${ROCM_VERSION}"
@@ -1036,6 +1174,7 @@ fi
 cat > "${OVERLAY_STAGE_DIR}/etc/bee-release" <<EOF
 BEE_ISO_VERSION=${ISO_VERSION_EFFECTIVE}
 BEE_AUDIT_VERSION=${AUDIT_VERSION_EFFECTIVE}
+BEE_BUILD_VARIANT=${BUILD_VARIANT}
 BEE_GPU_VENDOR=${BEE_GPU_VENDOR}
 BUILD_DATE=${BUILD_DATE}
 GIT_COMMIT=${GIT_COMMIT}
@@ -1046,6 +1185,11 @@ EOF
 
 # Write GPU vendor marker for hooks
 echo "${BEE_GPU_VENDOR}" > "${OVERLAY_STAGE_DIR}/etc/bee-gpu-vendor"
+if [ "$BEE_GPU_VENDOR" = "nvidia" ]; then
+    echo "${BEE_NVIDIA_MODULE_FLAVOR}" > "${OVERLAY_STAGE_DIR}/etc/bee-nvidia-modules-flavor"
+else
+    rm -f "${OVERLAY_STAGE_DIR}/etc/bee-nvidia-modules-flavor"
+fi
 
 # Patch motd with build info
 BEE_BUILD_INFO="${BUILD_DATE} git:${GIT_COMMIT} debian:${DEBIAN_VERSION} ${GPU_BUILD_INFO}"
@@ -1116,10 +1260,10 @@ fi
 
 # --- build ISO using live-build ---
 echo ""
-echo "=== building ISO (live-build, variant: ${BEE_GPU_VENDOR}) ==="
+echo "=== building ISO (variant: ${BUILD_VARIANT}) ==="
 
 # Export for auto/config
-BEE_GPU_VENDOR_UPPER="$(echo "${BEE_GPU_VENDOR}" | tr 'a-z' 'A-Z')"
+BEE_GPU_VENDOR_UPPER="$(echo "${BUILD_VARIANT}" | tr 'a-z-' 'A-Z_')"
 export BEE_GPU_VENDOR_UPPER
 
 cd "${LB_DIR}"
@@ -1151,9 +1295,10 @@ if [ -f "$ISO_RAW" ]; then
         fi
     fi
     validate_iso_memtest "$ISO_RAW"
+    validate_iso_nvidia_runtime "$ISO_RAW"
     cp "$ISO_RAW" "$ISO_OUT"
     echo ""
-    echo "=== done (${BEE_GPU_VENDOR}) ==="
+    echo "=== done (${BUILD_VARIANT}) ==="
     echo "ISO: $ISO_OUT"
     if command -v stat >/dev/null 2>&1; then
         ISO_SIZE_BYTES="$(stat -c '%s' "$ISO_OUT" 2>/dev/null || stat -f '%z' "$ISO_OUT")"

iso/builder/config/bootloaders/grub-pc/grub.cfg

@@ -7,6 +7,7 @@ echo "  █████╗  ███████║███████╗ ╚
 echo "  ██╔══╝  ██╔══██║╚════██║  ╚██╔╝  ╚════╝██╔══██╗██╔══╝  ██╔══╝"
 echo "  ███████╗██║  ██║███████║   ██║         ██████╔╝███████╗███████╗"
 echo "  ╚══════╝╚═╝  ╚═╝╚══════╝   ╚═╝         ╚═════╝ ╚══════╝╚══════╝"
+echo "  Hardware Audit LiveCD"
 echo ""
 
 menuentry "EASY-BEE" {
@@ -14,29 +15,21 @@ menuentry "EASY-BEE" {
     initrd  @INITRD_LIVE@
 }
 
-menuentry "EASY-BEE (graphics/KMS)" {
-    linux   @KERNEL_LIVE@ @APPEND_LIVE@ bee.display=kms bee.nvidia.mode=normal net.ifnames=0 biosdevname=0 mitigations=off transparent_hugepage=always numa_balancing=disable nowatchdog nosoftlockup
-    initrd  @INITRD_LIVE@
-}
+submenu "EASY-BEE (advanced options) -->" {
+    menuentry "EASY-BEE — GSP=off" {
+        linux   @KERNEL_LIVE@ @APPEND_LIVE@ nomodeset bee.nvidia.mode=gsp-off net.ifnames=0 biosdevname=0 mitigations=off transparent_hugepage=always numa_balancing=disable nowatchdog nosoftlockup
+        initrd  @INITRD_LIVE@
+    }
 
-menuentry "EASY-BEE (load to RAM)" {
-    linux   @KERNEL_LIVE@ @APPEND_LIVE@ toram nomodeset bee.nvidia.mode=normal net.ifnames=0 biosdevname=0 mitigations=off transparent_hugepage=always numa_balancing=disable nowatchdog nosoftlockup
-    initrd  @INITRD_LIVE@
-}
+    menuentry "EASY-BEE — KMS (no nomodeset)" {
+        linux   @KERNEL_LIVE@ @APPEND_LIVE@ bee.nvidia.mode=normal net.ifnames=0 biosdevname=0 mitigations=off transparent_hugepage=always numa_balancing=disable nowatchdog nosoftlockup
+        initrd  @INITRD_LIVE@
+    }
 
-menuentry "EASY-BEE (NVIDIA GSP=off)" {
-    linux   @KERNEL_LIVE@ @APPEND_LIVE@ nomodeset bee.nvidia.mode=gsp-off net.ifnames=0 biosdevname=0 mitigations=off transparent_hugepage=always numa_balancing=disable nowatchdog nosoftlockup
-    initrd  @INITRD_LIVE@
-}
-
-menuentry "EASY-BEE (graphics/KMS, GSP=off)" {
-    linux   @KERNEL_LIVE@ @APPEND_LIVE@ bee.display=kms bee.nvidia.mode=gsp-off net.ifnames=0 biosdevname=0 mitigations=off transparent_hugepage=always numa_balancing=disable nowatchdog nosoftlockup
-    initrd  @INITRD_LIVE@
-}
-
-menuentry "EASY-BEE (fail-safe)" {
-    linux   @KERNEL_LIVE@ @APPEND_LIVE@ bee.nvidia.mode=gsp-off memtest noapic noapm nodma nomce nolapic nosmp vga=normal net.ifnames=0 biosdevname=0
-    initrd  @INITRD_LIVE@
+    menuentry "EASY-BEE — fail-safe" {
+        linux   @KERNEL_LIVE@ @APPEND_LIVE@ nomodeset bee.nvidia.mode=gsp-off noapic noapm nodma nomce nolapic nosmp vga=normal net.ifnames=0 biosdevname=0
+        initrd  @INITRD_LIVE@
+    }
 }
 
 if [ "${grub_platform}" = "efi" ]; then

iso/builder/config/bootloaders/grub-pc/theme.cfg

@@ -1,9 +1,9 @@
 set color_normal=light-gray/black
-set color_highlight=white/dark-gray
+set color_highlight=yellow/black
 
 if [ -e /boot/grub/splash.png ]; then
     set theme=/boot/grub/live-theme/theme.txt
 else
-    set menu_color_normal=cyan/black
-    set menu_color_highlight=white/dark-gray
+    set menu_color_normal=yellow/black
+    set menu_color_highlight=white/brown
 fi

iso/builder/config/hooks/normal/9000-bee-setup.hook.chroot

@@ -30,6 +30,8 @@ systemctl enable bee-preflight.service
 systemctl enable bee-audit.service
 systemctl enable bee-web.service
 systemctl enable bee-sshsetup.service
+systemctl enable bee-selfheal.timer
+systemctl enable bee-boot-status.service
 systemctl enable ssh.service
 systemctl enable lightdm.service 2>/dev/null || true
 systemctl enable qemu-guest-agent.service 2>/dev/null || true
@@ -58,6 +60,8 @@ chmod +x /usr/local/bin/bee-sshsetup   2>/dev/null || true
 chmod +x /usr/local/bin/bee-smoketest  2>/dev/null || true
 chmod +x /usr/local/bin/bee            2>/dev/null || true
 chmod +x /usr/local/bin/bee-log-run    2>/dev/null || true
+chmod +x /usr/local/bin/bee-selfheal      2>/dev/null || true
+chmod +x /usr/local/bin/bee-boot-status  2>/dev/null || true
 if [ "$GPU_VENDOR" = "nvidia" ]; then
     chmod +x /usr/local/bin/bee-nvidia-load 2>/dev/null || true
     chmod +x /usr/local/bin/bee-gpu-burn 2>/dev/null || true

iso/builder/config/hooks/normal/9001-wallpaper.hook.chroot

@@ -0,0 +1,117 @@
+#!/bin/sh
+# 9001-wallpaper.hook.chroot — generate /usr/share/bee/wallpaper.png inside chroot
+set -e
+echo "=== generating bee wallpaper ==="
+mkdir -p /usr/share/bee
+
+python3 - <<'PYEOF'
+from PIL import Image, ImageDraw, ImageFont, ImageFilter
+import os
+
+W, H = 1920, 1080
+
+ASCII_ART = [
+    "  ███████╗ █████╗ ███████╗██╗   ██╗      ██████╗ ███████╗███████╗",
+    "  ██╔════╝██╔══██╗██╔════╝╚██╗ ██╔╝      ██╔══██╗██╔════╝██╔════╝",
+    "  █████╗  ███████║███████╗ ╚████╔╝ █████╗██████╔╝█████╗  █████╗",
+    "  ██╔══╝  ██╔══██║╚════██║  ╚██╔╝  ╚════╝██╔══██╗██╔══╝  ██╔══╝",
+    "  ███████╗██║  ██║███████║   ██║         ██████╔╝███████╗███████╗",
+    "  ╚══════╝╚═╝  ╚═╝╚══════╝   ╚═╝         ╚═════╝ ╚══════╝╚══════╝",
+]
+SUBTITLE = "  Hardware Audit LiveCD"
+
+FG = (0xF6, 0xD0, 0x47)
+FG_DIM = (0xD4, 0xA9, 0x1C)
+SHADOW = (0x5E, 0x47, 0x05)
+SUB = (0x96, 0x7A, 0x17)
+BG = (0x05, 0x05, 0x05)
+
+MONO_FONT_CANDIDATES = [
+    '/usr/share/fonts/truetype/dejavu/DejaVuSansMono-Bold.ttf',
+    '/usr/share/fonts/truetype/liberation2/LiberationMono-Bold.ttf',
+    '/usr/share/fonts/truetype/liberation/LiberationMono-Bold.ttf',
+    '/usr/share/fonts/truetype/freefont/FreeMonoBold.ttf',
+]
+SUB_FONT_CANDIDATES = [
+    '/usr/share/fonts/truetype/dejavu/DejaVuSans-Bold.ttf',
+    '/usr/share/fonts/truetype/liberation2/LiberationSans-Bold.ttf',
+    '/usr/share/fonts/truetype/liberation/LiberationSans-Bold.ttf',
+    '/usr/share/fonts/truetype/freefont/FreeSansBold.ttf',
+]
+
+
+def load_font(candidates, size):
+    for path in candidates:
+        if os.path.exists(path):
+            return ImageFont.truetype(path, size)
+    return ImageFont.load_default()
+
+
+def mono_metrics(font):
+    probe = Image.new('L', (W, H), 0)
+    draw = ImageDraw.Draw(probe)
+    char_w = int(round(draw.textlength("M", font=font)))
+    bb = draw.textbbox((0, 0), "Mg", font=font)
+    char_h = bb[3] - bb[1]
+    return char_w, char_h
+
+
+def render_ascii_mask(font, lines, char_w, char_h, line_gap):
+    width = max(len(line) for line in lines) * char_w
+    height = len(lines) * char_h + line_gap * (len(lines) - 1)
+    mask = Image.new('L', (width, height), 0)
+    draw = ImageDraw.Draw(mask)
+    for row, line in enumerate(lines):
+        y = row * (char_h + line_gap)
+        for col, ch in enumerate(line):
+            if ch == ' ':
+                continue
+            x = col * char_w
+            draw.text((x, y), ch, font=font, fill=255)
+    return mask
+
+
+img = Image.new('RGB', (W, H), BG)
+draw = ImageDraw.Draw(img)
+
+# Soft amber glow under the logo without depending on font rendering.
+glow = Image.new('RGBA', (W, H), (0, 0, 0, 0))
+glow_draw = ImageDraw.Draw(glow)
+glow_draw.ellipse((360, 250, 1560, 840), fill=(180, 120, 10, 56))
+glow_draw.ellipse((520, 340, 1400, 760), fill=(255, 190, 40, 36))
+glow = glow.filter(ImageFilter.GaussianBlur(60))
+img = Image.alpha_composite(img.convert('RGBA'), glow)
+
+TARGET_LOGO_W = 400
+max_chars = max(len(line) for line in ASCII_ART)
+_probe_font = load_font(MONO_FONT_CANDIDATES, 64)
+_probe_cw, _ = mono_metrics(_probe_font)
+font_size_logo = max(6, int(64 * TARGET_LOGO_W / (_probe_cw * max_chars)))
+font_logo = load_font(MONO_FONT_CANDIDATES, font_size_logo)
+char_w, char_h = mono_metrics(font_logo)
+logo_mask = render_ascii_mask(font_logo, ASCII_ART, char_w, char_h, 2)
+logo_w, logo_h = logo_mask.size
+logo_x = (W - logo_w) // 2
+logo_y = 380
+
+sh_off = max(1, font_size_logo // 6)
+shadow_mask = logo_mask.filter(ImageFilter.GaussianBlur(1))
+img.paste(SHADOW, (logo_x + sh_off * 2, logo_y + sh_off * 2), shadow_mask)
+img.paste(FG_DIM, (logo_x + sh_off, logo_y + sh_off), logo_mask)
+img.paste(FG, (logo_x, logo_y), logo_mask)
+
+font_sub = load_font(SUB_FONT_CANDIDATES, 30)
+sub_bb = draw.textbbox((0, 0), SUBTITLE, font=font_sub)
+sub_x = (W - (sub_bb[2] - sub_bb[0])) // 2
+sub_y = logo_y + logo_h + 48
+draw = ImageDraw.Draw(img)
+draw.text((sub_x + 2, sub_y + 2), SUBTITLE, font=font_sub, fill=(35, 28, 6))
+draw.text((sub_x, sub_y), SUBTITLE, font=font_sub, fill=SUB)
+
+img = img.convert('RGB')
+
+img.save('/usr/share/bee/wallpaper.png', optimize=True)
+print('wallpaper written: /usr/share/bee/wallpaper.png')
+PYEOF
+
+echo "=== wallpaper done ==="

iso/builder/config/hooks/normal/9010-fix-toram.hook.chroot

@@ -0,0 +1,41 @@
+#!/bin/sh
+# 9010-fix-toram.hook.chroot — patch live-boot toram to work with tmpfs (no O_DIRECT)
+#
+# live-boot tries "losetup --replace --direct-io=on" when re-associating the
+# loop device to the RAM copy in /dev/shm.  tmpfs does not support O_DIRECT,
+# so the ioctl returns EINVAL and the verification step fails.
+#
+# The patch replaces the replace call so that if --direct-io=on fails it falls
+# back to a plain replace without direct-io, and also relaxes the verification
+# to a warning so the boot continues even when re-association is imperfect.
+set -e
+
+TORAM_SCRIPT="/usr/lib/live/boot/9990-toram-todisk.sh"
+
+if [ ! -f "${TORAM_SCRIPT}" ]; then
+    echo "9010-fix-toram: ${TORAM_SCRIPT} not found, skipping"
+    exit 0
+fi
+
+echo "9010-fix-toram: patching ${TORAM_SCRIPT}"
+
+# Replace any losetup --replace call that includes --direct-io=on with a
+# version that first tries with direct-io, then retries without it.
+#
+# The sed expression turns:
+#   losetup --replace ... --direct-io=on LOOP FILE
+# into a shell snippet that tries both, silently.
+#
+# We also downgrade the fatal "Task finished with error." block to a warning
+# so the boot continues if re-association fails (squashfs still accessible).
+
+# 1. Strip --direct-io=on from the losetup --replace call so it works on tmpfs.
+sed -i 's/losetup --replace --direct-io=on/losetup --replace/g' "${TORAM_SCRIPT}"
+sed -i 's/losetup --replace --direct-io/losetup --replace/g' "${TORAM_SCRIPT}"
+
+# 2. Turn the hard error into a warning so boot continues.
+#    live-boot prints this exact string when verification fails.
+sed -i 's/echo "Task finished with error\."/echo "Warning: toram re-association failed, continuing boot (squashfs still in RAM)"/' "${TORAM_SCRIPT}"
+
+echo "9010-fix-toram: patch applied"
+grep -n "losetup" "${TORAM_SCRIPT}" | head -20 || true

iso/builder/config/package-lists/bee-nvidia.list.chroot

@@ -1,6 +1,10 @@
-# NVIDIA DCGM (Data Center GPU Manager) — dcgmi diag for acceptance testing.
-# DCGM 4 is packaged per CUDA major. The image ships NVIDIA driver 590 with CUDA 13 userspace,
-# so install the CUDA 13 build plus proprietary diagnostic components explicitly.
+# NVIDIA DCGM (Data Center GPU Manager).
+# Validate uses dcgmi diagnostics; Burn uses dcgmproftester as the official
+# NVIDIA max-compute recipe. The smoketest/runtime contract treats
+# dcgmproftester as required in the LiveCD.
+# DCGM 4 is packaged per CUDA major. The image ships NVIDIA driver 590 with
+# CUDA 13 userspace, so install the CUDA 13 build plus proprietary components
+# explicitly.
 datacenter-gpu-manager-4-cuda13=1:%%DCGM_VERSION%%
 datacenter-gpu-manager-4-proprietary=1:%%DCGM_VERSION%%
 datacenter-gpu-manager-4-proprietary-cuda13=1:%%DCGM_VERSION%%

iso/builder/config/package-lists/bee.list.chroot

@@ -60,9 +60,15 @@ qrencode
 # Local desktop (openbox + chromium kiosk)
 openbox
 tint2
+feh
+python3-pil
 xorg
 xterm
 chromium
+mousepad
+pcmanfm
+ristretto
+mupdf
 xserver-xorg-video-fbdev
 xserver-xorg-video-vesa
 lightdm

iso/builder/smoketest.sh

@@ -27,6 +27,7 @@ echo ""
 KVER=$(uname -r)
 info "kernel: $KVER"
 NVIDIA_BOOT_MODE="normal"
+NVIDIA_MODULES_FLAVOR="proprietary"
 for arg in $(cat /proc/cmdline 2>/dev/null); do
     case "$arg" in
         bee.nvidia.mode=*)
@@ -34,7 +35,11 @@ for arg in $(cat /proc/cmdline 2>/dev/null); do
             ;;
     esac
 done
+if [ -f /etc/bee-nvidia-modules-flavor ]; then
+    NVIDIA_MODULES_FLAVOR="$(tr -d '[:space:]' </etc/bee-nvidia-modules-flavor 2>/dev/null || echo proprietary)"
+fi
 info "nvidia boot mode: ${NVIDIA_BOOT_MODE}"
+info "nvidia modules flavor: ${NVIDIA_MODULES_FLAVOR}"
 
 # --- PATH & binaries ---
 echo "-- PATH & binaries --"
@@ -52,6 +57,31 @@ else
     fail "nvidia-smi: NOT FOUND"
 fi
 
+if p=$(PATH="/usr/local/bin:$PATH" command -v dcgmi 2>/dev/null); then
+    ok "dcgmi found: $p"
+else
+    fail "dcgmi: NOT FOUND"
+fi
+
+if p=$(PATH="/usr/local/bin:$PATH" command -v nv-hostengine 2>/dev/null); then
+    ok "nv-hostengine found: $p"
+else
+    fail "nv-hostengine: NOT FOUND"
+fi
+
+DCGM_PROFTESTER=""
+for tool in dcgmproftester dcgmproftester13 dcgmproftester12 dcgmproftester11; do
+    if p=$(PATH="/usr/local/bin:$PATH" command -v "$tool" 2>/dev/null); then
+        DCGM_PROFTESTER="$p"
+        break
+    fi
+done
+if [ -n "$DCGM_PROFTESTER" ]; then
+    ok "dcgmproftester found: $DCGM_PROFTESTER"
+else
+    fail "dcgmproftester: NOT FOUND"
+fi
+
 for tool in bee-gpu-burn bee-john-gpu-stress bee-nccl-gpu-stress all_reduce_perf; do
     if p=$(PATH="/usr/local/bin:$PATH" command -v "$tool" 2>/dev/null); then
         ok "$tool found: $p"
@@ -60,6 +90,12 @@ for tool in bee-gpu-burn bee-john-gpu-stress bee-nccl-gpu-stress all_reduce_perf
     fi
 done
 
+if p=$(PATH="/usr/local/bin:$PATH" command -v nvbandwidth 2>/dev/null); then
+    ok "nvbandwidth found: $p"
+else
+    warn "nvbandwidth: NOT FOUND"
+fi
+
 echo ""
 echo "-- NVIDIA modules --"
 KO_DIR="/usr/local/lib/nvidia"
@@ -79,10 +115,12 @@ fi
 for mod in nvidia_modeset nvidia_uvm; do
     if /sbin/lsmod 2>/dev/null | grep -q "^$mod "; then
         ok "module loaded: $mod"
-    elif [ "${NVIDIA_BOOT_MODE}" = "normal" ] || [ "${NVIDIA_BOOT_MODE}" = "full" ]; then
+    elif [ "${NVIDIA_MODULES_FLAVOR}" = "proprietary" ] && { [ "${NVIDIA_BOOT_MODE}" = "normal" ] || [ "${NVIDIA_BOOT_MODE}" = "full" ]; }; then
         fail "module NOT loaded in normal mode: $mod"
-    else
+    elif [ "${NVIDIA_MODULES_FLAVOR}" = "proprietary" ]; then
         warn "module not loaded in GSP-off mode: $mod"
+    else
+        fail "module NOT loaded: $mod"
     fi
 done
 
@@ -98,10 +136,12 @@ done
 
 if [ -e /dev/nvidia-uvm ]; then
     ok "/dev/nvidia-uvm exists"
-elif [ "${NVIDIA_BOOT_MODE}" = "normal" ] || [ "${NVIDIA_BOOT_MODE}" = "full" ]; then
+elif [ "${NVIDIA_MODULES_FLAVOR}" = "proprietary" ] && { [ "${NVIDIA_BOOT_MODE}" = "normal" ] || [ "${NVIDIA_BOOT_MODE}" = "full" ]; }; then
     fail "/dev/nvidia-uvm missing in normal mode"
-else
+elif [ "${NVIDIA_MODULES_FLAVOR}" = "proprietary" ]; then
     warn "/dev/nvidia-uvm missing — CUDA stress path may be unavailable until loaded on demand"
+else
+    fail "/dev/nvidia-uvm missing"
 fi
 
 echo ""
@@ -171,6 +211,12 @@ for svc in bee-nvidia bee-network bee-preflight bee-audit bee-web; do
     fi
 done
 
+if systemctl is-active --quiet bee-selfheal.timer 2>/dev/null; then
+    ok "timer active: bee-selfheal.timer"
+else
+    fail "timer NOT active: bee-selfheal.timer"
+fi
+
 echo ""
 echo "-- runtime health --"
 if [ -f /appdata/bee/export/runtime-health.json ] && [ -s /appdata/bee/export/runtime-health.json ]; then

iso/overlay/etc/systemd/system/bee-audit.service

@@ -1,7 +1,6 @@
 [Unit]
 Description=Bee: hardware audit
 After=bee-preflight.service bee-network.service bee-nvidia.service
-Before=bee-web.service
 
 [Service]
 Type=oneshot

iso/overlay/etc/systemd/system/bee-boot-status.service

@@ -0,0 +1,18 @@
+[Unit]
+Description=Bee: boot status display
+After=systemd-user-sessions.service
+Before=getty@tty1.service
+
+[Service]
+Type=oneshot
+RemainAfterExit=no
+ExecStart=/usr/local/bin/bee-boot-status
+TTYPath=/dev/tty1
+StandardInput=tty
+StandardOutput=tty
+StandardError=tty
+TTYReset=yes
+TTYVHangup=yes
+
+[Install]
+WantedBy=multi-user.target

iso/overlay/etc/systemd/system/bee-selfheal.service

@@ -0,0 +1,9 @@
+[Unit]
+Description=Bee: periodic runtime self-heal
+After=bee-web.service bee-audit.service bee-preflight.service
+
+[Service]
+Type=oneshot
+ExecStart=/usr/local/bin/bee-log-run /appdata/bee/export/bee-selfheal.log /usr/local/bin/bee-selfheal
+StandardOutput=journal
+StandardError=journal

iso/overlay/etc/systemd/system/bee-selfheal.timer

@@ -0,0 +1,11 @@
+[Unit]
+Description=Bee: run self-heal checks periodically
+
+[Timer]
+OnBootSec=45sec
+OnUnitActiveSec=60sec
+AccuracySec=15sec
+Unit=bee-selfheal.service
+
+[Install]
+WantedBy=timers.target

iso/overlay/etc/systemd/system/bee-web.service

@@ -1,12 +1,12 @@
 [Unit]
 Description=Bee: hardware audit web viewer
-After=bee-audit.service
+StartLimitIntervalSec=0
 
 [Service]
 Type=simple
 ExecStart=/usr/local/bin/bee-log-run /appdata/bee/export/bee-web.log /usr/local/bin/bee web --listen :80 --audit-path /appdata/bee/export/bee-audit.json --export-dir /appdata/bee/export --title "Bee Hardware Audit"
 Restart=always
-RestartSec=2
+RestartSec=3
 StandardOutput=journal
 StandardError=journal
 LimitMEMLOCK=infinity

iso/overlay/etc/systemd/system/getty@tty1.service.d/wait-bee.conf

@@ -0,0 +1,2 @@
+[Unit]
+After=bee-boot-status.service

iso/overlay/etc/systemd/system/lightdm.service.d/bee-display-mode.conf

@@ -1,6 +1,4 @@
 [Unit]
-Wants=bee-preflight.service
-After=bee-preflight.service
 
 [Service]
 ExecStartPre=/usr/local/bin/bee-display-mode

iso/overlay/usr/local/bin/bee-boot-status

@@ -0,0 +1,89 @@
+#!/bin/sh
+# bee-boot-status — boot progress display on tty1.
+# Shows live service status until all bee services are done or failed,
+# then exits so getty can show the login prompt.
+
+CRITICAL="bee-preflight bee-nvidia bee-audit"
+ALL="bee-sshsetup ssh bee-network bee-nvidia bee-preflight bee-audit bee-web"
+
+svc_state() { systemctl is-active "$1.service" 2>/dev/null || echo "inactive"; }
+
+svc_icon() {
+    case "$(svc_state "$1")" in
+        active)       printf '\033[32m[  OK  ]\033[0m' ;;
+        failed)       printf '\033[31m[ FAIL ]\033[0m' ;;
+        activating)   printf '\033[33m[  ..  ]\033[0m' ;;
+        deactivating) printf '\033[33m[ stop ]\033[0m' ;;
+        inactive)     printf '\033[90m[      ]\033[0m' ;;
+        *)            printf '\033[90m[  ?   ]\033[0m' ;;
+    esac
+}
+
+svc_detail() {
+    local svc="$1" state
+    state="$(svc_state "$svc")"
+    case "$state" in
+        failed)
+            local res
+            res="$(systemctl show -p Result "$svc.service" 2>/dev/null | cut -d= -f2)"
+            [ -n "$res" ] && [ "$res" != "success" ] && printf '  \033[31m(%s)\033[0m' "$res"
+            ;;
+        activating)
+            local line
+            line="$(journalctl -u "$svc.service" -n 1 --no-pager --output=cat 2>/dev/null | cut -c1-55)"
+            [ -n "$line" ] && printf '  \033[90m%s\033[0m' "$line"
+            ;;
+    esac
+}
+
+all_critical_done() {
+    for svc in $CRITICAL; do
+        case "$(svc_state "$svc")" in
+            active|failed|inactive) ;;
+            *) return 1 ;;
+        esac
+    done
+    return 0
+}
+
+while true; do
+    # move to top-left and clear screen
+    printf '\033[H\033[2J'
+
+    printf '\n'
+    printf '  \033[33m███████╗ █████╗ ███████╗██╗   ██╗      ██████╗ ███████╗███████╗\033[0m\n'
+    printf '  \033[33m██╔════╝██╔══██╗██╔════╝╚██╗ ██╔╝      ██╔══██╗██╔════╝██╔════╝\033[0m\n'
+    printf '  \033[33m█████╗  ███████║███████╗ ╚████╔╝ █████╗██████╔╝█████╗  █████╗\033[0m\n'
+    printf '  \033[33m██╔══╝  ██╔══██║╚════██║  ╚██╔╝  ╚════╝██╔══██╗██╔══╝  ██╔══╝\033[0m\n'
+    printf '  \033[33m███████╗██║  ██║███████║   ██║         ██████╔╝███████╗███████╗\033[0m\n'
+    printf '  \033[33m╚══════╝╚═╝  ╚═╝╚══════╝   ╚═╝         ╚═════╝ ╚══════╝╚══════╝\033[0m\n'
+    printf '  Hardware Audit LiveCD\n'
+    printf '\n'
+
+    for svc in $ALL; do
+        printf '  %s  %-20s%s\n' "$(svc_icon "$svc")" "$svc" "$(svc_detail "$svc")"
+    done
+    printf '\n'
+
+    # Network
+    ips="$(ip -4 addr show scope global 2>/dev/null | awk '/inet /{printf "  %-16s %s\n", $NF, $2}')"
+    if [ -n "$ips" ]; then
+        printf '  \033[1mNetwork:\033[0m\n'
+        printf '%s\n' "$ips"
+        printf '\n'
+    fi
+
+    if all_critical_done; then
+        printf '  \033[1;32mSystem ready.\033[0m  Audit is running in the background.\n'
+        first_ip="$(ip -4 addr show scope global 2>/dev/null | awk '/inet /{print $2}' | cut -d/ -f1 | head -1)"
+        if [ -n "$first_ip" ]; then
+            printf '  Web UI: \033[1mhttp://%s/\033[0m\n' "$first_ip"
+        fi
+        printf '\n'
+        sleep 3
+        break
+    fi
+
+    printf '  \033[90mStarting up...\033[0m\n'
+    sleep 3
+done

iso/overlay/usr/local/bin/bee-dcgmproftester-staggered

@@ -0,0 +1,110 @@
+#!/bin/sh
+set -eu
+
+SECONDS=300
+STAGGER_SECONDS=180
+DEVICES=""
+EXCLUDE=""
+
+usage() {
+    echo "usage: $0 [--seconds N] [--stagger-seconds N] [--devices 0,1] [--exclude 2,3]" >&2
+    exit 2
+}
+
+normalize_list() {
+    echo "${1:-}" | tr ',' '\n' | sed 's/[[:space:]]//g' | awk 'NF' | sort -n | uniq | paste -sd, -
+}
+
+contains_csv() {
+    needle="$1"
+    haystack="${2:-}"
+    echo ",${haystack}," | grep -q ",${needle},"
+}
+
+resolve_dcgmproftester() {
+    for candidate in dcgmproftester dcgmproftester13 dcgmproftester12 dcgmproftester11; do
+        if command -v "${candidate}" >/dev/null 2>&1; then
+            command -v "${candidate}"
+            return 0
+        fi
+    done
+    return 1
+}
+
+while [ "$#" -gt 0 ]; do
+    case "$1" in
+        --seconds|-t) [ "$#" -ge 2 ] || usage; SECONDS="$2"; shift 2 ;;
+        --stagger-seconds) [ "$#" -ge 2 ] || usage; STAGGER_SECONDS="$2"; shift 2 ;;
+        --devices) [ "$#" -ge 2 ] || usage; DEVICES="$2"; shift 2 ;;
+        --exclude) [ "$#" -ge 2 ] || usage; EXCLUDE="$2"; shift 2 ;;
+        *) usage ;;
+    esac
+done
+
+PROF=$(resolve_dcgmproftester) || { echo "dcgmproftester not found in PATH" >&2; exit 1; }
+ALL_DEVICES=$(nvidia-smi --query-gpu=index --format=csv,noheader,nounits 2>/dev/null | sed 's/[[:space:]]//g' | awk 'NF' | paste -sd, -)
+[ -n "${ALL_DEVICES}" ] || { echo "nvidia-smi found no NVIDIA GPUs" >&2; exit 1; }
+
+DEVICES=$(normalize_list "${DEVICES}")
+EXCLUDE=$(normalize_list "${EXCLUDE}")
+SELECTED="${DEVICES}"
+if [ -z "${SELECTED}" ]; then
+    SELECTED="${ALL_DEVICES}"
+fi
+
+FINAL=""
+for id in $(echo "${SELECTED}" | tr ',' ' '); do
+    [ -n "${id}" ] || continue
+    if contains_csv "${id}" "${EXCLUDE}"; then
+        continue
+    fi
+    if [ -z "${FINAL}" ]; then
+        FINAL="${id}"
+    else
+        FINAL="${FINAL},${id}"
+    fi
+done
+
+[ -n "${FINAL}" ] || { echo "no NVIDIA GPUs selected after filters" >&2; exit 1; }
+
+echo "loader=dcgmproftester-staggered"
+echo "selected_gpus=${FINAL}"
+echo "stagger_seconds=${STAGGER_SECONDS}"
+
+TMP_DIR=$(mktemp -d)
+trap 'rm -rf "${TMP_DIR}"' EXIT INT TERM
+
+GPU_COUNT=$(echo "${FINAL}" | tr ',' '\n' | awk 'NF' | wc -l | tr -d '[:space:]')
+gpu_pos=0
+WORKERS=""
+for id in $(echo "${FINAL}" | tr ',' ' '); do
+    gpu_pos=$((gpu_pos + 1))
+    log="${TMP_DIR}/gpu-${id}.log"
+    extra_sec=$(( STAGGER_SECONDS * (GPU_COUNT - gpu_pos) ))
+    gpu_seconds=$(( SECONDS + extra_sec ))
+    echo "starting gpu ${id} seconds=${gpu_seconds}"
+    CUDA_VISIBLE_DEVICES="${id}" "${PROF}" --no-dcgm-validation -t 1004 -d "${gpu_seconds}" >"${log}" 2>&1 &
+    pid=$!
+    WORKERS="${WORKERS} ${pid}:${id}:${log}"
+    if [ "${STAGGER_SECONDS}" -gt 0 ] && [ "${gpu_pos}" -lt "${GPU_COUNT}" ]; then
+        sleep "${STAGGER_SECONDS}"
+    fi
+done
+
+status=0
+for spec in ${WORKERS}; do
+    pid=${spec%%:*}
+    rest=${spec#*:}
+    id=${rest%%:*}
+    log=${rest#*:}
+    if wait "${pid}"; then
+        echo "gpu ${id} finished: OK"
+    else
+        rc=$?
+        echo "gpu ${id} finished: FAILED rc=${rc}"
+        status=1
+    fi
+    sed "s/^/[gpu ${id}] /" "${log}" || true
+done
+
+exit "${status}"

iso/overlay/usr/local/bin/bee-gpu-burn

@@ -2,13 +2,14 @@
 set -eu
 
 SECONDS=5
+STAGGER_SECONDS=0
 SIZE_MB=0
 DEVICES=""
 EXCLUDE=""
 WORKER="/usr/local/lib/bee/bee-gpu-burn-worker"
 
 usage() {
-    echo "usage: $0 [--seconds N] [--size-mb N] [--devices 0,1] [--exclude 2,3]" >&2
+    echo "usage: $0 [--seconds N] [--stagger-seconds N] [--size-mb N] [--devices 0,1] [--exclude 2,3]" >&2
     exit 2
 }
 
@@ -25,6 +26,7 @@ contains_csv() {
 while [ "$#" -gt 0 ]; do
     case "$1" in
         --seconds|-t) [ "$#" -ge 2 ] || usage; SECONDS="$2"; shift 2 ;;
+        --stagger-seconds) [ "$#" -ge 2 ] || usage; STAGGER_SECONDS="$2"; shift 2 ;;
         --size-mb|-m) [ "$#" -ge 2 ] || usage; SIZE_MB="$2"; shift 2 ;;
         --devices) [ "$#" -ge 2 ] || usage; DEVICES="$2"; shift 2 ;;
         --exclude) [ "$#" -ge 2 ] || usage; EXCLUDE="$2"; shift 2 ;;
@@ -61,12 +63,18 @@ done
 
 echo "loader=bee-gpu-burn"
 echo "selected_gpus=${FINAL}"
+echo "stagger_seconds=${STAGGER_SECONDS}"
+
+export CUDA_DEVICE_ORDER="PCI_BUS_ID"
 
 TMP_DIR=$(mktemp -d)
 trap 'rm -rf "${TMP_DIR}"' EXIT INT TERM
 
+GPU_COUNT=$(echo "${FINAL}" | tr ',' '\n' | awk 'NF' | wc -l | tr -d '[:space:]')
+gpu_pos=0
 WORKERS=""
 for id in $(echo "${FINAL}" | tr ',' ' '); do
+    gpu_pos=$((gpu_pos + 1))
     log="${TMP_DIR}/gpu-${id}.log"
     gpu_size_mb="${SIZE_MB}"
     if [ "${gpu_size_mb}" -le 0 ] 2>/dev/null; then
@@ -77,10 +85,16 @@ for id in $(echo "${FINAL}" | tr ',' ' '); do
             gpu_size_mb=512
         fi
     fi
-    echo "starting gpu ${id} size=${gpu_size_mb}MB"
-    "${WORKER}" --device "${id}" --seconds "${SECONDS}" --size-mb "${gpu_size_mb}" >"${log}" 2>&1 &
+    extra_sec=$(( STAGGER_SECONDS * (GPU_COUNT - gpu_pos) ))
+    gpu_seconds=$(( SECONDS + extra_sec ))
+    echo "starting gpu ${id} size=${gpu_size_mb}MB seconds=${gpu_seconds}"
+    CUDA_VISIBLE_DEVICES="${id}" \
+        "${WORKER}" --device 0 --seconds "${gpu_seconds}" --size-mb "${gpu_size_mb}" >"${log}" 2>&1 &
     pid=$!
     WORKERS="${WORKERS} ${pid}:${id}:${log}"
+    if [ "${STAGGER_SECONDS}" -gt 0 ] && [ "${gpu_pos}" -lt "${GPU_COUNT}" ]; then
+        sleep "${STAGGER_SECONDS}"
+    fi
 done
 
 status=0

iso/overlay/usr/local/bin/bee-john-gpu-stress

@@ -1,17 +1,19 @@
 #!/bin/sh
 set -eu
 
-SECONDS=300
+DURATION_SEC=300
+STAGGER_SECONDS=0
 DEVICES=""
 EXCLUDE=""
 FORMAT=""
+TEST_SLICE_SECONDS=300
 JOHN_DIR="/usr/local/lib/bee/john/run"
 JOHN_BIN="${JOHN_DIR}/john"
 export OCL_ICD_VENDORS="/etc/OpenCL/vendors"
 export LD_LIBRARY_PATH="/usr/lib:/usr/local/lib${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}"
 
 usage() {
-    echo "usage: $0 [--seconds N] [--devices 0,1] [--exclude 2,3] [--format name]" >&2
+    echo "usage: $0 [--seconds N] [--stagger-seconds N] [--devices 0,1] [--exclude 2,3] [--format name]" >&2
     exit 2
 }
 
@@ -116,7 +118,8 @@ ensure_opencl_ready() {
 
 while [ "$#" -gt 0 ]; do
     case "$1" in
-        --seconds|-t) [ "$#" -ge 2 ] || usage; SECONDS="$2"; shift 2 ;;
+        --seconds|-t) [ "$#" -ge 2 ] || usage; DURATION_SEC="$2"; shift 2 ;;
+        --stagger-seconds) [ "$#" -ge 2 ] || usage; STAGGER_SECONDS="$2"; shift 2 ;;
         --devices) [ "$#" -ge 2 ] || usage; DEVICES="$2"; shift 2 ;;
         --exclude) [ "$#" -ge 2 ] || usage; EXCLUDE="$2"; shift 2 ;;
         --format) [ "$#" -ge 2 ] || usage; FORMAT="$2"; shift 2 ;;
@@ -151,19 +154,25 @@ done
 
 [ -n "${FINAL}" ] || { echo "no NVIDIA GPUs selected after filters" >&2; exit 1; }
 
+export CUDA_DEVICE_ORDER="PCI_BUS_ID"
+export CUDA_VISIBLE_DEVICES="${FINAL}"
+
 JOHN_DEVICES=""
+local_id=1
 for id in $(echo "${FINAL}" | tr ',' ' '); do
-    opencl_id=$((id + 1))
+    opencl_id="${local_id}"
     if [ -z "${JOHN_DEVICES}" ]; then
         JOHN_DEVICES="${opencl_id}"
     else
         JOHN_DEVICES="${JOHN_DEVICES},${opencl_id}"
     fi
+    local_id=$((local_id + 1))
 done
 
 echo "loader=john"
 echo "selected_gpus=${FINAL}"
 echo "john_devices=${JOHN_DEVICES}"
+echo "stagger_seconds=${STAGGER_SECONDS}"
 
 cd "${JOHN_DIR}"
 
@@ -189,14 +198,58 @@ CHOSEN_FORMAT=$(choose_format) || {
     exit 1
 }
 
-echo "format=${CHOSEN_FORMAT}"
+run_john_loop() {
+    opencl_id="$1"
+    deadline="$2"
+    round=0
+    while :; do
+        now=$(date +%s)
+        remaining=$((deadline - now))
+        if [ "${remaining}" -le 0 ]; then
+            break
+        fi
+        round=$((round + 1))
+        slice="${remaining}"
+        if [ "${slice}" -gt "${TEST_SLICE_SECONDS}" ]; then
+            slice="${TEST_SLICE_SECONDS}"
+        fi
+        echo "device=${opencl_id} round=${round} remaining_sec=${remaining} slice_sec=${slice}"
+        ./john --test="${slice}" --format="${CHOSEN_FORMAT}" --devices="${opencl_id}" || return 1
+    done
+}
+
 PIDS=""
+cleanup() {
+    rc=$?
+    trap - EXIT INT TERM
+    for pid in ${PIDS}; do
+        kill "${pid}" 2>/dev/null || true
+    done
+    for pid in ${PIDS}; do
+        wait "${pid}" 2>/dev/null || true
+    done
+    exit "${rc}"
+}
+trap cleanup EXIT INT TERM
+
+echo "format=${CHOSEN_FORMAT}"
+echo "target_seconds=${DURATION_SEC}"
+echo "slice_seconds=${TEST_SLICE_SECONDS}"
+TOTAL_DEVICES=$(echo "${JOHN_DEVICES}" | tr ',' '\n' | awk 'NF' | wc -l | tr -d '[:space:]')
 _first=1
+pos=0
 for opencl_id in $(echo "${JOHN_DEVICES}" | tr ',' ' '); do
+    pos=$((pos + 1))
     [ "${_first}" = "1" ] || sleep 3
     _first=0
-    ./john --test="${SECONDS}" --format="${CHOSEN_FORMAT}" --devices="${opencl_id}" &
-    PIDS="${PIDS} $!"
+    extra_sec=$(( STAGGER_SECONDS * (TOTAL_DEVICES - pos) ))
+    deadline=$(( $(date +%s) + DURATION_SEC + extra_sec ))
+    run_john_loop "${opencl_id}" "${deadline}" &
+    pid=$!
+    PIDS="${PIDS} ${pid}"
+    if [ "${STAGGER_SECONDS}" -gt 0 ] && [ "${pos}" -lt "${TOTAL_DEVICES}" ]; then
+        sleep "${STAGGER_SECONDS}"
+    fi
 done
 FAIL=0
 for pid in ${PIDS}; do

iso/overlay/usr/local/bin/bee-nccl-gpu-stress

@@ -70,6 +70,8 @@ echo "gpu_count=${GPU_COUNT}"
 echo "range=${MIN_BYTES}..${MAX_BYTES}"
 echo "iters=${ITERS}"
 
+export CUDA_DEVICE_ORDER="PCI_BUS_ID"
+
 deadline=$(( $(date +%s) + SECONDS ))
 round=0
 

iso/overlay/usr/local/bin/bee-nvidia-load

@@ -6,10 +6,28 @@ NVIDIA_KO_DIR="/usr/local/lib/nvidia"
 
 log() { echo "[bee-nvidia] $*"; }
 
+read_nvidia_modules_flavor() {
+    if [ -f /etc/bee-nvidia-modules-flavor ]; then
+        flavor="$(tr -d '[:space:]' </etc/bee-nvidia-modules-flavor 2>/dev/null)"
+        case "$flavor" in
+            open|proprietary)
+                echo "$flavor"
+                return 0
+                ;;
+        esac
+    fi
+    echo "proprietary"
+}
+
 log "kernel: $(uname -r)"
 
-# Skip if no NVIDIA GPU present (PCI vendor 10de)
-if ! lspci -nn 2>/dev/null | grep -qi '10de:'; then
+# Skip if no NVIDIA display/compute GPU is present.
+# Match only display-class PCI functions (0300 VGA, 0302 3D controller) from vendor 10de.
+have_nvidia_gpu() {
+    lspci -Dn 2>/dev/null | awk '$2 ~ /^03(00|02):$/ && $3 ~ /^10de:/ { found=1; exit } END { exit(found ? 0 : 1) }'
+}
+
+if ! have_nvidia_gpu; then
     log "no NVIDIA GPU detected — skipping module load"
     exit 0
 fi
@@ -40,6 +58,8 @@ if [ -z "$nvidia_mode" ]; then
     nvidia_mode="normal"
 fi
 log "boot mode: $nvidia_mode"
+nvidia_modules_flavor="$(read_nvidia_modules_flavor)"
+log "modules flavor: $nvidia_modules_flavor"
 
 load_module() {
     mod="$1"
@@ -50,11 +70,93 @@ load_module() {
         log "WARN: not found: $ko"
         return 1
     fi
-    if insmod "$ko" "$@"; then
+    if timeout 90 insmod "$ko" "$@"; then
         log "loaded: $mod $*"
         return 0
     fi
-    log "WARN: failed to load: $mod"
+    log "WARN: failed to load: $mod (exit $?)"
+    dmesg | tail -n 10 | sed 's/^/  dmesg: /' || true
+    return 1
+}
+
+nvidia_is_functional() {
+    grep -q ' nvidiactl$' /proc/devices 2>/dev/null
+}
+
+load_module_with_gsp_fallback() {
+    ko="$NVIDIA_KO_DIR/nvidia.ko"
+    if [ ! -f "$ko" ]; then
+        log "ERROR: not found: $ko"
+        return 1
+    fi
+
+    # Run insmod in background — on some converted SXM→PCIe cards GSP enters an
+    # infinite crash/reload loop and insmod never returns. We check for successful
+    # initialization by polling /proc/devices for nvidiactl instead of waiting for
+    # insmod to exit.
+    log "loading nvidia (GSP enabled, timeout 90s)"
+    insmod "$ko" &
+    _insmod_pid=$!
+
+    _waited=0
+    while [ $_waited -lt 90 ]; do
+        if nvidia_is_functional; then
+            log "loaded: nvidia (GSP enabled, ${_waited}s)"
+            echo "gsp-on" > /run/bee-nvidia-mode
+            return 0
+        fi
+        # Check if insmod exited with an error before timeout
+        if ! kill -0 "$_insmod_pid" 2>/dev/null; then
+            wait "$_insmod_pid"
+            _rc=$?
+            if [ $_rc -ne 0 ]; then
+                log "nvidia load failed (exit $_rc)"
+                dmesg | tail -n 10 | sed 's/^/  dmesg: /' || true
+                return 1
+            fi
+            # insmod exited 0 but nvidiactl not yet in /proc/devices — give it a moment
+            sleep 2
+            if nvidia_is_functional; then
+                log "loaded: nvidia (GSP enabled, ${_waited}s)"
+                return 0
+            fi
+            log "insmod exited 0 but nvidiactl missing — treating as failure"
+            return 1
+        fi
+        sleep 1
+        _waited=$((_waited + 1))
+    done
+
+    # GSP init timed out — kill the hanging insmod and attempt gsp-off fallback
+    log "nvidia GSP init timed out after 90s"
+    kill "$_insmod_pid" 2>/dev/null || true
+    wait "$_insmod_pid" 2>/dev/null || true
+
+    # Attempt to unload the partially-initialized module
+    if ! rmmod nvidia 2>/dev/null; then
+        # Module is stuck in the kernel — cannot reload with different params.
+        # User must reboot and select bee.nvidia.mode=gsp-off at boot menu.
+        log "ERROR: rmmod nvidia failed (EBUSY) — module stuck in kernel"
+        log "ERROR: reboot and select 'EASY-BEE (advanced) -> GSP=off' in boot menu"
+        echo "gsp-stuck" > /run/bee-nvidia-mode
+        return 1
+    fi
+
+    sleep 2
+    log "retrying with NVreg_EnableGpuFirmware=0"
+    log "WARNING: GSP disabled — power management will run via CPU path, not GPU firmware"
+
+    if insmod "$ko" NVreg_EnableGpuFirmware=0; then
+        if nvidia_is_functional; then
+            log "loaded: nvidia (GSP disabled)"
+            echo "gsp-off" > /run/bee-nvidia-mode
+            return 0
+        fi
+        log "insmod gsp-off exited 0 but nvidiactl missing"
+        return 1
+    fi
+
+    log "nvidia load failed (GSP=off)"
     dmesg | tail -n 10 | sed 's/^/  dmesg: /' || true
     return 1
 }
@@ -68,37 +170,54 @@ load_host_module() {
     return 1
 }
 
-case "$nvidia_mode" in
-    normal|full)
-        if ! load_module nvidia; then
-            exit 1
-        fi
-        # nvidia-modeset on some server kernels needs ACPI video helper symbols
-        # exported by the generic "video" module. Best-effort only; compute paths
-        # remain functional even if display-related modules stay absent.
-        load_host_module video || true
-        load_module nvidia-modeset || true
-        load_module nvidia-uvm || true
-        ;;
-    gsp-off|safe)
-        # NVIDIA documents that GSP firmware is enabled by default on newer GPUs and can
-        # be disabled via NVreg_EnableGpuFirmware=0. Safe mode keeps the live ISO on the
-        # conservative path for platforms where full boot-time GSP init is unstable.
-        if ! load_module nvidia NVreg_EnableGpuFirmware=0; then
-            exit 1
-        fi
-        log "GSP-off mode: skipping nvidia-modeset and nvidia-uvm during boot"
-        ;;
-    nomsi|*)
-        # nomsi: disable MSI-X/MSI interrupts — use when RmInitAdapter fails with
-        # "Failed to enable MSI-X" on one or more GPUs (IOMMU group interrupt limits).
-        # NVreg_EnableMSI=0 forces legacy INTx interrupts for all GPUs.
-        if ! load_module nvidia NVreg_EnableGpuFirmware=0 NVreg_EnableMSI=0; then
-            exit 1
-        fi
-        log "nomsi mode: MSI-X disabled (NVreg_EnableMSI=0), skipping nvidia-modeset and nvidia-uvm"
-        ;;
-esac
+if [ "$nvidia_modules_flavor" = "open" ]; then
+    case "$nvidia_mode" in
+        gsp-off|safe|nomsi)
+            log "ignoring boot mode ${nvidia_mode} for open NVIDIA modules"
+            ;;
+    esac
+    if ! load_module nvidia; then
+        exit 1
+    fi
+    # nvidia-modeset on some server kernels needs ACPI video helper symbols
+    # exported by the generic "video" module. Best-effort only; compute paths
+    # remain functional even if display-related modules stay absent.
+    load_host_module video || true
+    load_module nvidia-modeset || true
+    load_module nvidia-uvm || true
+else
+    case "$nvidia_mode" in
+        normal|full)
+            if ! load_module_with_gsp_fallback; then
+                exit 1
+            fi
+            # nvidia-modeset on some server kernels needs ACPI video helper symbols
+            # exported by the generic "video" module. Best-effort only; compute paths
+            # remain functional even if display-related modules stay absent.
+            load_host_module video || true
+            load_module nvidia-modeset || true
+            load_module nvidia-uvm || true
+            ;;
+        gsp-off|safe)
+            # NVIDIA documents that GSP firmware is enabled by default on newer GPUs and can
+            # be disabled via NVreg_EnableGpuFirmware=0. Safe mode keeps the live ISO on the
+            # conservative path for platforms where full boot-time GSP init is unstable.
+            if ! load_module nvidia NVreg_EnableGpuFirmware=0; then
+                exit 1
+            fi
+            log "GSP-off mode: skipping nvidia-modeset and nvidia-uvm during boot"
+            ;;
+        nomsi|*)
+            # nomsi: disable MSI-X/MSI interrupts — use when RmInitAdapter fails with
+            # "Failed to enable MSI-X" on one or more GPUs (IOMMU group interrupt limits).
+            # NVreg_EnableMSI=0 forces legacy INTx interrupts for all GPUs.
+            if ! load_module nvidia NVreg_EnableGpuFirmware=0 NVreg_EnableMSI=0; then
+                exit 1
+            fi
+            log "nomsi mode: MSI-X disabled (NVreg_EnableMSI=0), skipping nvidia-modeset and nvidia-uvm"
+            ;;
+    esac
+fi
 
 # Create /dev/nvidia* device nodes (udev rules absent since we use .run installer)
 nvidia_major=$(grep -m1 ' nvidiactl$' /proc/devices | awk '{print $1}')
@@ -127,14 +246,45 @@ fi
 ldconfig 2>/dev/null || true
 log "ldconfig refreshed"
 
+# Keep persistence mode enabled across the session so dcgmi / stress tools do
+# not fail with deployment warnings on otherwise healthy GPUs.
+if command -v nvidia-smi >/dev/null 2>&1; then
+    if nvidia-smi -pm 1 >/dev/null 2>&1; then
+        log "enabled NVIDIA persistence mode"
+    else
+        log "WARN: failed to enable NVIDIA persistence mode"
+    fi
+else
+    log "WARN: nvidia-smi not found — cannot enable persistence mode"
+fi
+
 # Start DCGM host engine so dcgmi can discover GPUs.
-# nv-hostengine must run before any dcgmi command — without it, dcgmi reports
-# "group is empty" even when GPUs and modules are present.
-# Skip if already running (e.g. started by a dcgm systemd service or prior boot).
+# nv-hostengine must run after the NVIDIA modules and device nodes are ready.
+# If it started too early (for example via systemd before bee-nvidia-load), it can
+# keep a stale empty inventory and dcgmi diag later reports no testable entities.
 if command -v nv-hostengine >/dev/null 2>&1; then
     if pgrep -x nv-hostengine >/dev/null 2>&1; then
-        log "nv-hostengine already running — skipping"
-    else
+        if command -v pkill >/dev/null 2>&1; then
+            pkill -x nv-hostengine >/dev/null 2>&1 || true
+            tries=0
+            while pgrep -x nv-hostengine >/dev/null 2>&1; do
+                tries=$((tries + 1))
+                if [ "${tries}" -ge 10 ]; then
+                    log "WARN: nv-hostengine is still running after restart request"
+                    break
+                fi
+                sleep 1
+            done
+            if pgrep -x nv-hostengine >/dev/null 2>&1; then
+                log "WARN: keeping existing nv-hostengine process"
+            else
+                log "nv-hostengine restarted"
+            fi
+        else
+            log "WARN: pkill not found — cannot refresh nv-hostengine inventory"
+        fi
+    fi
+    if ! pgrep -x nv-hostengine >/dev/null 2>&1; then
         nv-hostengine
         log "nv-hostengine started"
     fi

iso/overlay/usr/local/bin/bee-openbox-session

@@ -7,16 +7,24 @@ xset s off
 xset -dpms
 xset s noblank
 
+# Set desktop background.
+if [ -f /usr/share/bee/wallpaper.png ]; then
+    feh --bg-fill /usr/share/bee/wallpaper.png
+else
+    xsetroot -solid '#f6c90e'
+fi
+
 tint2 &
 
-# Wait up to 120s for bee-web to bind. The web server starts immediately now
-# (audit is deferred), so this should succeed in a few seconds on most hardware.
-i=0
-while [ $i -lt 120 ]; do
-    if curl -sf http://localhost/healthz >/dev/null 2>&1; then break; fi
+# Wait up to 60s for bee-web before opening Chromium.
+# Without this Chromium gets connection-refused and shows a blank page.
+_i=0
+while [ $_i -lt 60 ]; do
+    curl -sf http://localhost/healthz >/dev/null 2>&1 && break
     sleep 1
-    i=$((i+1))
+    _i=$((_i+1))
 done
+unset _i
 
 chromium \
     --disable-infobars \
@@ -24,7 +32,8 @@ chromium \
     --no-first-run \
     --disable-session-crashed-bubble \
     --disable-features=TranslateUI \
+    --user-data-dir=/tmp/bee-chrome \
     --start-maximized \
-    http://localhost/ &
+    http://localhost/loading &
 
 exec openbox

iso/overlay/usr/local/bin/bee-selfheal

@@ -0,0 +1,99 @@
+#!/bin/bash
+# bee-selfheal — periodic best-effort recovery for critical live ISO services.
+
+set -u
+
+LOG_PREFIX="bee-selfheal"
+EXPORT_DIR="/appdata/bee/export"
+AUDIT_JSON="${EXPORT_DIR}/bee-audit.json"
+RUNTIME_JSON="${EXPORT_DIR}/runtime-health.json"
+LOCK_DIR="/run/bee-selfheal.lock"
+
+log() {
+    echo "[${LOG_PREFIX}] $*"
+}
+
+have_nvidia_gpu() {
+    lspci -Dn 2>/dev/null | awk '$2 ~ /^03(00|02):$/ && $3 ~ /^10de:/ { found=1; exit } END { exit(found ? 0 : 1) }'
+}
+
+service_active() {
+    systemctl is-active --quiet "$1" 2>/dev/null
+}
+
+restart_service() {
+    local svc="$1"
+    if systemctl restart "$svc" >/dev/null 2>&1; then
+        log "restarted ${svc}"
+        return 0
+    fi
+    log "WARN: failed to restart ${svc}"
+    return 1
+}
+
+file_ready() {
+    [ -s "$1" ]
+}
+
+artifact_state() {
+    local path="$1"
+    if [ -s "${path}" ]; then
+        echo "ready"
+        return 0
+    fi
+    if [ -e "${path}.tmp" ]; then
+        echo "interrupted"
+        return 0
+    fi
+    echo "missing"
+}
+
+web_healthy() {
+    bash -c 'exec 3<>/dev/tcp/127.0.0.1/80 && printf "GET /healthz HTTP/1.0\r\nHost: localhost\r\n\r\n" >&3 && grep -q "^ok$" <&3' \
+        >/dev/null 2>&1
+}
+
+mkdir -p "${EXPORT_DIR}" /run
+
+if ! mkdir "${LOCK_DIR}" 2>/dev/null; then
+    log "another self-heal run is already active"
+    exit 0
+fi
+trap 'rmdir "${LOCK_DIR}" >/dev/null 2>&1 || true' EXIT
+
+log "start"
+
+if have_nvidia_gpu && [ ! -e /dev/nvidia0 ]; then
+    log "NVIDIA GPU detected but /dev/nvidia0 is missing"
+    restart_service bee-nvidia.service || true
+fi
+
+runtime_state="$(artifact_state "${RUNTIME_JSON}")"
+if [ "${runtime_state}" != "ready" ]; then
+    if [ "${runtime_state}" = "interrupted" ]; then
+        log "runtime-health.json.tmp exists — interrupted runtime-health write detected"
+    else
+        log "runtime-health.json missing or empty"
+    fi
+    restart_service bee-preflight.service || true
+fi
+
+audit_state="$(artifact_state "${AUDIT_JSON}")"
+if [ "${audit_state}" != "ready" ]; then
+    if [ "${audit_state}" = "interrupted" ]; then
+        log "bee-audit.json.tmp exists — interrupted audit write detected"
+    else
+        log "bee-audit.json missing or empty"
+    fi
+    restart_service bee-audit.service || true
+fi
+
+if ! service_active bee-web.service; then
+    log "bee-web.service is not active"
+    restart_service bee-web.service || true
+elif ! web_healthy; then
+    log "bee-web health check failed"
+    restart_service bee-web.service || true
+fi
+
+log "done"