Add NVIDIA benchmark reporting flow

Tasks are now started simultaneously when multiple are enqueued (e.g.
Run All). The worker drains all pending tasks at once and launches each
in its own goroutine, waiting via WaitGroup. kmsg watcher updated to
use a shared event window with a reference counter across concurrent tasks.

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

PLAN.md

@@ -343,9 +343,9 @@ Planned code shape:
 - `bee tui` can rerun the audit manually
 - `bee tui` can export the latest audit JSON to removable media
 - `bee tui` can show health summary and run NVIDIA/memory/storage acceptance tests
-- NVIDIA SAT now includes a lightweight in-image GPU stress step via `bee-gpu-stress`
+- NVIDIA SAT now includes a lightweight in-image GPU stress step via `bee-gpu-burn`
 - SAT summaries now expose `overall_status` plus per-job `OK/FAILED/UNSUPPORTED`
-- Memory/GPU SAT runtime defaults can be overridden via `BEE_MEMTESTER_*` and `BEE_GPU_STRESS_*`
+- Memory SAT runtime defaults can be overridden via `BEE_MEMTESTER_*`
 - removable export requires explicit target selection, mount, confirmation, copy, and cleanup
 
 ### 2.6 — Vendor utilities and optional assets

audit/Makefile

@@ -1,5 +1,7 @@
 LISTEN ?= :8080
 AUDIT_PATH ?=
+VERSION ?= $(shell sh ./scripts/resolve-version.sh)
+GO_LDFLAGS := -X main.Version=$(VERSION)
 
 RUN_ARGS := web --listen $(LISTEN)
 ifneq ($(AUDIT_PATH),)
@@ -9,10 +11,10 @@ endif
 .PHONY: run build test
 
 run:
-	go run ./cmd/bee $(RUN_ARGS)
+	go run -ldflags "$(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
@@ -100,6 +89,7 @@ func printRootUsage(w io.Writer) {
   bee support-bundle --output stdout|file:<path>
   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
@@ -356,6 +348,7 @@ func runSAT(args []string, stdout, stderr io.Writer) int {
 	fs := flag.NewFlagSet("sat", flag.ContinueOnError)
 	fs.SetOutput(stderr)
 	duration := fs.Int("duration", 0, "stress-ng duration in seconds (cpu only; default: 60)")
+	diagLevel := fs.Int("diag-level", 0, "DCGM diagnostic level for nvidia (1=quick, 2=medium, 3=targeted stress, 4=extended stress; default: 1)")
 	if err := fs.Parse(args[1:]); err != nil {
 		if err == flag.ErrHelp {
 			return 0
@@ -370,7 +363,7 @@ func runSAT(args []string, stdout, stderr io.Writer) int {
 	target := args[0]
 	if target != "nvidia" && target != "memory" && target != "storage" && target != "cpu" {
 		fmt.Fprintf(stderr, "bee sat: unknown target %q\n", target)
-		fmt.Fprintln(stderr, "usage: bee sat nvidia|memory|storage|cpu [--duration <seconds>]")
+		fmt.Fprintln(stderr, "usage: bee sat nvidia|memory|storage|cpu [--duration <seconds>] [--diag-level <1-4>]")
 		return 2
 	}
 
@@ -382,7 +375,12 @@ func runSAT(args []string, stdout, stderr io.Writer) int {
 	logLine := func(s string) { fmt.Fprintln(os.Stderr, s) }
 	switch target {
 	case "nvidia":
-		archive, err = application.RunNvidiaAcceptancePack("", logLine)
+		level := *diagLevel
+		if level > 0 {
+			_, err = application.RunNvidiaAcceptancePackWithOptions(context.Background(), "", level, nil, logLine)
+		} else {
+			archive, err = application.RunNvidiaAcceptancePack("", logLine)
+		}
 	case "memory":
 		archive, err = application.RunMemoryAcceptancePackCtx(context.Background(), "", logLine)
 	case "storage":
@@ -401,3 +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 {
@@ -80,6 +83,7 @@ type installer interface {
 	ListInstallDisks() ([]platform.InstallDisk, error)
 	InstallToDisk(ctx context.Context, device string, logFile string) error
 	IsLiveMediaInRAM() bool
+	LiveBootSource() platform.LiveBootSource
 	RunInstallToRAM(ctx context.Context, logFunc func(string)) error
 }
 
@@ -100,6 +104,10 @@ func (a *App) IsLiveMediaInRAM() bool {
 	return a.installer.IsLiveMediaInRAM()
 }
 
+func (a *App) LiveBootSource() platform.LiveBootSource {
+	return a.installer.LiveBootSource()
+}
+
 func (a *App) RunInstallToRAM(ctx context.Context, logFunc func(string)) error {
 	return a.installer.RunInstallToRAM(ctx, logFunc)
 }
@@ -107,6 +115,8 @@ 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)
+	RunNvidiaBenchmark(ctx context.Context, baseDir string, opts platform.NvidiaBenchmarkOptions, 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)
 	RunCPUAcceptancePack(ctx context.Context, baseDir string, durationSec int, logFunc func(string)) (string, error)
@@ -130,7 +140,7 @@ type runtimeChecker interface {
 }
 
 func New(platform *platform.System) *App {
-	return &App{
+	a := &App{
 		network:   platform,
 		services:  platform,
 		exports:   platform,
@@ -139,19 +149,32 @@ 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,
 // and returns the updated JSON. Used by the web UI to serve always-fresh status.
 func ApplySATOverlay(auditJSON []byte) ([]byte, error) {
-	var snap schema.HardwareIngestRequest
-	if err := json.Unmarshal(auditJSON, &snap); err != nil {
+	snap, err := readAuditSnapshot(auditJSON)
+	if err != nil {
 		return nil, err
 	}
-	applyLatestSATStatuses(&snap.Hardware, DefaultSATBaseDir)
+	applyLatestSATStatuses(&snap.Hardware, DefaultSATBaseDir, nil)
 	return json.MarshalIndent(snap, "", "  ")
 }
 
+func readAuditSnapshot(auditJSON []byte) (schema.HardwareIngestRequest, error) {
+	var snap schema.HardwareIngestRequest
+	if err := json.Unmarshal(auditJSON, &snap); err != nil {
+		return schema.HardwareIngestRequest{}, err
+	}
+	collector.NormalizeSnapshot(&snap.Hardware, snap.CollectedAt)
+	return snap, nil
+}
+
 func (a *App) RunAudit(runtimeMode runtimeenv.Mode, output string) (string, error) {
 	if runtimeMode == runtimeenv.ModeLiveCD {
 		if err := a.runtime.CaptureTechnicalDump(DefaultTechDumpDir); err != nil {
@@ -159,7 +182,7 @@ 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
 	}
@@ -174,10 +197,7 @@ 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
@@ -202,10 +222,7 @@ 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
@@ -275,6 +292,9 @@ func (a *App) ExportLatestAudit(target platform.RemovableTarget) (string, error)
 	if err != nil {
 		return "", err
 	}
+	if normalized, normErr := ApplySATOverlay(data); normErr == nil {
+		data = normalized
+	}
 	if err := os.WriteFile(tmpPath, data, 0644); err != nil {
 		return "", err
 	}
@@ -508,6 +528,28 @@ func (a *App) RunNvidiaAcceptancePackWithOptions(ctx context.Context, baseDir st
 	return ActionResult{Title: "NVIDIA DCGM", Body: body}, err
 }
 
+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) RunNvidiaStressPackCtx(ctx context.Context, baseDir string, opts platform.NvidiaStressOptions, logFunc func(string)) (string, error) {
+	if strings.TrimSpace(baseDir) == "" {
+		baseDir = DefaultSATBaseDir
+	}
+	return a.sat.RunNvidiaStressPack(ctx, baseDir, opts, logFunc)
+}
+
 func (a *App) RunMemoryAcceptancePack(baseDir string, logFunc func(string)) (string, error) {
 	return a.RunMemoryAcceptancePackCtx(context.Background(), baseDir, logFunc)
 }
@@ -721,6 +763,7 @@ func (a *App) HealthSummaryResult() ActionResult {
 	if err := json.Unmarshal(raw, &snapshot); err != nil {
 		return ActionResult{Title: "Health summary", Body: "Audit JSON is unreadable."}
 	}
+	collector.NormalizeSnapshot(&snapshot.Hardware, snapshot.CollectedAt)
 
 	summary := collector.BuildHealthSummary(snapshot.Hardware)
 	var body strings.Builder
@@ -755,6 +798,7 @@ func (a *App) MainBanner() string {
 	if err := json.Unmarshal(raw, &snapshot); err != nil {
 		return ""
 	}
+	collector.NormalizeSnapshot(&snapshot.Hardware, snapshot.CollectedAt)
 
 	var lines []string
 	if system := formatSystemLine(snapshot.Hardware.Board); system != "" {

audit/internal/app/app_test.go

@@ -120,14 +120,16 @@ func (f fakeTools) CheckTools(names []string) []platform.ToolStatus {
 }
 
 type fakeSAT struct {
-	runNvidiaFn      func(string) (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)
+	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)
 }
 
 func (f fakeSAT) RunNvidiaAcceptancePack(baseDir string, _ func(string)) (string, error) {
@@ -138,6 +140,20 @@ 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) RunNvidiaStressPack(_ context.Context, baseDir string, opts platform.NvidiaStressOptions, _ func(string)) (string, error) {
+	if f.runNvidiaStressFn != nil {
+		return f.runNvidiaStressFn(baseDir, opts)
+	}
+	return f.runNvidiaFn(baseDir)
+}
+
 func (f fakeSAT) ListNvidiaGPUs() ([]platform.NvidiaGPU, error) {
 	if f.listNvidiaGPUsFn != nil {
 		return f.listNvidiaGPUsFn()
@@ -652,13 +668,50 @@ func TestHealthSummaryResultIncludesCompactSATSummary(t *testing.T) {
 	}
 }
 
+func TestApplySATOverlayFiltersIgnoredLegacyDevices(t *testing.T) {
+	tmp := t.TempDir()
+	oldSATBaseDir := DefaultSATBaseDir
+	DefaultSATBaseDir = filepath.Join(tmp, "sat")
+	t.Cleanup(func() { DefaultSATBaseDir = oldSATBaseDir })
+
+	raw := `{
+	  "collected_at": "2026-03-15T10:00:00Z",
+	  "hardware": {
+	    "board": {"serial_number": "SRV123"},
+	    "storage": [
+	      {"model": "Virtual HDisk0", "serial_number": "AAAABBBBCCCC3"},
+	      {"model": "PASCARI", "serial_number": "DISK1", "status": "OK"}
+	    ],
+	    "pcie_devices": [
+	      {"device_class": "Co-processor", "model": "402xx Series QAT", "status": "OK"},
+	      {"device_class": "VideoController", "model": "NVIDIA H100", "status": "OK"}
+	    ]
+	  }
+	}`
+
+	got, err := ApplySATOverlay([]byte(raw))
+	if err != nil {
+		t.Fatalf("ApplySATOverlay error: %v", err)
+	}
+	text := string(got)
+	if contains(text, "Virtual HDisk0") {
+		t.Fatalf("overlaid audit should drop virtual hdisk:\n%s", text)
+	}
+	if contains(text, "\"device_class\": \"Co-processor\"") {
+		t.Fatalf("overlaid audit should drop co-processors:\n%s", text)
+	}
+	if !contains(text, "PASCARI") || !contains(text, "NVIDIA H100") {
+		t.Fatalf("overlaid audit should keep real devices:\n%s", text)
+	}
+}
+
 func TestBuildSupportBundleIncludesExportDirContents(t *testing.T) {
 	tmp := t.TempDir()
 	exportDir := filepath.Join(tmp, "export")
 	if err := os.MkdirAll(filepath.Join(exportDir, "bee-sat", "memory-run"), 0755); err != nil {
 		t.Fatal(err)
 	}
-	if err := os.WriteFile(filepath.Join(exportDir, "bee-audit.json"), []byte(`{"ok":true}`), 0644); err != nil {
+	if err := os.WriteFile(filepath.Join(exportDir, "bee-audit.json"), []byte(`{"collected_at":"2026-03-15T10:00:00Z","hardware":{"board":{"serial_number":"SRV123"},"storage":[{"model":"Virtual HDisk0","serial_number":"AAAABBBBCCCC3"},{"model":"PASCARI","serial_number":"DISK1"}],"pcie_devices":[{"device_class":"Co-processor","model":"402xx Series QAT"},{"device_class":"VideoController","model":"NVIDIA H100"}]}}`), 0644); err != nil {
 		t.Fatal(err)
 	}
 	if err := os.WriteFile(filepath.Join(exportDir, "bee-sat", "memory-run", "verbose.log"), []byte("sat verbose"), 0644); err != nil {
@@ -690,6 +743,7 @@ func TestBuildSupportBundleIncludesExportDirContents(t *testing.T) {
 
 	tr := tar.NewReader(gzr)
 	var names []string
+	var auditJSON string
 	for {
 		hdr, err := tr.Next()
 		if errors.Is(err, io.EOF) {
@@ -699,6 +753,33 @@ func TestBuildSupportBundleIncludesExportDirContents(t *testing.T) {
 			t.Fatalf("read tar entry: %v", err)
 		}
 		names = append(names, hdr.Name)
+		if contains(hdr.Name, "/export/bee-audit.json") {
+			body, err := io.ReadAll(tr)
+			if err != nil {
+				t.Fatalf("read audit entry: %v", err)
+			}
+			auditJSON = string(body)
+		}
+	}
+
+	for _, want := range []string{
+		"/system/ip-link.txt",
+		"/system/ip-link-stats.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
@@ -713,6 +794,12 @@ func TestBuildSupportBundleIncludesExportDirContents(t *testing.T) {
 	if !foundRaw {
 		t.Fatalf("support bundle missing raw SAT log, names=%v", names)
 	}
+	if contains(auditJSON, "Virtual HDisk0") || contains(auditJSON, "\"device_class\": \"Co-processor\"") {
+		t.Fatalf("support bundle should normalize ignored devices:\n%s", auditJSON)
+	}
+	if !contains(auditJSON, "PASCARI") || !contains(auditJSON, "NVIDIA H100") {
+		t.Fatalf("support bundle should keep real devices:\n%s", auditJSON)
+	}
 }
 
 func TestMainBanner(t *testing.T) {
@@ -726,6 +813,10 @@ func TestMainBanner(t *testing.T) {
 	product := "PowerEdge R760"
 	cpuModel := "Intel Xeon Gold 6430"
 	memoryType := "DDR5"
+	memorySerialA := "DIMM-A"
+	memorySerialB := "DIMM-B"
+	storageSerialA := "DISK-A"
+	storageSerialB := "DISK-B"
 	gpuClass := "VideoController"
 	gpuModel := "NVIDIA H100"
 
@@ -741,12 +832,12 @@ func TestMainBanner(t *testing.T) {
 				{Model: &cpuModel},
 			},
 			Memory: []schema.HardwareMemory{
-				{Present: &trueValue, SizeMB: intPtr(524288), Type: &memoryType},
-				{Present: &trueValue, SizeMB: intPtr(524288), Type: &memoryType},
+				{Present: &trueValue, SizeMB: intPtr(524288), Type: &memoryType, SerialNumber: &memorySerialA},
+				{Present: &trueValue, SizeMB: intPtr(524288), Type: &memoryType, SerialNumber: &memorySerialB},
 			},
 			Storage: []schema.HardwareStorage{
-				{Present: &trueValue, SizeGB: intPtr(3840)},
-				{Present: &trueValue, SizeGB: intPtr(3840)},
+				{Present: &trueValue, SizeGB: intPtr(3840), SerialNumber: &storageSerialA},
+				{Present: &trueValue, SizeGB: intPtr(3840), SerialNumber: &storageSerialB},
 			},
 			PCIeDevices: []schema.HardwarePCIeDevice{
 				{DeviceClass: &gpuClass, Model: &gpuModel},

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 +1,266 @@
+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", "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

@@ -9,7 +9,7 @@ import (
 	"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
 	}
@@ -28,6 +28,8 @@ 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 satSummary struct {
@@ -206,6 +208,86 @@ 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)
@@ -53,7 +53,7 @@ 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)

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",
 }
 
@@ -27,15 +29,118 @@ var supportBundleCommands = []struct {
 	cmd  []string
 }{
 	{name: "system/uname.txt", cmd: []string{"uname", "-a"}},
+	{name: "system/cmdline.txt", cmd: []string{"cat", "/proc/cmdline"}},
 	{name: "system/lsmod.txt", cmd: []string{"lsmod"}},
 	{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-tail.txt", cmd: []string{"sh", "-c", "dmesg | tail -n 200"}},
+	{name: "system/dmesg.txt", cmd: []string{"dmesg"}},
+	{name: "system/nvidia-smi-q.txt", cmd: []string{"nvidia-smi", "-q"}},
+	{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")
+  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/$f" 2>/dev/null)"
+  done
+done
+`}},
+	{name: "system/ethtool-info.txt", cmd: []string{"sh", "-c", `
+if ! command -v ethtool >/dev/null 2>&1; then
+  echo "ethtool not found"
+  exit 0
+fi
+found=0
+for path in /sys/class/net/*; do
+  [ -e "$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
+`}},
 }
 
+var supportBundleOptionalFiles = []struct {
+	name string
+	src  string
+}{
+	{name: "system/kern.log", src: "/var/log/kern.log"},
+	{name: "system/syslog.txt", src: "/var/log/syslog"},
+}
+
+const supportBundleGlob = "bee-support-*.tar.gz"
+
 func BuildSupportBundle(exportDir string) (string, error) {
 	exportDir = strings.TrimSpace(exportDir)
 	if exportDir == "" {
@@ -75,6 +180,9 @@ func BuildSupportBundle(exportDir string) (string, error) {
 			return "", err
 		}
 	}
+	for _, item := range supportBundleOptionalFiles {
+		_ = copyOptionalFile(item.src, filepath.Join(stageRoot, item.name))
+	}
 	if err := writeManifest(filepath.Join(stageRoot, "manifest.txt"), exportDir, stageRoot); err != nil {
 		return "", err
 	}
@@ -86,34 +194,64 @@ func BuildSupportBundle(exportDir string) (string, error) {
 	return archivePath, nil
 }
 
+func LatestSupportBundlePath() (string, error) {
+	return latestSupportBundlePath(os.TempDir())
+}
+
 func cleanupOldSupportBundles(dir string) error {
-	matches, err := filepath.Glob(filepath.Join(dir, "bee-support-*.tar.gz"))
+	matches, err := filepath.Glob(filepath.Join(dir, supportBundleGlob))
 	if err != nil {
 		return err
 	}
-	type entry struct {
-		path string
-		mod  time.Time
+	entries := supportBundleEntries(matches)
+	for path, mod := range entries {
+		if time.Since(mod) > 24*time.Hour {
+			_ = os.Remove(path)
+			delete(entries, path)
+		}
 	}
-	list := make([]entry, 0, len(matches))
+	ordered := orderSupportBundles(entries)
+	if len(ordered) > 3 {
+		for _, old := range ordered[3:] {
+			_ = os.Remove(old)
+		}
+	}
+	return nil
+}
+
+func latestSupportBundlePath(dir string) (string, error) {
+	matches, err := filepath.Glob(filepath.Join(dir, supportBundleGlob))
+	if err != nil {
+		return "", err
+	}
+	ordered := orderSupportBundles(supportBundleEntries(matches))
+	if len(ordered) == 0 {
+		return "", os.ErrNotExist
+	}
+	return ordered[0], nil
+}
+
+func supportBundleEntries(matches []string) map[string]time.Time {
+	entries := make(map[string]time.Time, len(matches))
 	for _, match := range matches {
 		info, err := os.Stat(match)
 		if err != nil {
 			continue
 		}
-		if time.Since(info.ModTime()) > 24*time.Hour {
-			_ = os.Remove(match)
-			continue
-		}
-		list = append(list, entry{path: match, mod: info.ModTime()})
+		entries[match] = info.ModTime()
 	}
-	sort.Slice(list, func(i, j int) bool { return list[i].mod.After(list[j].mod) })
-	if len(list) > 3 {
-		for _, old := range list[3:] {
-			_ = os.Remove(old.path)
-		}
+	return entries
+}
+
+func orderSupportBundles(entries map[string]time.Time) []string {
+	ordered := make([]string, 0, len(entries))
+	for path := range entries {
+		ordered = append(ordered, path)
 	}
-	return nil
+	sort.Slice(ordered, func(i, j int) bool {
+		return entries[ordered[i]].After(entries[ordered[j]])
+	})
+	return ordered
 }
 
 func writeJournalDump(dst string) error {
@@ -152,6 +290,24 @@ func writeCommandOutput(dst string, cmd []string) error {
 	return os.WriteFile(dst, raw, 0644)
 }
 
+func copyOptionalFile(src, dst string) error {
+	in, err := os.Open(src)
+	if err != nil {
+		return err
+	}
+	defer in.Close()
+	if err := os.MkdirAll(filepath.Dir(dst), 0755); err != nil {
+		return err
+	}
+	out, err := os.Create(dst)
+	if err != nil {
+		return err
+	}
+	defer out.Close()
+	_, err = io.Copy(out, in)
+	return err
+}
+
 func writeManifest(dst, exportDir, stageRoot string) error {
 	if err := os.MkdirAll(filepath.Dir(dst), 0755); err != nil {
 		return err
@@ -215,7 +371,7 @@ func copyDirContents(srcDir, dstDir string) error {
 }
 
 func copyExportDirForSupportBundle(srcDir, dstDir string) error {
-	return copyDirContentsFiltered(srcDir, dstDir, func(rel string, info os.FileInfo) bool {
+	if err := copyDirContentsFiltered(srcDir, dstDir, func(rel string, info os.FileInfo) bool {
 		cleanRel := filepath.ToSlash(strings.TrimPrefix(filepath.Clean(rel), "./"))
 		if cleanRel == "" {
 			return true
@@ -227,7 +383,25 @@ func copyExportDirForSupportBundle(srcDir, dstDir string) error {
 			return false
 		}
 		return true
-	})
+	}); err != nil {
+		return err
+	}
+	return normalizeSupportBundleAuditJSON(filepath.Join(dstDir, "bee-audit.json"))
+}
+
+func normalizeSupportBundleAuditJSON(path string) error {
+	data, err := os.ReadFile(path)
+	if err != nil {
+		if os.IsNotExist(err) {
+			return nil
+		}
+		return err
+	}
+	normalized, err := ApplySATOverlay(data)
+	if err != nil {
+		return nil
+	}
+	return os.WriteFile(path, normalized, 0644)
 }
 
 func copyDirContentsFiltered(srcDir, dstDir string, keep func(rel string, info os.FileInfo) bool) error {

audit/internal/collector/finalize.go

@@ -1,10 +1,18 @@
 package collector
 
-import "bee/audit/internal/schema"
+import (
+	"bee/audit/internal/schema"
+	"strings"
+)
+
+func NormalizeSnapshot(snap *schema.HardwareSnapshot, collectedAt string) {
+	finalizeSnapshot(snap, collectedAt)
+}
 
 func finalizeSnapshot(snap *schema.HardwareSnapshot, collectedAt string) {
 	snap.Memory = filterMemory(snap.Memory)
 	snap.Storage = filterStorage(snap.Storage)
+	snap.PCIeDevices = filterPCIe(snap.PCIeDevices)
 	snap.PowerSupplies = filterPSUs(snap.PowerSupplies)
 
 	setComponentStatusMetadata(snap, collectedAt)
@@ -33,11 +41,25 @@ func filterStorage(disks []schema.HardwareStorage) []schema.HardwareStorage {
 		if disk.SerialNumber == nil || *disk.SerialNumber == "" {
 			continue
 		}
+		if disk.Model != nil && isVirtualHDiskModel(*disk.Model) {
+			continue
+		}
 		out = append(out, disk)
 	}
 	return out
 }
 
+func filterPCIe(devs []schema.HardwarePCIeDevice) []schema.HardwarePCIeDevice {
+	out := make([]schema.HardwarePCIeDevice, 0, len(devs))
+	for _, dev := range devs {
+		if dev.DeviceClass != nil && strings.Contains(strings.ToLower(strings.TrimSpace(*dev.DeviceClass)), "co-processor") {
+			continue
+		}
+		out = append(out, dev)
+	}
+	return out
+}
+
 func filterPSUs(psus []schema.HardwarePowerSupply) []schema.HardwarePowerSupply {
 	out := make([]schema.HardwarePowerSupply, 0, len(psus))
 	for _, psu := range psus {

audit/internal/collector/finalize_test.go

@@ -10,6 +10,10 @@ func TestFinalizeSnapshotFiltersComponentsWithoutRequiredSerials(t *testing.T) {
 	present := true
 	status := statusOK
 	serial := "SN-1"
+	virtualModel := "Virtual HDisk1"
+	realModel := "PASCARI"
+	coProcessorClass := "Co-processor"
+	gpuClass := "VideoController"
 
 	snap := schema.HardwareSnapshot{
 		Memory: []schema.HardwareMemory{
@@ -17,9 +21,15 @@ func TestFinalizeSnapshotFiltersComponentsWithoutRequiredSerials(t *testing.T) {
 			{Present: &present, HardwareComponentStatus: schema.HardwareComponentStatus{Status: &status}},
 		},
 		Storage: []schema.HardwareStorage{
+			{Model: &virtualModel, SerialNumber: &serial, HardwareComponentStatus: schema.HardwareComponentStatus{Status: &status}},
 			{SerialNumber: &serial, HardwareComponentStatus: schema.HardwareComponentStatus{Status: &status}},
+			{Model: &realModel, SerialNumber: &serial, HardwareComponentStatus: schema.HardwareComponentStatus{Status: &status}},
 			{HardwareComponentStatus: schema.HardwareComponentStatus{Status: &status}},
 		},
+		PCIeDevices: []schema.HardwarePCIeDevice{
+			{DeviceClass: &coProcessorClass, HardwareComponentStatus: schema.HardwareComponentStatus{Status: &status}},
+			{DeviceClass: &gpuClass, HardwareComponentStatus: schema.HardwareComponentStatus{Status: &status}},
+		},
 		PowerSupplies: []schema.HardwarePowerSupply{
 			{SerialNumber: &serial, HardwareComponentStatus: schema.HardwareComponentStatus{Status: &status}},
 			{HardwareComponentStatus: schema.HardwareComponentStatus{Status: &status}},
@@ -31,9 +41,12 @@ func TestFinalizeSnapshotFiltersComponentsWithoutRequiredSerials(t *testing.T) {
 	if len(snap.Memory) != 1 || snap.Memory[0].StatusCheckedAt == nil || *snap.Memory[0].StatusCheckedAt != collectedAt {
 		t.Fatalf("memory finalize mismatch: %+v", snap.Memory)
 	}
-	if len(snap.Storage) != 1 || snap.Storage[0].StatusCheckedAt == nil || *snap.Storage[0].StatusCheckedAt != collectedAt {
+	if len(snap.Storage) != 2 || snap.Storage[0].StatusCheckedAt == nil || *snap.Storage[0].StatusCheckedAt != collectedAt {
 		t.Fatalf("storage finalize mismatch: %+v", snap.Storage)
 	}
+	if len(snap.PCIeDevices) != 1 || snap.PCIeDevices[0].DeviceClass == nil || *snap.PCIeDevices[0].DeviceClass != gpuClass {
+		t.Fatalf("pcie finalize mismatch: %+v", snap.PCIeDevices)
+	}
 	if len(snap.PowerSupplies) != 1 || snap.PowerSupplies[0].StatusCheckedAt == nil || *snap.PowerSupplies[0].StatusCheckedAt != collectedAt {
 		t.Fatalf("psu finalize mismatch: %+v", snap.PowerSupplies)
 	}

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

@@ -13,14 +13,18 @@ import (
 const nvidiaVendorID = 0x10de
 
 type nvidiaGPUInfo struct {
-	BDF            string
-	Serial         string
-	VBIOS          string
-	TemperatureC   *float64
-	PowerW         *float64
-	ECCUncorrected *int64
-	ECCCorrected   *int64
-	HWSlowdown     *bool
+	BDF                string
+	Serial             string
+	VBIOS              string
+	TemperatureC       *float64
+	PowerW             *float64
+	ECCUncorrected     *int64
+	ECCCorrected       *int64
+	HWSlowdown         *bool
+	PCIeLinkGenCurrent *int
+	PCIeLinkGenMax     *int
+	PCIeLinkWidthCur   *int
+	PCIeLinkWidthMax   *int
 }
 
 // enrichPCIeWithNVIDIA enriches NVIDIA PCIe devices with data from nvidia-smi.
@@ -94,7 +98,7 @@ func enrichPCIeWithNVIDIAData(devs []schema.HardwarePCIeDevice, gpuByBDF map[str
 func queryNVIDIAGPUs() (map[string]nvidiaGPUInfo, error) {
 	out, err := exec.Command(
 		"nvidia-smi",
-		"--query-gpu=index,pci.bus_id,serial,vbios_version,temperature.gpu,power.draw,ecc.errors.uncorrected.aggregate.total,ecc.errors.corrected.aggregate.total,clocks_throttle_reasons.hw_slowdown",
+		"--query-gpu=index,pci.bus_id,serial,vbios_version,temperature.gpu,power.draw,ecc.errors.uncorrected.aggregate.total,ecc.errors.corrected.aggregate.total,clocks_throttle_reasons.hw_slowdown,pcie.link.gen.current,pcie.link.gen.max,pcie.link.width.current,pcie.link.width.max",
 		"--format=csv,noheader,nounits",
 	).Output()
 	if err != nil {
@@ -118,8 +122,8 @@ func parseNVIDIASMIQuery(raw string) (map[string]nvidiaGPUInfo, error) {
 		if len(rec) == 0 {
 			continue
 		}
-		if len(rec) < 9 {
-			return nil, fmt.Errorf("unexpected nvidia-smi columns: got %d, want 9", len(rec))
+		if len(rec) < 13 {
+			return nil, fmt.Errorf("unexpected nvidia-smi columns: got %d, want 13", len(rec))
 		}
 
 		bdf := normalizePCIeBDF(rec[1])
@@ -128,14 +132,18 @@ func parseNVIDIASMIQuery(raw string) (map[string]nvidiaGPUInfo, error) {
 		}
 
 		info := nvidiaGPUInfo{
-			BDF:            bdf,
-			Serial:         strings.TrimSpace(rec[2]),
-			VBIOS:          strings.TrimSpace(rec[3]),
-			TemperatureC:   parseMaybeFloat(rec[4]),
-			PowerW:         parseMaybeFloat(rec[5]),
-			ECCUncorrected: parseMaybeInt64(rec[6]),
-			ECCCorrected:   parseMaybeInt64(rec[7]),
-			HWSlowdown:     parseMaybeBool(rec[8]),
+			BDF:                bdf,
+			Serial:             strings.TrimSpace(rec[2]),
+			VBIOS:              strings.TrimSpace(rec[3]),
+			TemperatureC:       parseMaybeFloat(rec[4]),
+			PowerW:             parseMaybeFloat(rec[5]),
+			ECCUncorrected:     parseMaybeInt64(rec[6]),
+			ECCCorrected:       parseMaybeInt64(rec[7]),
+			HWSlowdown:         parseMaybeBool(rec[8]),
+			PCIeLinkGenCurrent: parseMaybeInt(rec[9]),
+			PCIeLinkGenMax:     parseMaybeInt(rec[10]),
+			PCIeLinkWidthCur:   parseMaybeInt(rec[11]),
+			PCIeLinkWidthMax:   parseMaybeInt(rec[12]),
 		}
 		result[bdf] = info
 	}
@@ -167,6 +175,22 @@ func parseMaybeInt64(v string) *int64 {
 	return &n
 }
 
+func parseMaybeInt(v string) *int {
+	v = strings.TrimSpace(v)
+	if v == "" || strings.EqualFold(v, "n/a") || strings.EqualFold(v, "not supported") || strings.EqualFold(v, "[not supported]") {
+		return nil
+	}
+	n, err := strconv.Atoi(v)
+	if err != nil {
+		return nil
+	}
+	return &n
+}
+
+func pcieLinkGenLabel(gen int) string {
+	return fmt.Sprintf("Gen%d", gen)
+}
+
 func parseMaybeBool(v string) *bool {
 	v = strings.TrimSpace(strings.ToLower(v))
 	switch v {
@@ -231,4 +255,22 @@ func injectNVIDIATelemetry(dev *schema.HardwarePCIeDevice, info nvidiaGPUInfo) {
 	if info.HWSlowdown != nil {
 		dev.HWSlowdown = info.HWSlowdown
 	}
+	// Override PCIe link speed/width with nvidia-smi driver values.
+	// sysfs current_link_speed reflects the instantaneous physical link state and
+	// can show Gen1 when the GPU is idle due to ASPM power management. The driver
+	// knows the negotiated speed regardless of the current power state.
+	if info.PCIeLinkGenCurrent != nil {
+		s := pcieLinkGenLabel(*info.PCIeLinkGenCurrent)
+		dev.LinkSpeed = &s
+	}
+	if info.PCIeLinkGenMax != nil {
+		s := pcieLinkGenLabel(*info.PCIeLinkGenMax)
+		dev.MaxLinkSpeed = &s
+	}
+	if info.PCIeLinkWidthCur != nil {
+		dev.LinkWidth = info.PCIeLinkWidthCur
+	}
+	if info.PCIeLinkWidthMax != nil {
+		dev.MaxLinkWidth = info.PCIeLinkWidthMax
+	}
 }

audit/internal/collector/nvidia_test.go

@@ -6,7 +6,7 @@ import (
 )
 
 func TestParseNVIDIASMIQuery(t *testing.T) {
-	raw := "0, 00000000:65:00.0, GPU-SERIAL-1, 96.00.1F.00.02, 54, 210.33, 0, 5, Not Active\n"
+	raw := "0, 00000000:65:00.0, GPU-SERIAL-1, 96.00.1F.00.02, 54, 210.33, 0, 5, Not Active, 4, 4, 16, 16\n"
 	byBDF, err := parseNVIDIASMIQuery(raw)
 	if err != nil {
 		t.Fatalf("parse failed: %v", err)
@@ -28,6 +28,12 @@ func TestParseNVIDIASMIQuery(t *testing.T) {
 	if gpu.HWSlowdown == nil || *gpu.HWSlowdown {
 		t.Fatalf("hw slowdown: got %v, want false", gpu.HWSlowdown)
 	}
+	if gpu.PCIeLinkGenCurrent == nil || *gpu.PCIeLinkGenCurrent != 4 {
+		t.Fatalf("pcie link gen current: got %v, want 4", gpu.PCIeLinkGenCurrent)
+	}
+	if gpu.PCIeLinkGenMax == nil || *gpu.PCIeLinkGenMax != 4 {
+		t.Fatalf("pcie link gen max: got %v, want 4", gpu.PCIeLinkGenMax)
+	}
 }
 
 func TestNormalizePCIeBDF(t *testing.T) {

audit/internal/collector/pcie.go

@@ -59,6 +59,7 @@ func shouldIncludePCIeDevice(class, vendor, device string) bool {
 		"host bridge",
 		"isa bridge",
 		"pci bridge",
+		"co-processor",
 		"performance counter",
 		"performance counters",
 		"ram memory",

audit/internal/collector/pcie_filter_test.go

@@ -19,6 +19,7 @@ func TestShouldIncludePCIeDevice(t *testing.T) {
 		{name: "audio", class: "Audio device", want: false},
 		{name: "host bridge", class: "Host bridge", want: false},
 		{name: "pci bridge", class: "PCI bridge", want: false},
+		{name: "co-processor", class: "Co-processor", want: false},
 		{name: "smbus", class: "SMBus", want: false},
 		{name: "perf", class: "Performance counters", want: false},
 		{name: "non essential instrumentation", class: "Non-Essential Instrumentation", want: false},
@@ -76,6 +77,20 @@ func TestParseLspci_filtersAMDChipsetNoise(t *testing.T) {
 	}
 }
 
+func TestParseLspci_filtersCoProcessors(t *testing.T) {
+	input := "" +
+		"Slot:\t0000:01:00.0\nClass:\tCo-processor\nVendor:\tIntel Corporation\nDevice:\t402xx Series QAT\n\n" +
+		"Slot:\t0000:65:00.0\nClass:\tVGA compatible controller\nVendor:\tNVIDIA Corporation\nDevice:\tH100\n\n"
+
+	devs := parseLspci(input)
+	if len(devs) != 1 {
+		t.Fatalf("expected 1 remaining device, got %d", len(devs))
+	}
+	if devs[0].Model == nil || *devs[0].Model != "H100" {
+		t.Fatalf("unexpected remaining device: %+v", devs[0])
+	}
+}
+
 func TestPCIeJSONUsesSlotNotBDF(t *testing.T) {
 	input := "Slot:\t0000:65:00.0\nClass:\tVGA compatible controller\nVendor:\tNVIDIA Corporation\nDevice:\tH100\n\n"
 

audit/internal/collector/storage.go

@@ -77,11 +77,28 @@ func discoverStorageDevices() []lsblkDevice {
 		if dev.Type != "disk" {
 			continue
 		}
+		if isVirtualBMCDisk(dev) {
+			slog.Debug("storage: skipping BMC virtual disk", "name", dev.Name, "model", dev.Model)
+			continue
+		}
 		disks = append(disks, dev)
 	}
 	return disks
 }
 
+// isVirtualBMCDisk returns true for BMC/IPMI virtual USB mass storage devices
+// that appear as disks but are not real hardware (e.g. iDRAC Virtual HDisk*).
+// These have zero reported size, a generic fake serial, and a model name that
+// starts with "Virtual HDisk".
+func isVirtualBMCDisk(dev lsblkDevice) bool {
+	return isVirtualHDiskModel(dev.Model)
+}
+
+func isVirtualHDiskModel(model string) bool {
+	model = strings.ToLower(strings.TrimSpace(model))
+	return strings.HasPrefix(model, "virtual hdisk")
+}
+
 func lsblkDevices() []lsblkDevice {
 	out, err := exec.Command("lsblk", "-J", "-d",
 		"-o", "NAME,TYPE,SIZE,SERIAL,MODEL,TRAN,HCTL").Output()

audit/internal/platform/benchmark_report.go

@@ -0,0 +1,141 @@
+package platform
+
+import (
+	"fmt"
+	"strings"
+	"time"
+)
+
+func renderBenchmarkReport(result NvidiaBenchmarkResult) string {
+	var b strings.Builder
+	fmt.Fprintf(&b, "Bee NVIDIA Benchmark Report\n")
+	fmt.Fprintf(&b, "===========================\n\n")
+	fmt.Fprintf(&b, "Generated: %s\n", result.GeneratedAt.Format("2006-01-02 15:04:05 UTC"))
+	fmt.Fprintf(&b, "Host: %s\n", result.Hostname)
+	fmt.Fprintf(&b, "Profile: %s\n", result.BenchmarkProfile)
+	fmt.Fprintf(&b, "Overall status: %s\n", result.OverallStatus)
+	fmt.Fprintf(&b, "Selected GPUs: %s\n", joinIndexList(result.SelectedGPUIndices))
+	fmt.Fprintf(&b, "Normalization: %s\n\n", result.Normalization.Status)
+
+	if len(result.Findings) > 0 {
+		fmt.Fprintf(&b, "Executive Summary\n")
+		fmt.Fprintf(&b, "-----------------\n")
+		for _, finding := range result.Findings {
+			fmt.Fprintf(&b, "- %s\n", finding)
+		}
+		b.WriteString("\n")
+	}
+
+	if len(result.Warnings) > 0 {
+		fmt.Fprintf(&b, "Warnings\n")
+		fmt.Fprintf(&b, "--------\n")
+		for _, warning := range result.Warnings {
+			fmt.Fprintf(&b, "- %s\n", warning)
+		}
+		b.WriteString("\n")
+	}
+
+	fmt.Fprintf(&b, "Per GPU Scorecard\n")
+	fmt.Fprintf(&b, "-----------------\n")
+	for _, gpu := range result.GPUs {
+		fmt.Fprintf(&b, "GPU %d  %s\n", gpu.Index, gpu.Name)
+		fmt.Fprintf(&b, "  Status: %s\n", gpu.Status)
+		fmt.Fprintf(&b, "  Composite score: %.2f\n", gpu.Scores.CompositeScore)
+		fmt.Fprintf(&b, "  Compute score: %.2f\n", gpu.Scores.ComputeScore)
+		fmt.Fprintf(&b, "  Power sustain: %.1f\n", gpu.Scores.PowerSustainScore)
+		fmt.Fprintf(&b, "  Thermal sustain: %.1f\n", gpu.Scores.ThermalSustainScore)
+		fmt.Fprintf(&b, "  Stability: %.1f\n", gpu.Scores.StabilityScore)
+		if gpu.Scores.InterconnectScore > 0 {
+			fmt.Fprintf(&b, "  Interconnect: %.1f\n", gpu.Scores.InterconnectScore)
+		}
+		if len(gpu.DegradationReasons) > 0 {
+			fmt.Fprintf(&b, "  Degradation reasons: %s\n", strings.Join(gpu.DegradationReasons, ", "))
+		}
+		fmt.Fprintf(&b, "  Avg power/temp/clock: %.1f W / %.1f C / %.0f MHz\n", gpu.Steady.AvgPowerW, gpu.Steady.AvgTempC, gpu.Steady.AvgGraphicsClockMHz)
+		fmt.Fprintf(&b, "  P95 power/temp/clock: %.1f W / %.1f C / %.0f MHz\n", gpu.Steady.P95PowerW, gpu.Steady.P95TempC, gpu.Steady.P95GraphicsClockMHz)
+		if len(gpu.PrecisionResults) > 0 {
+			fmt.Fprintf(&b, "  Precision results:\n")
+			for _, precision := range gpu.PrecisionResults {
+				if precision.Supported {
+					fmt.Fprintf(&b, "    - %s: %.2f TOPS lanes=%d iterations=%d\n", precision.Name, precision.TeraOpsPerSec, precision.Lanes, precision.Iterations)
+				} else {
+					fmt.Fprintf(&b, "    - %s: unsupported (%s)\n", precision.Name, precision.Notes)
+				}
+			}
+		}
+		fmt.Fprintf(&b, "  Throttle counters (us): sw_power=%d sw_thermal=%d sync_boost=%d hw_thermal=%d hw_power_brake=%d\n",
+			gpu.Throttle.SWPowerCapUS,
+			gpu.Throttle.SWThermalSlowdownUS,
+			gpu.Throttle.SyncBoostUS,
+			gpu.Throttle.HWThermalSlowdownUS,
+			gpu.Throttle.HWPowerBrakeSlowdownUS,
+		)
+		if len(gpu.Notes) > 0 {
+			fmt.Fprintf(&b, "  Notes:\n")
+			for _, note := range gpu.Notes {
+				fmt.Fprintf(&b, "    - %s\n", note)
+			}
+		}
+		b.WriteString("\n")
+	}
+
+	if result.Interconnect != nil {
+		fmt.Fprintf(&b, "Interconnect\n")
+		fmt.Fprintf(&b, "------------\n")
+		fmt.Fprintf(&b, "Status: %s\n", result.Interconnect.Status)
+		if result.Interconnect.Supported {
+			fmt.Fprintf(&b, "Avg algbw / busbw: %.1f / %.1f GB/s\n", result.Interconnect.AvgAlgBWGBps, result.Interconnect.AvgBusBWGBps)
+			fmt.Fprintf(&b, "Max algbw / busbw: %.1f / %.1f GB/s\n", result.Interconnect.MaxAlgBWGBps, result.Interconnect.MaxBusBWGBps)
+		}
+		for _, note := range result.Interconnect.Notes {
+			fmt.Fprintf(&b, "- %s\n", note)
+		}
+		b.WriteString("\n")
+	}
+
+	fmt.Fprintf(&b, "Methodology\n")
+	fmt.Fprintf(&b, "-----------\n")
+	fmt.Fprintf(&b, "- Profile %s uses standardized baseline, warmup, steady-state, interconnect, and cooldown phases.\n", result.BenchmarkProfile)
+	fmt.Fprintf(&b, "- Single-GPU compute score comes from bee-gpu-burn cuBLASLt output when available.\n")
+	fmt.Fprintf(&b, "- Thermal and power limitations are inferred from NVIDIA clock event reason counters and sustained telemetry.\n")
+	fmt.Fprintf(&b, "- result.json is the canonical machine-readable source for this benchmark run.\n\n")
+
+	fmt.Fprintf(&b, "Raw Files\n")
+	fmt.Fprintf(&b, "---------\n")
+	fmt.Fprintf(&b, "- result.json\n")
+	fmt.Fprintf(&b, "- report.txt\n")
+	fmt.Fprintf(&b, "- summary.txt\n")
+	fmt.Fprintf(&b, "- verbose.log\n")
+	fmt.Fprintf(&b, "- gpu-*-baseline-metrics.csv/html/term.txt\n")
+	fmt.Fprintf(&b, "- gpu-*-warmup.log\n")
+	fmt.Fprintf(&b, "- gpu-*-steady.log\n")
+	fmt.Fprintf(&b, "- gpu-*-steady-metrics.csv/html/term.txt\n")
+	fmt.Fprintf(&b, "- gpu-*-cooldown-metrics.csv/html/term.txt\n")
+	if result.Interconnect != nil {
+		fmt.Fprintf(&b, "- nccl-all-reduce.log\n")
+	}
+	return b.String()
+}
+
+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 +1,132 @@
+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 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.",
+		"Composite score: 1176.00",
+		"fp16_tensor: 700.00 TOPS",
+	} {
+		if !strings.Contains(report, needle) {
+			t.Fatalf("report missing %q\n%s", needle, report)
+		}
+	}
+}

audit/internal/platform/benchmark_types.go

@@ -0,0 +1,132 @@
+package platform
+
+import "time"
+
+const (
+	NvidiaBenchmarkProfileStandard  = "standard"
+	NvidiaBenchmarkProfileStability = "stability"
+	NvidiaBenchmarkProfileOvernight = "overnight"
+)
+
+type NvidiaBenchmarkOptions struct {
+	Profile           string
+	SizeMB            int
+	GPUIndices        []int
+	ExcludeGPUIndices []int
+	RunNCCL           bool
+}
+
+type NvidiaBenchmarkResult struct {
+	BenchmarkVersion   string                       `json:"benchmark_version"`
+	GeneratedAt        time.Time                    `json:"generated_at"`
+	Hostname           string                       `json:"hostname,omitempty"`
+	BenchmarkProfile   string                       `json:"benchmark_profile"`
+	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"`
+}
+
+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"`
+	MaxGraphicsClockMHz    float64                    `json:"max_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"`
+}
+
+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"}

audit/internal/platform/install.go

@@ -11,10 +11,10 @@ import (
 
 // InstallDisk describes a candidate disk for installation.
 type InstallDisk struct {
-	Device      string   // e.g. /dev/sda
-	Model       string
-	Size        string   // human-readable, e.g. "500G"
-	SizeBytes   int64    // raw byte count from lsblk
+	Device       string // e.g. /dev/sda
+	Model        string
+	Size         string   // human-readable, e.g. "500G"
+	SizeBytes    int64    // raw byte count from lsblk
 	MountedParts []string // partition mount points currently active
 }
 
@@ -117,6 +117,61 @@ func findLiveBootDevice() string {
 	return "/dev/" + strings.TrimSpace(string(out2))
 }
 
+func mountSource(target string) string {
+	out, err := exec.Command("findmnt", "-n", "-o", "SOURCE", target).Output()
+	if err != nil {
+		return ""
+	}
+	return strings.TrimSpace(string(out))
+}
+
+func mountFSType(target string) string {
+	out, err := exec.Command("findmnt", "-n", "-o", "FSTYPE", target).Output()
+	if err != nil {
+		return ""
+	}
+	return strings.TrimSpace(string(out))
+}
+
+func blockDeviceType(device string) string {
+	if strings.TrimSpace(device) == "" {
+		return ""
+	}
+	out, err := exec.Command("lsblk", "-dn", "-o", "TYPE", device).Output()
+	if err != nil {
+		return ""
+	}
+	return strings.TrimSpace(string(out))
+}
+
+func blockDeviceTransport(device string) string {
+	if strings.TrimSpace(device) == "" {
+		return ""
+	}
+	out, err := exec.Command("lsblk", "-dn", "-o", "TRAN", device).Output()
+	if err != nil {
+		return ""
+	}
+	return strings.TrimSpace(string(out))
+}
+
+func inferLiveBootKind(fsType, source, deviceType, transport string) string {
+	switch {
+	case strings.EqualFold(strings.TrimSpace(fsType), "tmpfs"):
+		return "ram"
+	case strings.EqualFold(strings.TrimSpace(deviceType), "rom"):
+		return "cdrom"
+	case strings.EqualFold(strings.TrimSpace(transport), "usb"):
+		return "usb"
+	case strings.HasPrefix(strings.TrimSpace(source), "/dev/sr"):
+		return "cdrom"
+	case strings.HasPrefix(strings.TrimSpace(source), "/dev/"):
+		return "disk"
+	default:
+		return "unknown"
+	}
+}
+
 // MinInstallBytes returns the minimum recommended disk size for installation:
 // squashfs size × 1.5 to allow for extracted filesystem and bootloader.
 // Returns 0 if the squashfs is not available (non-live environment).

audit/internal/platform/install_to_ram.go

@@ -12,11 +12,40 @@ import (
 )
 
 func (s *System) IsLiveMediaInRAM() bool {
-	out, err := exec.Command("findmnt", "-n", "-o", "FSTYPE", "/run/live/medium").Output()
-	if err != nil {
+	fsType := mountFSType("/run/live/medium")
+	if fsType == "" {
 		return toramActive()
 	}
-	return strings.TrimSpace(string(out)) == "tmpfs"
+	return strings.EqualFold(fsType, "tmpfs")
+}
+
+func (s *System) LiveBootSource() LiveBootSource {
+	fsType := mountFSType("/run/live/medium")
+	source := mountSource("/run/live/medium")
+	device := findLiveBootDevice()
+	status := LiveBootSource{
+		InRAM:  strings.EqualFold(fsType, "tmpfs"),
+		Source: source,
+		Device: device,
+	}
+	if fsType == "" && source == "" && device == "" {
+		if toramActive() {
+			status.InRAM = true
+			status.Kind = "ram"
+			status.Source = "tmpfs"
+			return status
+		}
+		status.Kind = "unknown"
+		return status
+	}
+	status.Kind = inferLiveBootKind(fsType, source, blockDeviceType(device), blockDeviceTransport(device))
+	if status.Kind == "" {
+		status.Kind = "unknown"
+	}
+	if status.InRAM && strings.TrimSpace(status.Source) == "" {
+		status.Source = "tmpfs"
+	}
+	return status
 }
 
 func (s *System) RunInstallToRAM(ctx context.Context, logFunc func(string)) error {

audit/internal/platform/install_to_ram_test.go

@@ -0,0 +1,28 @@
+package platform
+
+import "testing"
+
+func TestInferLiveBootKind(t *testing.T) {
+	tests := []struct {
+		name       string
+		fsType     string
+		source     string
+		deviceType string
+		transport  string
+		want       string
+	}{
+		{name: "ram tmpfs", fsType: "tmpfs", source: "/dev/shm/bee-live", want: "ram"},
+		{name: "usb disk", source: "/dev/sdb1", deviceType: "disk", transport: "usb", want: "usb"},
+		{name: "cdrom rom", source: "/dev/sr0", deviceType: "rom", want: "cdrom"},
+		{name: "disk sata", source: "/dev/nvme0n1p1", deviceType: "disk", transport: "nvme", want: "disk"},
+		{name: "unknown", source: "overlay", want: "unknown"},
+	}
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			got := inferLiveBootKind(tc.fsType, tc.source, tc.deviceType, tc.transport)
+			if got != tc.want {
+				t.Fatalf("inferLiveBootKind(%q,%q,%q,%q)=%q want %q", tc.fsType, tc.source, tc.deviceType, tc.transport, got, tc.want)
+			}
+		})
+	}
+}

audit/internal/platform/kill_workers.go

@@ -0,0 +1,64 @@
+package platform
+
+import (
+	"fmt"
+	"os"
+	"strconv"
+	"strings"
+	"syscall"
+)
+
+// workerPatterns are substrings matched against /proc/<pid>/cmdline to identify
+// bee test worker processes that should be killed by KillTestWorkers.
+var workerPatterns = []string{
+	"bee-gpu-burn",
+	"stress-ng",
+	"stressapptest",
+	"memtester",
+}
+
+// KilledProcess describes a process that was sent SIGKILL.
+type KilledProcess struct {
+	PID  int    `json:"pid"`
+	Name string `json:"name"`
+}
+
+// KillTestWorkers scans /proc for running test worker processes and sends
+// SIGKILL to each one found. It returns a list of killed processes.
+// Errors for individual processes (e.g. already exited) are silently ignored.
+func KillTestWorkers() []KilledProcess {
+	entries, err := os.ReadDir("/proc")
+	if err != nil {
+		return nil
+	}
+
+	var killed []KilledProcess
+	for _, e := range entries {
+		if !e.IsDir() {
+			continue
+		}
+		pid, err := strconv.Atoi(e.Name())
+		if err != nil {
+			continue
+		}
+		cmdline, err := os.ReadFile(fmt.Sprintf("/proc/%d/cmdline", pid))
+		if err != nil {
+			continue
+		}
+		// /proc/*/cmdline uses NUL bytes as argument separators.
+		args := strings.SplitN(strings.ReplaceAll(string(cmdline), "\x00", " "), " ", 2)
+		exe := strings.TrimSpace(args[0])
+		base := exe
+		if idx := strings.LastIndexByte(exe, '/'); idx >= 0 {
+			base = exe[idx+1:]
+		}
+		for _, pat := range workerPatterns {
+			if strings.Contains(base, pat) || strings.Contains(exe, pat) {
+				_ = syscall.Kill(pid, syscall.SIGKILL)
+				killed = append(killed, KilledProcess{PID: pid, Name: base})
+				break
+			}
+		}
+	}
+	return killed
+}

audit/internal/platform/live_metrics.go

@@ -68,18 +68,20 @@ func SampleLiveMetrics() LiveMetricSample {
 
 // sampleCPULoadPct reads two /proc/stat snapshots 200ms apart and returns
 // the overall CPU utilisation percentage.
-var cpuStatPrev [2]uint64 // [total, idle]
-
 func sampleCPULoadPct() float64 {
-	total, idle := readCPUStat()
-	if total == 0 {
+	total0, idle0 := readCPUStat()
+	if total0 == 0 {
 		return 0
 	}
-	prevTotal, prevIdle := cpuStatPrev[0], cpuStatPrev[1]
-	cpuStatPrev = [2]uint64{total, idle}
-	if prevTotal == 0 {
+	time.Sleep(200 * time.Millisecond)
+	total1, idle1 := readCPUStat()
+	if total1 == 0 {
 		return 0
 	}
+	return cpuLoadPctBetween(total0, idle0, total1, idle1)
+}
+
+func cpuLoadPctBetween(prevTotal, prevIdle, total, idle uint64) float64 {
 	dt := float64(total - prevTotal)
 	di := float64(idle - prevIdle)
 	if dt <= 0 {

audit/internal/platform/live_metrics_test.go

@@ -42,3 +42,53 @@ func TestCompactAmbientTempName(t *testing.T) {
 		t.Fatalf("got %q", got)
 	}
 }
+
+func TestCPULoadPctBetween(t *testing.T) {
+	tests := []struct {
+		name      string
+		prevTotal uint64
+		prevIdle  uint64
+		total     uint64
+		idle      uint64
+		want      float64
+	}{
+		{
+			name:      "busy half",
+			prevTotal: 100,
+			prevIdle:  40,
+			total:     200,
+			idle:      90,
+			want:      50,
+		},
+		{
+			name:      "fully busy",
+			prevTotal: 100,
+			prevIdle:  40,
+			total:     200,
+			idle:      40,
+			want:      100,
+		},
+		{
+			name:      "no progress",
+			prevTotal: 100,
+			prevIdle:  40,
+			total:     100,
+			idle:      40,
+			want:      0,
+		},
+		{
+			name:      "idle delta larger than total clamps to zero",
+			prevTotal: 100,
+			prevIdle:  40,
+			total:     200,
+			idle:      150,
+			want:      0,
+		},
+	}
+
+	for _, tc := range tests {
+		if got := cpuLoadPctBetween(tc.prevTotal, tc.prevIdle, tc.total, tc.idle); got != tc.want {
+			t.Fatalf("%s: cpuLoadPctBetween(...)=%v want %v", tc.name, got, tc.want)
+		}
+	}
+}

audit/internal/platform/nvidia_stress.go

@@ -0,0 +1,203 @@
+package platform
+
+import (
+	"context"
+	"fmt"
+	"sort"
+	"strconv"
+	"strings"
+)
+
+func (s *System) RunNvidiaStressPack(ctx context.Context, baseDir string, opts NvidiaStressOptions, logFunc func(string)) (string, error) {
+	normalizeNvidiaStressOptions(&opts)
+
+	job, err := buildNvidiaStressJob(opts)
+	if err != nil {
+		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"}},
+		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)
+}
+
+func nvidiaStressArchivePrefix(loader string) string {
+	switch strings.TrimSpace(strings.ToLower(loader)) {
+	case NvidiaStressLoaderJohn:
+		return "gpu-nvidia-john"
+	case NvidiaStressLoaderNCCL:
+		return "gpu-nvidia-nccl"
+	default:
+		return "gpu-nvidia-burn"
+	}
+}
+
+func buildNvidiaStressJob(opts NvidiaStressOptions) (satJob, error) {
+	selected, err := resolveNvidiaGPUSelection(opts.GPUIndices, opts.ExcludeGPUIndices)
+	if err != nil {
+		return satJob{}, err
+	}
+
+	loader := strings.TrimSpace(strings.ToLower(opts.Loader))
+	switch loader {
+	case "", NvidiaStressLoaderBuiltin:
+		cmd := []string{
+			"bee-gpu-burn",
+			"--seconds", strconv.Itoa(opts.DurationSec),
+			"--size-mb", strconv.Itoa(opts.SizeMB),
+		}
+		if len(selected) > 0 {
+			cmd = append(cmd, "--devices", joinIndexList(selected))
+		}
+		return satJob{
+			name:       "03-bee-gpu-burn.log",
+			cmd:        cmd,
+			collectGPU: true,
+			gpuIndices: selected,
+		}, nil
+	case NvidiaStressLoaderJohn:
+		cmd := []string{
+			"bee-john-gpu-stress",
+			"--seconds", strconv.Itoa(opts.DurationSec),
+		}
+		if len(selected) > 0 {
+			cmd = append(cmd, "--devices", joinIndexList(selected))
+		}
+		return satJob{
+			name:       "03-john-gpu-stress.log",
+			cmd:        cmd,
+			collectGPU: true,
+			gpuIndices: selected,
+		}, nil
+	case NvidiaStressLoaderNCCL:
+		cmd := []string{
+			"bee-nccl-gpu-stress",
+			"--seconds", strconv.Itoa(opts.DurationSec),
+		}
+		if len(selected) > 0 {
+			cmd = append(cmd, "--devices", joinIndexList(selected))
+		}
+		return satJob{
+			name:       "03-bee-nccl-gpu-stress.log",
+			cmd:        cmd,
+			collectGPU: true,
+			gpuIndices: selected,
+		}, nil
+	default:
+		return satJob{}, fmt.Errorf("unknown NVIDIA stress loader %q", opts.Loader)
+	}
+}
+
+func normalizeNvidiaStressOptions(opts *NvidiaStressOptions) {
+	if opts.DurationSec <= 0 {
+		opts.DurationSec = 300
+	}
+	// SizeMB=0 means "auto" — bee-gpu-burn will query per-GPU memory at runtime.
+	switch strings.TrimSpace(strings.ToLower(opts.Loader)) {
+	case "", NvidiaStressLoaderBuiltin:
+		opts.Loader = NvidiaStressLoaderBuiltin
+	case NvidiaStressLoaderJohn:
+		opts.Loader = NvidiaStressLoaderJohn
+	case NvidiaStressLoaderNCCL:
+		opts.Loader = NvidiaStressLoaderNCCL
+	default:
+		opts.Loader = NvidiaStressLoaderBuiltin
+	}
+	opts.GPUIndices = dedupeSortedIndices(opts.GPUIndices)
+	opts.ExcludeGPUIndices = dedupeSortedIndices(opts.ExcludeGPUIndices)
+}
+
+func resolveNvidiaGPUSelection(include, exclude []int) ([]int, error) {
+	all, err := listNvidiaGPUIndices()
+	if err != nil {
+		return nil, err
+	}
+	if len(all) == 0 {
+		return nil, fmt.Errorf("nvidia-smi found no NVIDIA GPUs")
+	}
+
+	selected := all
+	if len(include) > 0 {
+		want := make(map[int]struct{}, len(include))
+		for _, idx := range include {
+			want[idx] = struct{}{}
+		}
+		selected = selected[:0]
+		for _, idx := range all {
+			if _, ok := want[idx]; ok {
+				selected = append(selected, idx)
+			}
+		}
+	}
+	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 after applying filters")
+	}
+	out := append([]int(nil), selected...)
+	sort.Ints(out)
+	return out, nil
+}
+
+func listNvidiaGPUIndices() ([]int, error) {
+	out, err := satExecCommand("nvidia-smi", "--query-gpu=index", "--format=csv,noheader,nounits").Output()
+	if err != nil {
+		return nil, fmt.Errorf("nvidia-smi: %w", err)
+	}
+	var indices []int
+	for _, line := range strings.Split(strings.TrimSpace(string(out)), "\n") {
+		line = strings.TrimSpace(line)
+		if line == "" {
+			continue
+		}
+		idx, err := strconv.Atoi(line)
+		if err != nil {
+			continue
+		}
+		indices = append(indices, idx)
+	}
+	return dedupeSortedIndices(indices), nil
+}
+
+func dedupeSortedIndices(values []int) []int {
+	if len(values) == 0 {
+		return nil
+	}
+	seen := make(map[int]struct{}, len(values))
+	out := make([]int, 0, len(values))
+	for _, value := range values {
+		if value < 0 {
+			continue
+		}
+		if _, ok := seen[value]; ok {
+			continue
+		}
+		seen[value] = struct{}{}
+		out = append(out, value)
+	}
+	sort.Ints(out)
+	return out
+}
+
+func joinIndexList(values []int) string {
+	parts := make([]string, 0, len(values))
+	for _, value := range values {
+		parts = append(parts, strconv.Itoa(value))
+	}
+	return strings.Join(parts, ",")
+}

audit/internal/platform/platform_stress.go

@@ -10,9 +10,11 @@ import (
 	"os"
 	"os/exec"
 	"path/filepath"
+	"runtime"
 	"strconv"
 	"strings"
 	"sync"
+	"syscall"
 	"time"
 )
 
@@ -24,7 +26,8 @@ type PlatformStressCycle struct {
 
 // PlatformStressOptions controls the thermal cycling test.
 type PlatformStressOptions struct {
-	Cycles []PlatformStressCycle
+	Cycles     []PlatformStressCycle
+	Components []string // if empty: run all; values: "cpu", "gpu"
 }
 
 // platformStressRow is one second of telemetry.
@@ -66,8 +69,11 @@ func (s *System) RunPlatformStress(
 		return "", fmt.Errorf("mkdir run dir: %w", err)
 	}
 
+	hasCPU := len(opts.Components) == 0 || containsComponent(opts.Components, "cpu")
+	hasGPU := len(opts.Components) == 0 || containsComponent(opts.Components, "gpu")
+
 	vendor := s.DetectGPUVendor()
-	logFunc(fmt.Sprintf("Platform Thermal Cycling — %d cycle(s), GPU vendor: %s", len(opts.Cycles), vendor))
+	logFunc(fmt.Sprintf("Platform Thermal Cycling — %d cycle(s), GPU vendor: %s, cpu=%v gpu=%v", len(opts.Cycles), vendor, hasCPU, hasGPU))
 
 	var rows []platformStressRow
 	start := time.Now()
@@ -86,27 +92,31 @@ func (s *System) RunPlatformStress(
 		var wg sync.WaitGroup
 
 		// CPU stress
-		wg.Add(1)
-		go func() {
-			defer wg.Done()
-			cpuCmd, err := buildCPUStressCmd(loadCtx)
-			if err != nil {
-				logFunc("CPU stress: " + err.Error())
-				return
-			}
-			_ = cpuCmd.Wait() // exits when loadCtx times out (SIGKILL)
-		}()
+		if hasCPU {
+			wg.Add(1)
+			go func() {
+				defer wg.Done()
+				cpuCmd, err := buildCPUStressCmd(loadCtx)
+				if err != nil {
+					logFunc("CPU stress: " + err.Error())
+					return
+				}
+				_ = cpuCmd.Wait() // exits when loadCtx times out (SIGKILL)
+			}()
+		}
 
 		// GPU stress
-		wg.Add(1)
-		go func() {
-			defer wg.Done()
-			gpuCmd := buildGPUStressCmd(loadCtx, vendor)
-			if gpuCmd == nil {
-				return
-			}
-			_ = gpuCmd.Wait()
-		}()
+		if hasGPU {
+			wg.Add(1)
+			go func() {
+				defer wg.Done()
+				gpuCmd := buildGPUStressCmd(loadCtx, vendor)
+				if gpuCmd == nil {
+					return
+				}
+				_ = gpuCmd.Wait()
+			}()
+		}
 
 		// Monitoring goroutine for load phase
 		loadRows := collectPhase(loadCtx, cycleNum, "load", start)
@@ -374,10 +384,17 @@ func buildCPUStressCmd(ctx context.Context) (*exec.Cmd, error) {
 		return nil, fmt.Errorf("stressapptest not found: %w", err)
 	}
 	// Use a very long duration; the context timeout will kill it at the right time.
-	cmd := exec.CommandContext(ctx, path, "-s", "86400", "-W", "--cc_test")
+	cmdArgs := []string{"-s", "86400", "-W", "--cc_test"}
+	if threads := platformStressCPUThreads(); threads > 0 {
+		cmdArgs = append(cmdArgs, "-m", strconv.Itoa(threads))
+	}
+	if mb := platformStressMemoryMB(); mb > 0 {
+		cmdArgs = append(cmdArgs, "-M", strconv.Itoa(mb))
+	}
+	cmd := exec.CommandContext(ctx, path, cmdArgs...)
 	cmd.Stdout = nil
 	cmd.Stderr = nil
-	if err := cmd.Start(); err != nil {
+	if err := startLowPriorityCmd(cmd, 15); err != nil {
 		return nil, fmt.Errorf("stressapptest start: %w", err)
 	}
 	return cmd, nil
@@ -418,22 +435,74 @@ func buildAMDGPUStressCmd(ctx context.Context) *exec.Cmd {
 	cmd := exec.CommandContext(ctx, rvsPath, "-c", cfgFile)
 	cmd.Stdout = nil
 	cmd.Stderr = nil
-	_ = cmd.Start()
+	_ = startLowPriorityCmd(cmd, 10)
 	return cmd
 }
 
 func buildNvidiaGPUStressCmd(ctx context.Context) *exec.Cmd {
-	path, err := satLookPath("bee-gpu-stress")
+	path, err := satLookPath("bee-gpu-burn")
+	if err != nil {
+		path, err = satLookPath("bee-gpu-stress")
+	}
 	if err != nil {
 		return nil
 	}
-	cmd := exec.CommandContext(ctx, path, "--seconds", "86400", "--size-mb", "64")
+	cmd := exec.CommandContext(ctx, path, "--seconds", "86400")
 	cmd.Stdout = nil
 	cmd.Stderr = nil
-	_ = cmd.Start()
+	_ = startLowPriorityCmd(cmd, 10)
 	return cmd
 }
 
+func startLowPriorityCmd(cmd *exec.Cmd, nice int) error {
+	if err := cmd.Start(); err != nil {
+		return err
+	}
+	if cmd.Process != nil {
+		_ = syscall.Setpriority(syscall.PRIO_PROCESS, cmd.Process.Pid, nice)
+	}
+	return nil
+}
+
+func platformStressCPUThreads() int {
+	if n := envInt("BEE_PLATFORM_STRESS_THREADS", 0); n > 0 {
+		return n
+	}
+	cpus := runtime.NumCPU()
+	switch {
+	case cpus <= 2:
+		return 1
+	case cpus <= 8:
+		return cpus - 1
+	default:
+		return cpus - 2
+	}
+}
+
+func platformStressMemoryMB() int {
+	if mb := envInt("BEE_PLATFORM_STRESS_MB", 0); mb > 0 {
+		return mb
+	}
+	free := freeMemBytes()
+	if free <= 0 {
+		return 0
+	}
+	mb := int((free * 60) / 100 / (1024 * 1024))
+	if mb < 1024 {
+		return 1024
+	}
+	return mb
+}
+
+func containsComponent(components []string, name string) bool {
+	for _, c := range components {
+		if c == name {
+			return true
+		}
+	}
+	return false
+}
+
 func packPlatformDir(dir, dest string) error {
 	f, err := os.Create(dest)
 	if err != nil {

audit/internal/platform/platform_stress_test.go

@@ -0,0 +1,34 @@
+package platform
+
+import (
+	"runtime"
+	"testing"
+)
+
+func TestPlatformStressCPUThreadsOverride(t *testing.T) {
+	t.Setenv("BEE_PLATFORM_STRESS_THREADS", "7")
+	if got := platformStressCPUThreads(); got != 7 {
+		t.Fatalf("platformStressCPUThreads=%d want 7", got)
+	}
+}
+
+func TestPlatformStressCPUThreadsDefaultLeavesHeadroom(t *testing.T) {
+	t.Setenv("BEE_PLATFORM_STRESS_THREADS", "")
+	got := platformStressCPUThreads()
+	if got < 1 {
+		t.Fatalf("platformStressCPUThreads=%d want >= 1", got)
+	}
+	if got > runtime.NumCPU() {
+		t.Fatalf("platformStressCPUThreads=%d want <= NumCPU=%d", got, runtime.NumCPU())
+	}
+	if runtime.NumCPU() > 2 && got >= runtime.NumCPU() {
+		t.Fatalf("platformStressCPUThreads=%d want headroom below NumCPU=%d", got, runtime.NumCPU())
+	}
+}
+
+func TestPlatformStressMemoryMBOverride(t *testing.T) {
+	t.Setenv("BEE_PLATFORM_STRESS_MB", "8192")
+	if got := platformStressMemoryMB(); got != 8192 {
+		t.Fatalf("platformStressMemoryMB=%d want 8192", got)
+	}
+}

audit/internal/platform/runtime.go

@@ -136,7 +136,10 @@ func (s *System) runtimeToolStatuses(vendor string) []ToolStatus {
 		tools = append(tools, s.CheckTools([]string{
 			"nvidia-smi",
 			"nvidia-bug-report.sh",
-			"bee-gpu-stress",
+			"bee-gpu-burn",
+			"bee-john-gpu-stress",
+			"bee-nccl-gpu-stress",
+			"all_reduce_perf",
 		})...)
 	case "amd":
 		tool := ToolStatus{Name: "rocm-smi"}
@@ -176,8 +179,8 @@ func (s *System) collectGPURuntimeHealth(vendor string, health *schema.RuntimeHe
 			health.DriverReady = true
 		}
 
-		if lookErr := exec.Command("sh", "-c", "command -v bee-gpu-stress >/dev/null 2>&1").Run(); lookErr == nil {
-			out, err := exec.Command("bee-gpu-stress", "--seconds", "1", "--size-mb", "1").CombinedOutput()
+		if _, lookErr := exec.LookPath("bee-gpu-burn"); lookErr == nil {
+			out, err := exec.Command("bee-gpu-burn", "--seconds", "1", "--size-mb", "1").CombinedOutput()
 			if err == nil {
 				health.CUDAReady = true
 			} else if strings.Contains(strings.ToLower(string(out)), "cuda_error_system_not_ready") {

audit/internal/platform/sat.go

@@ -12,6 +12,7 @@ import (
 	"os"
 	"os/exec"
 	"path/filepath"
+	"syscall"
 	"sort"
 	"strconv"
 	"strings"
@@ -285,7 +286,25 @@ 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 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 (s *System) RunMemoryAcceptancePack(ctx context.Context, baseDir string, logFunc func(string)) (string, error) {
@@ -425,14 +444,12 @@ type satStats struct {
 }
 
 func nvidiaSATJobs() []satJob {
-	seconds := envInt("BEE_GPU_STRESS_SECONDS", 5)
-	sizeMB := envInt("BEE_GPU_STRESS_SIZE_MB", 64)
 	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-stress.log", cmd: []string{"bee-gpu-stress", "--seconds", fmt.Sprintf("%d", seconds), "--size-mb", fmt.Sprintf("%d", sizeMB)}},
+		{name: "05-bee-gpu-burn.log", cmd: []string{"bee-gpu-burn", "--seconds", "5", "--size-mb", "64"}},
 	}
 }
 
@@ -533,6 +550,13 @@ func runSATCommandCtx(ctx context.Context, verboseLog, name string, cmd []string
 	}
 
 	c := exec.CommandContext(ctx, resolvedCmd[0], resolvedCmd[1:]...)
+	c.SysProcAttr = &syscall.SysProcAttr{Setpgid: true}
+	c.Cancel = func() error {
+		if c.Process != nil {
+			_ = syscall.Kill(-c.Process.Pid, syscall.SIGKILL)
+		}
+		return nil
+	}
 	if len(env) > 0 {
 		c.Env = append(os.Environ(), env...)
 	}
@@ -686,7 +710,11 @@ func resolveSATCommand(cmd []string) ([]string, error) {
 	case "rvs":
 		return resolveRVSCommand(cmd[1:]...)
 	}
-	return cmd, nil
+	path, err := satLookPath(cmd[0])
+	if err != nil {
+		return nil, fmt.Errorf("%s not found in PATH: %w", cmd[0], err)
+	}
+	return append([]string{path}, cmd[1:]...), nil
 }
 
 func resolveRVSCommand(args ...string) ([]string, error) {

audit/internal/platform/sat_fan_stress.go

@@ -51,6 +51,18 @@ type FanStressRow struct {
 	SysPowerW    float64 // DCMI system power reading
 }
 
+type cachedPowerReading struct {
+	Value     float64
+	UpdatedAt time.Time
+}
+
+var (
+	systemPowerCacheMu sync.Mutex
+	systemPowerCache   cachedPowerReading
+)
+
+const systemPowerHoldTTL = 15 * time.Second
+
 // RunFanStressTest runs a two-phase GPU stress test while monitoring fan speeds,
 // temperatures, and power draw every second. Exports metrics.csv and fan-sensors.csv.
 // Designed to reproduce case-04 fan-speed lag and detect GPU thermal throttling.
@@ -130,26 +142,21 @@ func (s *System) RunFanStressTest(ctx context.Context, baseDir string, opts FanS
 		stats.OK++
 	}
 
-	// loadPhase runs bee-gpu-stress for durSec; sampler stamps phaseName on each row.
+	// loadPhase runs bee-gpu-burn for durSec; sampler stamps phaseName on each row.
 	loadPhase := func(phaseName, stepName string, durSec int) {
 		if ctx.Err() != nil {
 			return
 		}
 		setPhase(phaseName)
-		var env []string
-		if len(opts.GPUIndices) > 0 {
-			ids := make([]string, len(opts.GPUIndices))
-			for i, idx := range opts.GPUIndices {
-				ids[i] = strconv.Itoa(idx)
-			}
-			env = []string{"CUDA_VISIBLE_DEVICES=" + strings.Join(ids, ",")}
-		}
 		cmd := []string{
-			"bee-gpu-stress",
+			"bee-gpu-burn",
 			"--seconds", strconv.Itoa(durSec),
 			"--size-mb", strconv.Itoa(opts.SizeMB),
 		}
-		out, err := runSATCommandCtx(ctx, verboseLog, stepName, cmd, env, nil)
+		if len(opts.GPUIndices) > 0 {
+			cmd = append(cmd, "--devices", joinIndexList(dedupeSortedIndices(opts.GPUIndices)))
+		}
+		out, err := runSATCommandCtx(ctx, verboseLog, stepName, cmd, nil, nil)
 		_ = os.WriteFile(filepath.Join(runDir, stepName+".log"), out, 0644)
 		if err != nil && err != context.Canceled && err.Error() != "signal: killed" {
 			fmt.Fprintf(&summary, "%s_status=FAILED\n", stepName)
@@ -323,8 +330,9 @@ func sampleFanSpeeds() ([]FanReading, error) {
 
 // parseFanSpeeds parses "ipmitool sdr type Fan" output.
 // Handles two formats:
-//   Old: "FAN1 | 2400.000 | RPM | ok"           (value in col[1], unit in col[2])
-//   New: "FAN1 | 41h | ok | 29.1 | 4340 RPM"   (value+unit combined in last col)
+//
+//	Old: "FAN1 | 2400.000 | RPM | ok"           (value in col[1], unit in col[2])
+//	New: "FAN1 | 41h | ok | 29.1 | 4340 RPM"   (value+unit combined in last col)
 func parseFanSpeeds(raw string) []FanReading {
 	var fans []FanReading
 	for _, line := range strings.Split(strings.TrimSpace(raw), "\n") {
@@ -512,11 +520,17 @@ func sampleCPUTempViaSensors() float64 {
 
 // sampleSystemPower reads system power draw via DCMI.
 func sampleSystemPower() float64 {
+	now := time.Now()
+	current := 0.0
 	out, err := exec.Command("ipmitool", "dcmi", "power", "reading").Output()
-	if err != nil {
-		return 0
+	if err == nil {
+		current = parseDCMIPowerReading(string(out))
 	}
-	return parseDCMIPowerReading(string(out))
+	systemPowerCacheMu.Lock()
+	defer systemPowerCacheMu.Unlock()
+	value, updated := effectiveSystemPowerReading(systemPowerCache, current, now)
+	systemPowerCache = updated
+	return value
 }
 
 // parseDCMIPowerReading extracts the instantaneous power reading from ipmitool dcmi output.
@@ -539,6 +553,17 @@ func parseDCMIPowerReading(raw string) float64 {
 	return 0
 }
 
+func effectiveSystemPowerReading(cache cachedPowerReading, current float64, now time.Time) (float64, cachedPowerReading) {
+	if current > 0 {
+		cache = cachedPowerReading{Value: current, UpdatedAt: now}
+		return current, cache
+	}
+	if cache.Value > 0 && !cache.UpdatedAt.IsZero() && now.Sub(cache.UpdatedAt) <= systemPowerHoldTTL {
+		return cache.Value, cache
+	}
+	return 0, cache
+}
+
 // analyzeThrottling returns true if any GPU reported an active throttle reason
 // during either load phase.
 func analyzeThrottling(rows []FanStressRow) bool {

audit/internal/platform/sat_fan_stress_test.go

@@ -1,6 +1,9 @@
 package platform
 
-import "testing"
+import (
+	"testing"
+	"time"
+)
 
 func TestParseFanSpeeds(t *testing.T) {
 	raw := "FAN1 | 2400.000 | RPM | ok\nFAN2 | 1800 RPM | ok | ok\nFAN3 | na | RPM | ns\n"
@@ -25,3 +28,40 @@ func TestFirstFanInputValue(t *testing.T) {
 		t.Fatalf("got=%v ok=%v", got, ok)
 	}
 }
+
+func TestParseDCMIPowerReading(t *testing.T) {
+	raw := `
+Instantaneous power reading:                   512 Watts
+Minimum during sampling period:               498 Watts
+`
+	if got := parseDCMIPowerReading(raw); got != 512 {
+		t.Fatalf("parseDCMIPowerReading()=%v want 512", got)
+	}
+}
+
+func TestEffectiveSystemPowerReading(t *testing.T) {
+	now := time.Now()
+	cache := cachedPowerReading{Value: 480, UpdatedAt: now.Add(-5 * time.Second)}
+
+	got, updated := effectiveSystemPowerReading(cache, 0, now)
+	if got != 480 {
+		t.Fatalf("got=%v want cached 480", got)
+	}
+	if updated.Value != 480 {
+		t.Fatalf("updated=%+v", updated)
+	}
+
+	got, updated = effectiveSystemPowerReading(cache, 530, now)
+	if got != 530 {
+		t.Fatalf("got=%v want 530", got)
+	}
+	if updated.Value != 530 {
+		t.Fatalf("updated=%+v", updated)
+	}
+
+	expired := cachedPowerReading{Value: 480, UpdatedAt: now.Add(-systemPowerHoldTTL - time.Second)}
+	got, _ = effectiveSystemPowerReading(expired, 0, now)
+	if got != 0 {
+		t.Fatalf("expired cache returned %v want 0", got)
+	}
+}

audit/internal/platform/sat_test.go

@@ -31,8 +31,8 @@ func TestRunNvidiaAcceptancePackIncludesGPUStress(t *testing.T) {
 	if len(jobs) != 5 {
 		t.Fatalf("jobs=%d want 5", len(jobs))
 	}
-	if got := jobs[4].cmd[0]; got != "bee-gpu-stress" {
-		t.Fatalf("gpu stress command=%q want bee-gpu-stress", got)
+	if got := jobs[4].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" {
 		t.Fatalf("bug report flag=%q want --output-file", got)
@@ -80,13 +80,10 @@ func TestAMDStressJobsIncludeBandwidthAndGST(t *testing.T) {
 	}
 }
 
-func TestNvidiaSATJobsUseEnvOverrides(t *testing.T) {
-	t.Setenv("BEE_GPU_STRESS_SECONDS", "9")
-	t.Setenv("BEE_GPU_STRESS_SIZE_MB", "96")
-
+func TestNvidiaSATJobsUseBuiltinBurnDefaults(t *testing.T) {
 	jobs := nvidiaSATJobs()
 	got := jobs[4].cmd
-	want := []string{"bee-gpu-stress", "--seconds", "9", "--size-mb", "96"}
+	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))
 	}
@@ -97,6 +94,126 @@ func TestNvidiaSATJobsUseEnvOverrides(t *testing.T) {
 	}
 }
 
+func TestBuildNvidiaStressJobUsesSelectedLoaderAndDevices(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 '0\n1\n2\n'")
+		}
+		return exec.Command(name, args...)
+	}
+	t.Cleanup(func() { satExecCommand = oldExecCommand })
+
+	job, err := buildNvidiaStressJob(NvidiaStressOptions{
+		DurationSec:       600,
+		Loader:            NvidiaStressLoaderJohn,
+		ExcludeGPUIndices: []int{1},
+	})
+	if err != nil {
+		t.Fatalf("buildNvidiaStressJob error: %v", err)
+	}
+	wantCmd := []string{"bee-john-gpu-stress", "--seconds", "600", "--devices", "0,2"}
+	if len(job.cmd) != len(wantCmd) {
+		t.Fatalf("cmd len=%d want %d (%v)", len(job.cmd), len(wantCmd), job.cmd)
+	}
+	for i := range wantCmd {
+		if job.cmd[i] != wantCmd[i] {
+			t.Fatalf("cmd[%d]=%q want %q", i, job.cmd[i], wantCmd[i])
+		}
+	}
+	if got := joinIndexList(job.gpuIndices); got != "0,2" {
+		t.Fatalf("gpuIndices=%q want 0,2", got)
+	}
+}
+
+func TestBuildNvidiaStressJobUsesNCCLLoader(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 '0\n1\n2\n'")
+		}
+		return exec.Command(name, args...)
+	}
+	t.Cleanup(func() { satExecCommand = oldExecCommand })
+
+	job, err := buildNvidiaStressJob(NvidiaStressOptions{
+		DurationSec: 120,
+		Loader:      NvidiaStressLoaderNCCL,
+		GPUIndices:  []int{2, 0},
+	})
+	if err != nil {
+		t.Fatalf("buildNvidiaStressJob error: %v", err)
+	}
+	wantCmd := []string{"bee-nccl-gpu-stress", "--seconds", "120", "--devices", "0,2"}
+	if len(job.cmd) != len(wantCmd) {
+		t.Fatalf("cmd len=%d want %d (%v)", len(job.cmd), len(wantCmd), job.cmd)
+	}
+	for i := range wantCmd {
+		if job.cmd[i] != wantCmd[i] {
+			t.Fatalf("cmd[%d]=%q want %q", i, job.cmd[i], wantCmd[i])
+		}
+	}
+	if got := joinIndexList(job.gpuIndices); got != "0,2" {
+		t.Fatalf("gpuIndices=%q want 0,2", got)
+	}
+}
+
+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 TestNvidiaStressArchivePrefixByLoader(t *testing.T) {
+	t.Parallel()
+
+	tests := []struct {
+		loader string
+		want   string
+	}{
+		{loader: NvidiaStressLoaderBuiltin, want: "gpu-nvidia-burn"},
+		{loader: NvidiaStressLoaderJohn, want: "gpu-nvidia-john"},
+		{loader: NvidiaStressLoaderNCCL, want: "gpu-nvidia-nccl"},
+		{loader: "", want: "gpu-nvidia-burn"},
+	}
+	for _, tt := range tests {
+		if got := nvidiaStressArchivePrefix(tt.loader); got != tt.want {
+			t.Fatalf("loader=%q prefix=%q want %q", tt.loader, got, tt.want)
+		}
+	}
+}
+
 func TestEnvIntFallback(t *testing.T) {
 	os.Unsetenv("BEE_MEMTESTER_SIZE_MB")
 	if got := envInt("BEE_MEMTESTER_SIZE_MB", 123); got != 123 {
@@ -122,8 +239,8 @@ 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: "failed", job: "bee-gpu-stress", out: "cuda error", err: errors.New("rc 1"), status: "FAILED"},
-		{name: "cuda not ready", job: "bee-gpu-stress", out: "cuInit failed: CUDA_ERROR_SYSTEM_NOT_READY", 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"},
 	}
 
 	for _, tt := range tests {
@@ -172,6 +289,44 @@ func TestResolveROCmSMICommandFromPATH(t *testing.T) {
 	}
 }
 
+func TestResolveSATCommandUsesLookPathForGenericTools(t *testing.T) {
+	oldLookPath := satLookPath
+	satLookPath = func(file string) (string, error) {
+		if file == "stress-ng" {
+			return "/usr/bin/stress-ng", nil
+		}
+		return "", exec.ErrNotFound
+	}
+	t.Cleanup(func() { satLookPath = oldLookPath })
+
+	cmd, err := resolveSATCommand([]string{"stress-ng", "--cpu", "0"})
+	if err != nil {
+		t.Fatalf("resolveSATCommand error: %v", err)
+	}
+	if len(cmd) != 3 {
+		t.Fatalf("cmd len=%d want 3 (%v)", len(cmd), cmd)
+	}
+	if cmd[0] != "/usr/bin/stress-ng" {
+		t.Fatalf("cmd[0]=%q want /usr/bin/stress-ng", cmd[0])
+	}
+}
+
+func TestResolveSATCommandFailsForMissingGenericTool(t *testing.T) {
+	oldLookPath := satLookPath
+	satLookPath = func(file string) (string, error) {
+		return "", exec.ErrNotFound
+	}
+	t.Cleanup(func() { satLookPath = oldLookPath })
+
+	_, err := resolveSATCommand([]string{"stress-ng", "--cpu", "0"})
+	if err == nil {
+		t.Fatal("expected error")
+	}
+	if !strings.Contains(err.Error(), "stress-ng not found in PATH") {
+		t.Fatalf("error=%q", err)
+	}
+}
+
 func TestResolveROCmSMICommandFallsBackToROCmTree(t *testing.T) {
 	tmp := t.TempDir()
 	execPath := filepath.Join(tmp, "opt", "rocm", "bin", "rocm-smi")

audit/internal/platform/types.go

@@ -2,6 +2,13 @@ package platform
 
 type System struct{}
 
+type LiveBootSource struct {
+	InRAM  bool   `json:"in_ram"`
+	Kind   string `json:"kind"`
+	Source string `json:"source,omitempty"`
+	Device string `json:"device,omitempty"`
+}
+
 type InterfaceInfo struct {
 	Name  string
 	State string
@@ -51,6 +58,20 @@ type ToolStatus struct {
 	OK   bool
 }
 
+const (
+	NvidiaStressLoaderBuiltin = "builtin"
+	NvidiaStressLoaderJohn    = "john"
+	NvidiaStressLoaderNCCL    = "nccl"
+)
+
+type NvidiaStressOptions struct {
+	DurationSec       int
+	SizeMB            int
+	Loader            string
+	GPUIndices        []int
+	ExcludeGPUIndices []int
+}
+
 func New() *System {
 	return &System{}
 }

audit/internal/webui/api.go

@@ -2,21 +2,26 @@ package webui
 
 import (
 	"bufio"
-	"context"
 	"encoding/json"
+	"errors"
 	"fmt"
 	"io"
 	"net/http"
+	"os"
 	"os/exec"
 	"path/filepath"
+	"regexp"
 	"strings"
 	"sync/atomic"
+	"syscall"
 	"time"
 
 	"bee/audit/internal/app"
 	"bee/audit/internal/platform"
 )
 
+var ansiEscapeRE = regexp.MustCompile(`\x1b\[[0-9;]*[a-zA-Z]|\x1b[()][A-Z0-9]|\x1b[DABC]`)
+
 // ── Job ID counter ────────────────────────────────────────────────────────────
 
 var jobCounter atomic.Uint64
@@ -58,6 +63,10 @@ func streamJob(w http.ResponseWriter, r *http.Request, j *jobState) {
 	if !sseStart(w) {
 		return
 	}
+	streamSubscribedJob(w, r, j)
+}
+
+func streamSubscribedJob(w http.ResponseWriter, r *http.Request, j *jobState) {
 	existing, ch := j.subscribe()
 	for _, line := range existing {
 		sseWrite(w, "", line)
@@ -81,31 +90,59 @@ func streamJob(w http.ResponseWriter, r *http.Request, j *jobState) {
 	}
 }
 
-// runCmdJob runs an exec.Cmd as a background job, streaming stdout+stderr lines.
-func runCmdJob(j *jobState, cmd *exec.Cmd) {
+// streamCmdJob runs an exec.Cmd and streams stdout+stderr lines into j.
+func streamCmdJob(j *jobState, cmd *exec.Cmd) error {
 	pr, pw := io.Pipe()
 	cmd.Stdout = pw
 	cmd.Stderr = pw
 
 	if err := cmd.Start(); err != nil {
-		j.finish(err.Error())
-		return
+		_ = pw.Close()
+		_ = pr.Close()
+		return err
+	}
+	// Lower the CPU scheduling priority of stress/audit subprocesses to nice+10
+	// so the X server and kernel interrupt handling remain responsive under load
+	// (prevents KVM/IPMI graphical console from freezing during GPU stress tests).
+	if cmd.Process != nil {
+		_ = syscall.Setpriority(syscall.PRIO_PROCESS, cmd.Process.Pid, 10)
 	}
 
+	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() {
-			j.append(scanner.Text())
+			// Split on \r to handle progress-bar style output (e.g. \r overwrites)
+			// and strip ANSI escape codes so logs are readable in the browser.
+			parts := strings.Split(scanner.Text(), "\r")
+			for _, part := range parts {
+				line := ansiEscapeRE.ReplaceAllString(part, "")
+				if line != "" {
+					j.append(line)
+				}
+			}
 		}
+		if err := scanner.Err(); err != nil && !errors.Is(err, io.ErrClosedPipe) {
+			scanDone <- err
+			return
+		}
+		scanDone <- nil
 	}()
 
 	err := cmd.Wait()
 	_ = pw.Close()
+	scanErr := <-scanDone
+	_ = pr.Close()
 	if err != nil {
-		j.finish(err.Error())
-	} else {
-		j.finish("")
+		return err
 	}
+	return scanErr
 }
 
 // ── Audit ─────────────────────────────────────────────────────────────────────
@@ -153,20 +190,23 @@ func (h *handler) handleAPISATRun(target string) http.HandlerFunc {
 		}
 
 		var body struct {
-			Duration    int    `json:"duration"`
-			DiagLevel   int    `json:"diag_level"`
-			GPUIndices  []int  `json:"gpu_indices"`
-			Profile     string `json:"profile"`
-			DisplayName string `json:"display_name"`
+			Duration           int      `json:"duration"`
+			DiagLevel          int      `json:"diag_level"`
+			GPUIndices         []int    `json:"gpu_indices"`
+			ExcludeGPUIndices  []int    `json:"exclude_gpu_indices"`
+			Loader             string   `json:"loader"`
+			Profile            string   `json:"profile"`
+			DisplayName        string   `json:"display_name"`
+			PlatformComponents []string `json:"platform_components"`
 		}
-		if r.ContentLength > 0 {
-			_ = json.NewDecoder(r.Body).Decode(&body)
+		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
+			}
 		}
 
-		name := taskNames[target]
-		if name == "" {
-			name = target
-		}
+		name := taskDisplayName(target, body.Profile, body.Loader)
 		t := &Task{
 			ID:        newJobID("sat-" + target),
 			Name:      name,
@@ -174,11 +214,14 @@ func (h *handler) handleAPISATRun(target string) http.HandlerFunc {
 			Status:    TaskPending,
 			CreatedAt: time.Now(),
 			params: taskParams{
-				Duration:    body.Duration,
-				DiagLevel:   body.DiagLevel,
-				GPUIndices:  body.GPUIndices,
-				BurnProfile: body.Profile,
-				DisplayName: body.DisplayName,
+				Duration:           body.Duration,
+				DiagLevel:          body.DiagLevel,
+				GPUIndices:         body.GPUIndices,
+				ExcludeGPUIndices:  body.ExcludeGPUIndices,
+				Loader:             body.Loader,
+				BurnProfile:        body.Profile,
+				DisplayName:        body.DisplayName,
+				PlatformComponents: body.PlatformComponents,
 			},
 		}
 		if strings.TrimSpace(body.DisplayName) != "" {
@@ -312,8 +355,10 @@ func (h *handler) handleAPINetworkStatus(w http.ResponseWriter, r *http.Request)
 		return
 	}
 	writeJSON(w, map[string]any{
-		"interfaces":    ifaces,
-		"default_route": h.opts.App.DefaultRoute(),
+		"interfaces":     ifaces,
+		"default_route":  h.opts.App.DefaultRoute(),
+		"pending_change": h.hasPendingNetworkChange(),
+		"rollback_in":    h.pendingNetworkRollbackIn(),
 	})
 }
 
@@ -392,17 +437,56 @@ func (h *handler) handleAPIExportList(w http.ResponseWriter, r *http.Request) {
 	writeJSON(w, entries)
 }
 
-func (h *handler) handleAPIExportBundle(w http.ResponseWriter, r *http.Request) {
-	archive, err := app.BuildSupportBundle(h.opts.ExportDir)
+func (h *handler) handleAPIExportUSBTargets(w http.ResponseWriter, _ *http.Request) {
+	if h.opts.App == nil {
+		writeError(w, http.StatusServiceUnavailable, "app not configured")
+		return
+	}
+	targets, err := h.opts.App.ListRemovableTargets()
 	if err != nil {
 		writeError(w, http.StatusInternalServerError, err.Error())
 		return
 	}
-	writeJSON(w, map[string]string{
-		"status": "ok",
-		"path":   archive,
-		"url":    "/export/support.tar.gz",
-	})
+	if targets == nil {
+		targets = []platform.RemovableTarget{}
+	}
+	writeJSON(w, targets)
+}
+
+func (h *handler) handleAPIExportUSBAudit(w http.ResponseWriter, r *http.Request) {
+	if h.opts.App == nil {
+		writeError(w, http.StatusServiceUnavailable, "app not configured")
+		return
+	}
+	var target platform.RemovableTarget
+	if err := json.NewDecoder(r.Body).Decode(&target); err != nil || target.Device == "" {
+		writeError(w, http.StatusBadRequest, "device is required")
+		return
+	}
+	result, err := h.opts.App.ExportLatestAuditResult(target)
+	if err != nil {
+		writeError(w, http.StatusInternalServerError, err.Error())
+		return
+	}
+	writeJSON(w, map[string]string{"status": "ok", "message": result.Body})
+}
+
+func (h *handler) handleAPIExportUSBBundle(w http.ResponseWriter, r *http.Request) {
+	if h.opts.App == nil {
+		writeError(w, http.StatusServiceUnavailable, "app not configured")
+		return
+	}
+	var target platform.RemovableTarget
+	if err := json.NewDecoder(r.Body).Decode(&target); err != nil || target.Device == "" {
+		writeError(w, http.StatusBadRequest, "device is required")
+		return
+	}
+	result, err := h.opts.App.ExportSupportBundleResult(target)
+	if err != nil {
+		writeError(w, http.StatusInternalServerError, err.Error())
+		return
+	}
+	writeJSON(w, map[string]string{"status": "ok", "message": result.Body})
 }
 
 // ── GPU presence ──────────────────────────────────────────────────────────────
@@ -420,6 +504,26 @@ func (h *handler) handleAPIGPUPresence(w http.ResponseWriter, r *http.Request) {
 	})
 }
 
+// ── GPU tools ─────────────────────────────────────────────────────────────────
+
+func (h *handler) handleAPIGPUTools(w http.ResponseWriter, _ *http.Request) {
+	type toolEntry struct {
+		ID        string `json:"id"`
+		Available bool   `json:"available"`
+		Vendor    string `json:"vendor"` // "nvidia" | "amd"
+	}
+	_, nvidiaErr := os.Stat("/dev/nvidia0")
+	_, amdErr := os.Stat("/dev/kfd")
+	nvidiaUp := nvidiaErr == nil
+	amdUp := amdErr == nil
+	writeJSON(w, []toolEntry{
+		{ID: "bee-gpu-burn", Available: nvidiaUp, Vendor: "nvidia"},
+		{ID: "john", Available: nvidiaUp, Vendor: "nvidia"},
+		{ID: "nccl", Available: nvidiaUp, Vendor: "nvidia"},
+		{ID: "rvs", Available: amdUp, Vendor: "amd"},
+	})
+}
+
 // ── System ────────────────────────────────────────────────────────────────────
 
 func (h *handler) handleAPIRAMStatus(w http.ResponseWriter, r *http.Request) {
@@ -427,9 +531,9 @@ func (h *handler) handleAPIRAMStatus(w http.ResponseWriter, r *http.Request) {
 		writeError(w, http.StatusServiceUnavailable, "app not configured")
 		return
 	}
-	inRAM := h.opts.App.IsLiveMediaInRAM()
+	status := h.opts.App.LiveBootSource()
 	w.Header().Set("Content-Type", "application/json")
-	_ = json.NewEncoder(w).Encode(map[string]bool{"in_ram": inRAM})
+	_ = json.NewEncoder(w).Encode(status)
 }
 
 func (h *handler) handleAPIInstallToRAM(w http.ResponseWriter, r *http.Request) {
@@ -437,10 +541,7 @@ func (h *handler) handleAPIInstallToRAM(w http.ResponseWriter, r *http.Request)
 		writeError(w, http.StatusServiceUnavailable, "app not configured")
 		return
 	}
-	h.installMu.Lock()
-	installRunning := h.installJob != nil && !h.installJob.isDone()
-	h.installMu.Unlock()
-	if installRunning {
+	if globalQueue.hasActiveTarget("install") {
 		writeError(w, http.StatusConflict, "install to disk is already running")
 		return
 	}
@@ -555,39 +656,43 @@ func (h *handler) handleAPIInstallRun(w http.ResponseWriter, r *http.Request) {
 		writeError(w, http.StatusConflict, "install to RAM task is already pending or running")
 		return
 	}
-
-	h.installMu.Lock()
-	if h.installJob != nil && !h.installJob.isDone() {
-		h.installMu.Unlock()
-		writeError(w, http.StatusConflict, "install already running")
+	if globalQueue.hasActiveTarget("install") {
+		writeError(w, http.StatusConflict, "install task is already pending or running")
 		return
 	}
-	j := &jobState{}
-	h.installJob = j
-	h.installMu.Unlock()
-
-	logFile := platform.InstallLogPath(req.Device)
-	go runCmdJob(j, exec.CommandContext(context.Background(), "bee-install", req.Device, logFile))
-
-	w.WriteHeader(http.StatusNoContent)
-}
-
-func (h *handler) handleAPIInstallStream(w http.ResponseWriter, r *http.Request) {
-	h.installMu.Lock()
-	j := h.installJob
-	h.installMu.Unlock()
-	if j == nil {
-		if !sseStart(w) {
-			return
-		}
-		sseWrite(w, "done", "")
-		return
+	t := &Task{
+		ID:        newJobID("install"),
+		Name:      "Install to Disk",
+		Target:    "install",
+		Priority:  20,
+		Status:    TaskPending,
+		CreatedAt: time.Now(),
+		params: taskParams{
+			Device: req.Device,
+		},
 	}
-	streamJob(w, r, j)
+	globalQueue.enqueue(t)
+	writeJSON(w, map[string]string{"task_id": t.ID, "job_id": t.ID})
 }
 
 // ── Metrics SSE ───────────────────────────────────────────────────────────────
 
+func (h *handler) handleAPIMetricsLatest(w http.ResponseWriter, r *http.Request) {
+	sample, ok := h.latestMetric()
+	if !ok {
+		w.Header().Set("Content-Type", "application/json")
+		_, _ = w.Write([]byte("{}"))
+		return
+	}
+	b, err := json.Marshal(sample)
+	if err != nil {
+		http.Error(w, err.Error(), http.StatusInternalServerError)
+		return
+	}
+	w.Header().Set("Content-Type", "application/json")
+	_, _ = w.Write(b)
+}
+
 func (h *handler) handleAPIMetricsStream(w http.ResponseWriter, r *http.Request) {
 	if !sseStart(w) {
 		return
@@ -629,13 +734,7 @@ func (h *handler) feedRings(sample platform.LiveMetricSample) {
 	h.ringMemLoad.push(sample.MemLoadPct)
 
 	h.ringsMu.Lock()
-	for i, fan := range sample.Fans {
-		for len(h.ringFans) <= i {
-			h.ringFans = append(h.ringFans, newMetricsRing(120))
-			h.fanNames = append(h.fanNames, fan.Name)
-		}
-		h.ringFans[i].push(float64(fan.RPM))
-	}
+	h.pushFanRings(sample.Fans)
 	for _, gpu := range sample.GPUs {
 		idx := gpu.GPUIndex
 		for len(h.gpuRings) <= idx {
@@ -654,6 +753,51 @@ func (h *handler) feedRings(sample platform.LiveMetricSample) {
 	h.ringsMu.Unlock()
 }
 
+func (h *handler) pushFanRings(fans []platform.FanReading) {
+	if len(fans) == 0 && len(h.ringFans) == 0 {
+		return
+	}
+	fanValues := make(map[string]float64, len(fans))
+	for _, fan := range fans {
+		if fan.Name == "" {
+			continue
+		}
+		fanValues[fan.Name] = fan.RPM
+		found := false
+		for i, name := range h.fanNames {
+			if name == fan.Name {
+				found = true
+				if i >= len(h.ringFans) {
+					h.ringFans = append(h.ringFans, newMetricsRing(120))
+				}
+				break
+			}
+		}
+		if !found {
+			h.fanNames = append(h.fanNames, fan.Name)
+			h.ringFans = append(h.ringFans, newMetricsRing(120))
+		}
+	}
+	for i, ring := range h.ringFans {
+		if ring == nil {
+			continue
+		}
+		name := ""
+		if i < len(h.fanNames) {
+			name = h.fanNames[i]
+		}
+		if rpm, ok := fanValues[name]; ok {
+			ring.push(rpm)
+			continue
+		}
+		if last, ok := ring.latest(); ok {
+			ring.push(last)
+			continue
+		}
+		ring.push(0)
+	}
+}
+
 func (h *handler) pushNamedMetricRing(dst *[]*namedMetricsRing, name string, value float64) {
 	if name == "" {
 		return
@@ -732,7 +876,10 @@ func (h *handler) applyPendingNetworkChange(apply func() (app.ActionResult, erro
 		return result, err
 	}
 
-	pnc := &pendingNetChange{snapshot: snapshot}
+	pnc := &pendingNetChange{
+		snapshot: snapshot,
+		deadline: time.Now().Add(netRollbackTimeout),
+	}
 	pnc.timer = time.AfterFunc(netRollbackTimeout, func() {
 		_ = h.opts.App.RestoreNetworkSnapshot(snapshot)
 		h.pendingNetMu.Lock()
@@ -749,6 +896,25 @@ func (h *handler) applyPendingNetworkChange(apply func() (app.ActionResult, erro
 	return result, nil
 }
 
+func (h *handler) hasPendingNetworkChange() bool {
+	h.pendingNetMu.Lock()
+	defer h.pendingNetMu.Unlock()
+	return h.pendingNet != nil
+}
+
+func (h *handler) pendingNetworkRollbackIn() int {
+	h.pendingNetMu.Lock()
+	defer h.pendingNetMu.Unlock()
+	if h.pendingNet == nil {
+		return 0
+	}
+	remaining := int(time.Until(h.pendingNet.deadline).Seconds())
+	if remaining < 1 {
+		return 1
+	}
+	return remaining
+}
+
 func (h *handler) handleAPINetworkConfirm(w http.ResponseWriter, _ *http.Request) {
 	h.pendingNetMu.Lock()
 	pnc := h.pendingNet
@@ -790,3 +956,108 @@ 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

@@ -0,0 +1,92 @@
+package webui
+
+import (
+	"net/http/httptest"
+	"strings"
+	"testing"
+
+	"bee/audit/internal/app"
+	"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
+	globalQueue.tasks = nil
+	globalQueue.mu.Unlock()
+	t.Cleanup(func() {
+		globalQueue.mu.Lock()
+		globalQueue.tasks = originalTasks
+		globalQueue.mu.Unlock()
+	})
+
+	h := &handler{opts: HandlerOptions{App: &app.App{}}}
+	req := httptest.NewRequest("POST", "/api/sat/cpu/run", strings.NewReader(`{"profile":"smoke"}`))
+	req.ContentLength = -1
+	rec := httptest.NewRecorder()
+
+	h.handleAPISATRun("cpu").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) != 1 {
+		t.Fatalf("tasks=%d want 1", len(globalQueue.tasks))
+	}
+	if got := globalQueue.tasks[0].params.BurnProfile; got != "smoke" {
+		t.Fatalf("burn profile=%q want smoke", got)
+	}
+}
+
+
+func TestPushFanRingsTracksByNameAndCarriesForwardMissingSamples(t *testing.T) {
+	h := &handler{}
+	h.pushFanRings([]platform.FanReading{
+		{Name: "FAN_A", RPM: 4200},
+		{Name: "FAN_B", RPM: 5100},
+	})
+	h.pushFanRings([]platform.FanReading{
+		{Name: "FAN_B", RPM: 5200},
+	})
+
+	if len(h.fanNames) != 2 || h.fanNames[0] != "FAN_A" || h.fanNames[1] != "FAN_B" {
+		t.Fatalf("fanNames=%v", h.fanNames)
+	}
+	aVals, _ := h.ringFans[0].snapshot()
+	bVals, _ := h.ringFans[1].snapshot()
+	if len(aVals) != 2 || len(bVals) != 2 {
+		t.Fatalf("fan ring lengths: A=%d B=%d", len(aVals), len(bVals))
+	}
+	if aVals[1] != 4200 {
+		t.Fatalf("FAN_A should carry forward last value, got %v", aVals)
+	}
+	if bVals[1] != 5200 {
+		t.Fatalf("FAN_B should use latest sampled value, got %v", bVals)
+	}
+}

audit/internal/webui/jobs.go

@@ -84,12 +84,12 @@ 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
 }
 

audit/internal/webui/kmsg_watcher.go

@@ -0,0 +1,241 @@
+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-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

@@ -4,7 +4,11 @@ import (
 	"database/sql"
 	"encoding/csv"
 	"io"
+	"os"
+	"path/filepath"
+	"sort"
 	"strconv"
+	"strings"
 	"time"
 
 	"bee/audit/internal/platform"
@@ -20,6 +24,9 @@ type MetricsDB struct {
 
 // 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 {
+		return nil, err
+	}
 	db, err := sql.Open("sqlite", path+"?_journal=WAL&_busy_timeout=5000")
 	if err != nil {
 		return nil, err
@@ -48,6 +55,8 @@ 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 (
@@ -64,6 +73,38 @@ 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
 }
 
@@ -85,8 +126,8 @@ 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
@@ -115,7 +156,7 @@ func (m *MetricsDB) Write(s platform.LiveMetricSample) error {
 
 // LoadRecent returns up to n samples in chronological order (oldest first).
 func (m *MetricsDB) LoadRecent(n int) ([]platform.LiveMetricSample, error) {
-	return m.loadSamples(`SELECT ts,cpu_load_pct,mem_load_pct,power_w FROM sys_metrics ORDER BY ts DESC LIMIT ?`, n)
+	return m.loadSamples(`SELECT ts,cpu_load_pct,mem_load_pct,power_w FROM (SELECT ts,cpu_load_pct,mem_load_pct,power_w FROM sys_metrics ORDER BY ts DESC LIMIT ?) ORDER BY ts`, n)
 }
 
 // LoadAll returns all persisted samples in chronological order (oldest first).
@@ -132,7 +173,7 @@ func (m *MetricsDB) loadSamples(query string, args ...any) ([]platform.LiveMetri
 	defer rows.Close()
 
 	type sysRow struct {
-		ts          int64
+		ts            int64
 		cpu, mem, pwr float64
 	}
 	var sysRows []sysRow
@@ -146,20 +187,18 @@ func (m *MetricsDB) loadSamples(query string, args ...any) ([]platform.LiveMetri
 	if len(sysRows) == 0 {
 		return nil, nil
 	}
-	// Reverse to chronological order
-	for i, j := 0, len(sysRows)-1; i < j; i, j = i+1, j-1 {
-		sysRows[i], sysRows[j] = sysRows[j], sysRows[i]
-	}
-
 	// Collect min/max ts for range query
 	minTS := sysRows[0].ts
 	maxTS := sysRows[len(sysRows)-1].ts
 
 	// Load GPU rows in range
-	type gpuKey struct{ ts int64; idx int }
+	type gpuKey struct {
+		ts  int64
+		idx int
+	}
 	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 {
@@ -167,14 +206,17 @@ 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
 			}
 		}
 	}
 
 	// Load fan rows in range
-	type fanKey struct{ ts int64; name string }
+	type fanKey struct {
+		ts   int64
+		name string
+	}
 	fanData := map[fanKey]float64{}
 	fRows, err := m.db.Query(
 		`SELECT ts,name,rpm FROM fan_metrics WHERE ts>=? AND ts<=?`, minTS, maxTS,
@@ -192,7 +234,10 @@ func (m *MetricsDB) loadSamples(query string, args ...any) ([]platform.LiveMetri
 	}
 
 	// Load temp rows in range
-	type tempKey struct{ ts int64; name string }
+	type tempKey struct {
+		ts   int64
+		name string
+	}
 	tempData := map[tempKey]platform.TempReading{}
 	tRows, err := m.db.Query(
 		`SELECT ts,name,grp,celsius FROM temp_metrics WHERE ts>=? AND ts<=?`, minTS, maxTS,
@@ -208,7 +253,9 @@ func (m *MetricsDB) loadSamples(query string, args ...any) ([]platform.LiveMetri
 		}
 	}
 
-	// Collect unique GPU indices and fan names from loaded data (preserve order)
+	// Collect unique GPU indices and fan/temp names from loaded data.
+	// Sort each list so that sample reconstruction is deterministic regardless
+	// of Go's non-deterministic map iteration order.
 	seenGPU := map[int]bool{}
 	var gpuIndices []int
 	for k := range gpuData {
@@ -217,6 +264,8 @@ func (m *MetricsDB) loadSamples(query string, args ...any) ([]platform.LiveMetri
 			gpuIndices = append(gpuIndices, k.idx)
 		}
 	}
+	sort.Ints(gpuIndices)
+
 	seenFan := map[string]bool{}
 	var fanNames []string
 	for k := range fanData {
@@ -225,6 +274,8 @@ func (m *MetricsDB) loadSamples(query string, args ...any) ([]platform.LiveMetri
 			fanNames = append(fanNames, k.name)
 		}
 	}
+	sort.Strings(fanNames)
+
 	seenTemp := map[string]bool{}
 	var tempNames []string
 	for k := range tempData {
@@ -233,6 +284,7 @@ func (m *MetricsDB) loadSamples(query string, args ...any) ([]platform.LiveMetri
 			tempNames = append(tempNames, k.name)
 		}
 	}
+	sort.Strings(tempNames)
 
 	samples := make([]platform.LiveMetricSample, len(sysRows))
 	for i, r := range sysRows {
@@ -266,7 +318,8 @@ 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
@@ -277,13 +330,13 @@ 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{
@@ -299,9 +352,11 @@ 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)
 	}

audit/internal/webui/metricsdb_test.go

@@ -0,0 +1,145 @@
+package webui
+
+import (
+	"database/sql"
+	"path/filepath"
+	"testing"
+	"time"
+
+	"bee/audit/internal/platform"
+	_ "modernc.org/sqlite"
+)
+
+func TestMetricsDBLoadSamplesKeepsChronologicalRangeForGPUs(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 < 3; i++ {
+		err := db.Write(platform.LiveMetricSample{
+			Timestamp:  base.Add(time.Duration(i) * time.Second),
+			CPULoadPct: float64(10 + i),
+			MemLoadPct: float64(20 + i),
+			PowerW:     float64(300 + i),
+			GPUs: []platform.GPUMetricRow{
+				{GPUIndex: 0, PowerW: float64(100 + i)},
+				{GPUIndex: 2, PowerW: float64(200 + i)},
+			},
+		})
+		if err != nil {
+			t.Fatalf("Write(%d): %v", i, err)
+		}
+	}
+
+	all, err := db.LoadAll()
+	if err != nil {
+		t.Fatalf("LoadAll: %v", err)
+	}
+	if len(all) != 3 {
+		t.Fatalf("LoadAll len=%d want 3", len(all))
+	}
+	for i, sample := range all {
+		if len(sample.GPUs) != 2 {
+			t.Fatalf("LoadAll sample %d GPUs=%v want 2 rows", i, sample.GPUs)
+		}
+		if sample.GPUs[0].GPUIndex != 0 || sample.GPUs[0].PowerW != float64(100+i) {
+			t.Fatalf("LoadAll sample %d GPU0=%+v", i, sample.GPUs[0])
+		}
+		if sample.GPUs[1].GPUIndex != 2 || sample.GPUs[1].PowerW != float64(200+i) {
+			t.Fatalf("LoadAll sample %d GPU1=%+v", i, sample.GPUs[1])
+		}
+	}
+
+	recent, err := db.LoadRecent(2)
+	if err != nil {
+		t.Fatalf("LoadRecent: %v", err)
+	}
+	if len(recent) != 2 {
+		t.Fatalf("LoadRecent len=%d want 2", len(recent))
+	}
+	if !recent[0].Timestamp.Before(recent[1].Timestamp) {
+		t.Fatalf("LoadRecent timestamps not ascending: %v >= %v", recent[0].Timestamp, recent[1].Timestamp)
+	}
+	for i, sample := range recent {
+		if len(sample.GPUs) != 2 {
+			t.Fatalf("LoadRecent sample %d GPUs=%v want 2 rows", i, sample.GPUs)
+		}
+	}
+}
+
+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)
+	}
+}

audit/internal/webui/server_test.go

@@ -34,6 +34,23 @@ 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 TestChartDataFromSamplesUsesFullHistory(t *testing.T) {
 	samples := []platform.LiveMetricSample{
 		{
@@ -89,6 +106,242 @@ func TestChartDataFromSamplesUsesFullHistory(t *testing.T) {
 	}
 }
 
+func TestChartDataFromSamplesKeepsStableGPUSeriesOrder(t *testing.T) {
+	samples := []platform.LiveMetricSample{
+		{
+			Timestamp: time.Now().Add(-2 * time.Minute),
+			GPUs: []platform.GPUMetricRow{
+				{GPUIndex: 7, PowerW: 170},
+				{GPUIndex: 2, PowerW: 120},
+				{GPUIndex: 0, PowerW: 100},
+			},
+		},
+		{
+			Timestamp: time.Now().Add(-1 * time.Minute),
+			GPUs: []platform.GPUMetricRow{
+				{GPUIndex: 0, PowerW: 101},
+				{GPUIndex: 7, PowerW: 171},
+				{GPUIndex: 2, PowerW: 121},
+			},
+		},
+	}
+
+	datasets, names, _, title, _, _, ok := chartDataFromSamples("gpu-all-power", samples)
+	if !ok {
+		t.Fatal("chartDataFromSamples returned ok=false")
+	}
+	if title != "GPU Power" {
+		t.Fatalf("title=%q", title)
+	}
+	wantNames := []string{"GPU 0", "GPU 2", "GPU 7"}
+	if len(names) != len(wantNames) {
+		t.Fatalf("names len=%d want %d: %v", len(names), len(wantNames), names)
+	}
+	for i := range wantNames {
+		if names[i] != wantNames[i] {
+			t.Fatalf("names[%d]=%q want %q; full=%v", i, names[i], wantNames[i], names)
+		}
+	}
+	if got := datasets[0]; len(got) != 2 || got[0] != 100 || got[1] != 101 {
+		t.Fatalf("GPU 0 dataset=%v want [100 101]", got)
+	}
+	if got := datasets[1]; len(got) != 2 || got[0] != 120 || got[1] != 121 {
+		t.Fatalf("GPU 2 dataset=%v want [120 121]", got)
+	}
+	if got := datasets[2]; len(got) != 2 || got[0] != 170 || got[1] != 171 {
+		t.Fatalf("GPU 7 dataset=%v want [170 171]", got)
+	}
+}
+
+func TestChartDataFromSamplesIncludesGPUClockCharts(t *testing.T) {
+	samples := []platform.LiveMetricSample{
+		{
+			Timestamp: time.Now().Add(-2 * time.Minute),
+			GPUs: []platform.GPUMetricRow{
+				{GPUIndex: 0, ClockMHz: 1400, MemClockMHz: 2600},
+				{GPUIndex: 3, ClockMHz: 1500, MemClockMHz: 2800},
+			},
+		},
+		{
+			Timestamp: time.Now().Add(-1 * time.Minute),
+			GPUs: []platform.GPUMetricRow{
+				{GPUIndex: 0, ClockMHz: 1410, MemClockMHz: 2610},
+				{GPUIndex: 3, ClockMHz: 1510, MemClockMHz: 2810},
+			},
+		},
+	}
+
+	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)
+	}
+
+	datasets, names, _, title, _, _, ok = chartDataFromSamples("gpu-all-memclock", samples)
+	if !ok {
+		t.Fatal("gpu-all-memclock returned ok=false")
+	}
+	if title != "GPU Memory 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[0][0]; got != 2600 {
+		t.Fatalf("GPU 0 memory clock=%v want 2600", got)
+	}
+}
+
+func TestNormalizePowerSeriesHoldsLastPositive(t *testing.T) {
+	got := normalizePowerSeries([]float64{0, 480, 0, 0, 510, 0})
+	want := []float64{0, 480, 480, 480, 510, 510}
+	if len(got) != len(want) {
+		t.Fatalf("len=%d want %d", len(got), len(want))
+	}
+	for i := range want {
+		if got[i] != want[i] {
+			t.Fatalf("got[%d]=%v want %v", i, got[i], want[i])
+		}
+	}
+}
+
+func TestRenderMetricsUsesBufferedChartRefresh(t *testing.T) {
+	body := renderMetrics()
+	if !strings.Contains(body, "const probe = new Image();") {
+		t.Fatalf("metrics page should preload chart images before swap: %s", body)
+	}
+	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 include GPU memory clock chart: %s", body)
+	}
+	if !strings.Contains(body, `renderGPUOverviewCards(indices)`) {
+		t.Fatalf("metrics page should build per-GPU chart cards dynamically: %s", body)
+	}
+}
+
+func TestChartLegendVisible(t *testing.T) {
+	if !chartLegendVisible(8) {
+		t.Fatal("legend should stay visible for charts with up to 8 series")
+	}
+	if chartLegendVisible(9) {
+		t.Fatal("legend should be hidden for charts with more than 8 series")
+	}
+}
+
+func TestChartYAxisNumber(t *testing.T) {
+	tests := []struct {
+		in   float64
+		want string
+	}{
+		{in: 999, want: "999"},
+		{in: 1000, want: "1к"},
+		{in: 1370, want: "1,4к"},
+		{in: 1500, want: "1,5к"},
+		{in: 1700, want: "1,7к"},
+		{in: 2000, want: "2к"},
+		{in: 9999, want: "10к"},
+		{in: 10200, want: "10к"},
+		{in: -1500, want: "-1,5к"},
+	}
+	for _, tc := range tests {
+		if got := chartYAxisNumber(tc.in); got != tc.want {
+			t.Fatalf("chartYAxisNumber(%v)=%q want %q", tc.in, got, tc.want)
+		}
+	}
+}
+
+func TestChartCanvasHeight(t *testing.T) {
+	if got := chartCanvasHeight(4); got != 360 {
+		t.Fatalf("chartCanvasHeight(4)=%d want 360", got)
+	}
+	if got := chartCanvasHeight(12); got != 288 {
+		t.Fatalf("chartCanvasHeight(12)=%d want 288", got)
+	}
+}
+
+func TestNormalizeFanSeriesHoldsLastPositive(t *testing.T) {
+	got := normalizeFanSeries([]float64{4200, 0, 0, 4300, 0})
+	want := []float64{4200, 4200, 4200, 4300, 4300}
+	if len(got) != len(want) {
+		t.Fatalf("len=%d want %d", len(got), len(want))
+	}
+	for i := range want {
+		if got[i] != want[i] {
+			t.Fatalf("got[%d]=%v want %v", i, got[i], want[i])
+		}
+	}
+}
+
+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)
+	r1.push(1000)
+	r1.push(1100)
+	r2.push(1200)
+	r2.push(1300)
+
+	datasets, names, labels := snapshotFanRings([]*metricsRing{r1, r2}, []string{"FAN_A", "FAN_B"})
+	if len(datasets) != 2 {
+		t.Fatalf("datasets=%d want 2", len(datasets))
+	}
+	if len(names) != 2 || names[0] != "FAN_A RPM" || names[1] != "FAN_B RPM" {
+		t.Fatalf("names=%v", names)
+	}
+	if len(labels) != 2 {
+		t.Fatalf("labels=%v want 2 entries", labels)
+	}
+	if labels[0] == "" || labels[1] == "" {
+		t.Fatalf("labels should contain timeline values, got %v", labels)
+	}
+}
+
+func TestRenderNetworkInlineSyncsPendingState(t *testing.T) {
+	body := renderNetworkInline()
+	if !strings.Contains(body, "d.pending_change") {
+		t.Fatalf("network UI should read pending network state from API: %s", body)
+	}
+	if !strings.Contains(body, "setInterval(loadNetwork, 5000)") {
+		t.Fatalf("network UI should periodically refresh network state: %s", body)
+	}
+	if !strings.Contains(body, "showNetPending(NET_ROLLBACK_SECS)") {
+		t.Fatalf("network UI should show pending confirmation immediately on apply: %s", body)
+	}
+}
+
 func TestRootRendersDashboard(t *testing.T) {
 	dir := t.TempDir()
 	path := filepath.Join(dir, "audit.json")
@@ -101,9 +354,10 @@ 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()
@@ -118,6 +372,11 @@ 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)
 	}
@@ -136,6 +395,33 @@ func TestRootRendersDashboard(t *testing.T) {
 	}
 }
 
+func TestRootShowsRunAuditButtonWhenSnapshotMissing(t *testing.T) {
+	dir := t.TempDir()
+	exportDir := filepath.Join(dir, "export")
+	if err := os.MkdirAll(exportDir, 0755); err != nil {
+		t.Fatal(err)
+	}
+
+	handler := NewHandler(HandlerOptions{
+		Title:     "Bee Hardware Audit",
+		AuditPath: filepath.Join(dir, "missing-audit.json"),
+		ExportDir: exportDir,
+	})
+
+	rec := httptest.NewRecorder()
+	handler.ServeHTTP(rec, httptest.NewRequest(http.MethodGet, "/", nil))
+	if rec.Code != http.StatusOK {
+		t.Fatalf("status=%d", rec.Code)
+	}
+	body := rec.Body.String()
+	if !strings.Contains(body, `Run Audit`) {
+		t.Fatalf("dashboard missing run audit button: %s", body)
+	}
+	if strings.Contains(body, `No audit data`) {
+		t.Fatalf("dashboard still shows empty audit badge: %s", body)
+	}
+}
+
 func TestAuditPageRendersViewerFrameAndActions(t *testing.T) {
 	dir := t.TempDir()
 	path := filepath.Join(dir, "audit.json")
@@ -158,6 +444,84 @@ func TestAuditPageRendersViewerFrameAndActions(t *testing.T) {
 	}
 }
 
+func TestTasksPageRendersLogModalAndPaginationControls(t *testing.T) {
+	handler := NewHandler(HandlerOptions{})
+	rec := httptest.NewRecorder()
+	handler.ServeHTTP(rec, httptest.NewRequest(http.MethodGet, "/tasks", nil))
+	if rec.Code != http.StatusOK {
+		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, `_taskPageSize = 50`) {
+		t.Fatalf("tasks page missing pagination size config: %s", body)
+	}
+	if !strings.Contains(body, `Previous</button>`) || !strings.Contains(body, `Next</button>`) {
+		t.Fatalf("tasks page missing pagination controls: %s", body)
+	}
+}
+
+func TestToolsPageRendersRestartGPUDriversButton(t *testing.T) {
+	handler := NewHandler(HandlerOptions{})
+	rec := httptest.NewRecorder()
+	handler.ServeHTTP(rec, httptest.NewRequest(http.MethodGet, "/tools", nil))
+	if rec.Code != http.StatusOK {
+		t.Fatalf("status=%d", rec.Code)
+	}
+	body := rec.Body.String()
+	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, `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 TestTasksPageRendersScrollableLogModal(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"}`), 0644); err != nil {
+		t.Fatal(err)
+	}
+
+	handler := NewHandler(HandlerOptions{
+		Title:     "Bee Hardware Audit",
+		AuditPath: path,
+		ExportDir: exportDir,
+	})
+
+	rec := httptest.NewRecorder()
+	handler.ServeHTTP(rec, httptest.NewRequest(http.MethodGet, "/tasks", nil))
+	if rec.Code != http.StatusOK {
+		t.Fatalf("status=%d", rec.Code)
+	}
+	body := rec.Body.String()
+	if !strings.Contains(body, `height:calc(100vh - 32px)`) {
+		t.Fatalf("tasks page missing bounded log modal height: %s", body)
+	}
+	if !strings.Contains(body, `flex:1;min-height:0;overflow:hidden`) {
+		t.Fatalf("tasks page missing log modal overflow guard: %s", body)
+	}
+	if !strings.Contains(body, `height:100%;min-height:0;overflow:auto`) {
+		t.Fatalf("tasks page missing scrollable log wrapper: %s", body)
+	}
+}
+
 func TestViewerRendersLatestSnapshot(t *testing.T) {
 	dir := t.TempDir()
 	path := filepath.Join(dir, "audit.json")
@@ -232,6 +596,17 @@ func TestSupportBundleEndpointReturnsArchive(t *testing.T) {
 	if err := os.WriteFile(filepath.Join(exportDir, "bee-audit.log"), []byte("audit log"), 0644); err != nil {
 		t.Fatal(err)
 	}
+	archive, err := os.CreateTemp(os.TempDir(), "bee-support-server-test-*.tar.gz")
+	if err != nil {
+		t.Fatal(err)
+	}
+	t.Cleanup(func() { _ = os.Remove(archive.Name()) })
+	if _, err := archive.WriteString("support-bundle"); err != nil {
+		t.Fatal(err)
+	}
+	if err := archive.Close(); err != nil {
+		t.Fatal(err)
+	}
 
 	handler := NewHandler(HandlerOptions{ExportDir: exportDir})
 	rec := httptest.NewRecorder()

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/tasks.go

@@ -4,10 +4,14 @@ import (
 	"context"
 	"encoding/json"
 	"fmt"
+	"log/slog"
 	"net/http"
 	"os"
+	"os/exec"
 	"path/filepath"
+	"runtime/debug"
 	"sort"
+	"strings"
 	"sync"
 	"time"
 
@@ -24,36 +28,74 @@ const (
 	TaskCancelled = "cancelled"
 )
 
-// taskNames maps target → human-readable name.
+// taskNames maps target → human-readable name for validate (SAT) runs.
 var taskNames = map[string]string{
-	"nvidia":         "NVIDIA SAT",
-	"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)",
+	"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",
-	"install":        "Install to Disk",
-	"install-to-ram": "Install to RAM",
+	"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.
+var burnNames = map[string]string{
+	"nvidia": "NVIDIA Burn-in",
+	"memory": "Memory Burn-in",
+	"cpu":    "CPU Burn-in",
+	"amd":    "AMD GPU Burn-in",
+}
+
+func nvidiaStressTaskName(loader string) string {
+	switch strings.TrimSpace(strings.ToLower(loader)) {
+	case platform.NvidiaStressLoaderJohn:
+		return "NVIDIA GPU Stress (John/OpenCL)"
+	case platform.NvidiaStressLoaderNCCL:
+		return "NVIDIA GPU Stress (NCCL)"
+	default:
+		return "NVIDIA GPU Stress (bee-gpu-burn)"
+	}
+}
+
+func taskDisplayName(target, profile, loader string) string {
+	name := taskNames[target]
+	if profile != "" {
+		if n, ok := burnNames[target]; ok {
+			name = n
+		}
+	}
+	if target == "nvidia-stress" {
+		name = nvidiaStressTaskName(loader)
+	}
+	if name == "" {
+		name = target
+	}
+	return name
 }
 
 // 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"`
-	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"`
 
 	// runtime fields (not serialised)
 	job    *jobState
@@ -62,12 +104,15 @@ 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"`
-	GPUIndices  []int  `json:"gpu_indices,omitempty"`
-	BurnProfile string `json:"burn_profile,omitempty"`
-	DisplayName string `json:"display_name,omitempty"`
-	Device      string `json:"device,omitempty"` // for install
+	Duration           int      `json:"duration,omitempty"`
+	DiagLevel          int      `json:"diag_level,omitempty"`
+	GPUIndices         []int    `json:"gpu_indices,omitempty"`
+	ExcludeGPUIndices  []int    `json:"exclude_gpu_indices,omitempty"`
+	Loader             string   `json:"loader,omitempty"`
+	BurnProfile        string   `json:"burn_profile,omitempty"`
+	DisplayName        string   `json:"display_name,omitempty"`
+	Device             string   `json:"device,omitempty"` // for install
+	PlatformComponents []string `json:"platform_components,omitempty"`
 }
 
 type persistedTask struct {
@@ -130,13 +175,14 @@ func resolvePlatformStressPreset(profile string) platform.PlatformStressOptions
 
 // 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)}
@@ -162,6 +208,9 @@ var (
 	runAMDMemBandwidthPackCtx = func(a *app.App, ctx context.Context, baseDir string, logFunc func(string)) (string, error) {
 		return a.RunAMDMemBandwidthPackCtx(ctx, baseDir, logFunc)
 	}
+	runNvidiaStressPackCtx = func(a *app.App, ctx context.Context, baseDir string, opts platform.NvidiaStressOptions, logFunc func(string)) (string, error) {
+		return a.RunNvidiaStressPackCtx(ctx, baseDir, opts, logFunc)
+	}
 	runAMDStressPackCtx = func(a *app.App, ctx context.Context, baseDir string, durationSec int, logFunc func(string)) (string, error) {
 		return a.RunAMDStressPackCtx(ctx, baseDir, durationSec, logFunc)
 	}
@@ -171,6 +220,10 @@ var (
 	runSATStressPackCtx = func(a *app.App, ctx context.Context, baseDir string, durationSec int, logFunc func(string)) (string, error) {
 		return a.RunSATStressPackCtx(ctx, baseDir, durationSec, logFunc)
 	}
+	buildSupportBundle = app.BuildSupportBundle
+	installCommand     = func(ctx context.Context, device string, logPath string) *exec.Cmd {
+		return exec.CommandContext(ctx, "bee-install", device, logPath)
+	}
 )
 
 // enqueue adds a task to the queue and notifies the worker.
@@ -241,6 +294,30 @@ func (q *taskQueue) findJob(id string) (*jobState, bool) {
 	return t.job, true
 }
 
+type taskStreamSource struct {
+	status  string
+	errMsg  string
+	logPath string
+	job     *jobState
+}
+
+func (q *taskQueue) taskStreamSource(id string) (taskStreamSource, bool) {
+	q.mu.Lock()
+	defer q.mu.Unlock()
+	for _, t := range q.tasks {
+		if t.ID != id {
+			continue
+		}
+		return taskStreamSource{
+			status:  t.Status,
+			errMsg:  t.ErrMsg,
+			logPath: t.LogPath,
+			job:     t.job,
+		}, true
+	}
+	return taskStreamSource{}, false
+}
+
 func (q *taskQueue) hasActiveTarget(target string) bool {
 	q.mu.Lock()
 	defer q.mu.Unlock()
@@ -255,15 +332,19 @@ func (q *taskQueue) hasActiveTarget(target string) bool {
 	return false
 }
 
-// snapshot returns a copy of all tasks sorted for display (running first, then pending by priority, then done by doneAt desc).
+// snapshot returns a copy of all tasks sorted for display with newest tasks first.
 func (q *taskQueue) snapshot() []Task {
 	q.mu.Lock()
 	defer q.mu.Unlock()
 	out := make([]Task, len(q.tasks))
 	for i, t := range q.tasks {
 		out[i] = *t
+		out[i].ElapsedSec = taskElapsedSec(&out[i], time.Now())
 	}
 	sort.SliceStable(out, func(i, j int) bool {
+		if !out[i].CreatedAt.Equal(out[j].CreatedAt) {
+			return out[i].CreatedAt.After(out[j].CreatedAt)
+		}
 		si := statusOrder(out[i].Status)
 		sj := statusOrder(out[j].Status)
 		if si != sj {
@@ -272,7 +353,7 @@ func (q *taskQueue) snapshot() []Task {
 		if out[i].Priority != out[j].Priority {
 			return out[i].Priority > out[j].Priority
 		}
-		return out[i].CreatedAt.Before(out[j].CreatedAt)
+		return out[i].Name < out[j].Name
 	})
 	return out
 }
@@ -298,7 +379,7 @@ 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()
@@ -313,44 +394,90 @@ 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()
-
-			q.runTask(t, j, ctx)
+		func() {
+			setCPUGovernor("performance")
+			defer setCPUGovernor("powersave")
 
+			// Drain all pending tasks and start them in parallel.
 			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
+			var batch []*Task
+			for {
+				t := q.nextPending()
+				if t == nil {
+					break
 				}
+				now := time.Now()
+				t.Status = TaskRunning
+				t.StartedAt = &now
+				t.DoneAt = nil
+				t.ErrMsg = ""
+				j := newTaskJobState(t.LogPath)
+				t.job = j
+				batch = append(batch, t)
+			}
+			if len(batch) > 0 {
+				q.persistLocked()
 			}
-			q.prune()
-			q.persistLocked()
 			q.mu.Unlock()
-		}
-		setCPUGovernor("powersave")
+
+			var wg sync.WaitGroup
+			for _, t := range batch {
+				t := t
+				j := t.job
+				taskCtx, taskCancel := context.WithCancel(context.Background())
+				j.cancel = taskCancel
+				wg.Add(1)
+				goRecoverOnce("task "+t.Target, func() {
+					defer wg.Done()
+					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)
+					}
+
+					q.runTask(t, j, taskCtx)
+
+					if q.kmsgWatcher != nil {
+						q.kmsgWatcher.NotifyTaskFinished(t.ID)
+					}
+
+					q.mu.Lock()
+					now2 := time.Now()
+					t.DoneAt = &now2
+					if t.Status == TaskRunning {
+						if j.err != "" {
+							t.Status = TaskFailed
+							t.ErrMsg = j.err
+						} else {
+							t.Status = TaskDone
+						}
+					}
+					q.persistLocked()
+					q.mu.Unlock()
+				})
+			}
+			wg.Wait()
+
+			if len(batch) > 0 {
+				q.mu.Lock()
+				q.prune()
+				q.persistLocked()
+				q.mu.Unlock()
+			}
+		}()
+
 	}
 }
 
@@ -368,9 +495,9 @@ func setCPUGovernor(governor string) {
 
 // runTask executes the work for a task, writing output to j.
 func (q *taskQueue) runTask(t *Task, j *jobState, ctx context.Context) {
-	if q.opts == nil || q.opts.App == nil {
-		j.append("ERROR: app not configured")
-		j.finish("app not configured")
+	if q.opts == nil {
+		j.append("ERROR: handler options not configured")
+		j.finish("handler options not configured")
 		return
 	}
 	a := q.opts.App
@@ -387,6 +514,10 @@ func (q *taskQueue) runTask(t *Task, j *jobState, ctx context.Context) {
 
 	switch t.Target {
 	case "nvidia":
+		if a == nil {
+			err = fmt.Errorf("app not configured")
+			break
+		}
 		diagLevel := t.params.DiagLevel
 		if t.params.BurnProfile != "" && diagLevel <= 0 {
 			diagLevel = resolveBurnPreset(t.params.BurnProfile).NvidiaDiag
@@ -403,11 +534,38 @@ func (q *taskQueue) runTask(t *Task, j *jobState, ctx context.Context) {
 		} else {
 			archive, err = a.RunNvidiaAcceptancePack("", 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,
+		}, j.append)
 	case "memory":
+		if a == nil {
+			err = fmt.Errorf("app not configured")
+			break
+		}
 		archive, err = runMemoryAcceptancePackCtx(a, ctx, "", j.append)
 	case "storage":
+		if a == nil {
+			err = fmt.Errorf("app not configured")
+			break
+		}
 		archive, err = runStorageAcceptancePackCtx(a, ctx, "", j.append)
 	case "cpu":
+		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
@@ -415,35 +573,69 @@ func (q *taskQueue) runTask(t *Task, j *jobState, ctx context.Context) {
 		if dur <= 0 {
 			dur = 60
 		}
+		j.append(fmt.Sprintf("CPU stress duration: %ds", dur))
 		archive, err = runCPUAcceptancePackCtx(a, ctx, "", dur, j.append)
 	case "amd":
+		if a == nil {
+			err = fmt.Errorf("app not configured")
+			break
+		}
 		archive, err = runAMDAcceptancePackCtx(a, ctx, "", j.append)
 	case "amd-mem":
+		if a == nil {
+			err = fmt.Errorf("app not configured")
+			break
+		}
 		archive, err = runAMDMemIntegrityPackCtx(a, ctx, "", j.append)
 	case "amd-bandwidth":
+		if a == nil {
+			err = fmt.Errorf("app not configured")
+			break
+		}
 		archive, err = runAMDMemBandwidthPackCtx(a, ctx, "", j.append)
 	case "amd-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 = runAMDStressPackCtx(a, ctx, "", dur, j.append)
 	case "memory-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 = runMemoryStressPackCtx(a, ctx, "", dur, j.append)
 	case "sat-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 = runSATStressPackCtx(a, ctx, "", dur, j.append)
 	case "platform-stress":
+		if a == nil {
+			err = fmt.Errorf("app not configured")
+			break
+		}
 		opts := resolvePlatformStressPreset(t.params.BurnProfile)
+		opts.Components = t.params.PlatformComponents
 		archive, err = a.RunPlatformStress(ctx, "", opts, j.append)
 	case "audit":
+		if a == nil {
+			err = fmt.Errorf("app not configured")
+			break
+		}
 		result, e := a.RunAuditNow(q.opts.RuntimeMode)
 		if e != nil {
 			err = e
@@ -452,7 +644,22 @@ func (q *taskQueue) runTask(t *Task, j *jobState, ctx context.Context) {
 				j.append(line)
 			}
 		}
+	case "support-bundle":
+		j.append("Building support bundle...")
+		archive, err = buildSupportBundle(q.opts.ExportDir)
+	case "install":
+		if strings.TrimSpace(t.params.Device) == "" {
+			err = fmt.Errorf("device is required")
+			break
+		}
+		installLogPath := platform.InstallLogPath(t.params.Device)
+		j.append("Install log: " + installLogPath)
+		err = streamCmdJob(j, installCommand(ctx, t.params.Device, installLogPath))
 	case "install-to-ram":
+		if a == nil {
+			err = fmt.Errorf("app not configured")
+			break
+		}
 		err = a.RunInstallToRAM(ctx, j.append)
 	default:
 		j.append("ERROR: unknown target: " + t.Target)
@@ -460,6 +667,19 @@ 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.")
@@ -476,6 +696,13 @@ 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) {
@@ -585,23 +812,83 @@ func (h *handler) handleAPITasksCancelAll(w http.ResponseWriter, _ *http.Request
 	writeJSON(w, map[string]int{"cancelled": n})
 }
 
+func (h *handler) handleAPITasksKillWorkers(w http.ResponseWriter, _ *http.Request) {
+	// Cancel all queued/running tasks in the queue first.
+	globalQueue.mu.Lock()
+	now := time.Now()
+	cancelled := 0
+	for _, t := range globalQueue.tasks {
+		switch t.Status {
+		case TaskPending:
+			t.Status = TaskCancelled
+			t.DoneAt = &now
+			cancelled++
+		case TaskRunning:
+			if t.job != nil {
+				t.job.abort()
+			}
+			t.Status = TaskCancelled
+			t.DoneAt = &now
+			cancelled++
+		}
+	}
+	globalQueue.persistLocked()
+	globalQueue.mu.Unlock()
+
+	// Kill orphaned test worker processes at the OS level.
+	killed := platform.KillTestWorkers()
+	writeJSON(w, map[string]any{
+		"cancelled": cancelled,
+		"killed":    len(killed),
+		"processes": killed,
+	})
+}
+
 func (h *handler) handleAPITasksStream(w http.ResponseWriter, r *http.Request) {
 	id := r.PathValue("id")
-	// Wait up to 5s for the task to get a job (it may be pending)
-	deadline := time.Now().Add(5 * time.Second)
-	var j *jobState
-	for time.Now().Before(deadline) {
-		if jj, ok := globalQueue.findJob(id); ok {
-			j = jj
-			break
-		}
-		time.Sleep(200 * time.Millisecond)
-	}
-	if j == nil {
-		http.Error(w, "task not found or not yet started", http.StatusNotFound)
+	src, ok := globalQueue.taskStreamSource(id)
+	if !ok {
+		http.Error(w, "task not found", http.StatusNotFound)
 		return
 	}
-	streamJob(w, r, j)
+	if src.job != nil {
+		streamJob(w, r, src.job)
+		return
+	}
+	if src.status == TaskDone || src.status == TaskFailed || src.status == TaskCancelled {
+		j := newTaskJobState(src.logPath)
+		j.finish(src.errMsg)
+		streamJob(w, r, j)
+		return
+	}
+	if !sseStart(w) {
+		return
+	}
+	sseWrite(w, "", "Task is queued. Waiting for worker...")
+	ticker := time.NewTicker(200 * time.Millisecond)
+	defer ticker.Stop()
+	for {
+		select {
+		case <-ticker.C:
+			src, ok = globalQueue.taskStreamSource(id)
+			if !ok {
+				sseWrite(w, "done", "task not found")
+				return
+			}
+			if src.job != nil {
+				streamSubscribedJob(w, r, src.job)
+				return
+			}
+			if src.status == TaskDone || src.status == TaskFailed || src.status == TaskCancelled {
+				j := newTaskJobState(src.logPath)
+				j.finish(src.errMsg)
+				streamSubscribedJob(w, r, j)
+				return
+			}
+		case <-r.Context().Done():
+			return
+		}
+	}
 }
 
 func (q *taskQueue) assignTaskLogPathLocked(t *Task) {
@@ -638,8 +925,18 @@ func (q *taskQueue) loadLocked() {
 			params:    pt.Params,
 		}
 		q.assignTaskLogPathLocked(t)
-		if t.Status == TaskPending || t.Status == TaskRunning {
-			t.Status = TaskPending
+		if t.Status == TaskRunning {
+			// The task was interrupted by a bee-web restart. Child processes
+			// (e.g. bee-gpu-burn-worker) survive the restart in their own
+			// process groups and cannot be cancelled retroactively. Mark the
+			// task as failed so the user can decide whether to re-run it
+			// rather than blindly re-launching duplicate workers.
+			now := time.Now()
+			t.Status = TaskFailed
+			t.DoneAt = &now
+			t.ErrMsg = "interrupted by bee-web restart"
+		} else if t.Status == TaskPending {
+			t.StartedAt = nil
 			t.DoneAt = nil
 			t.ErrMsg = ""
 		}
@@ -679,3 +976,21 @@ func (q *taskQueue) persistLocked() {
 	}
 	_ = os.Rename(tmp, q.statePath)
 }
+
+func taskElapsedSec(t *Task, now time.Time) int {
+	if t == nil || t.StartedAt == nil || t.StartedAt.IsZero() {
+		return 0
+	}
+	start := *t.StartedAt
+	if !t.CreatedAt.IsZero() && start.Before(t.CreatedAt) {
+		start = t.CreatedAt
+	}
+	end := now
+	if t.DoneAt != nil && !t.DoneAt.IsZero() {
+		end = *t.DoneAt
+	}
+	if end.Before(start) {
+		return 0
+	}
+	return int(end.Sub(start).Round(time.Second) / time.Second)
+}

audit/internal/webui/tasks_test.go

@@ -2,8 +2,12 @@ package webui
 
 import (
 	"context"
+	"net/http"
+	"net/http/httptest"
 	"os"
+	"os/exec"
 	"path/filepath"
+	"strings"
 	"testing"
 	"time"
 
@@ -22,21 +26,34 @@ func TestTaskQueuePersistsAndRecoversPendingTasks(t *testing.T) {
 	}
 
 	started := time.Now().Add(-time.Minute)
-	task := &Task{
-		ID:        "task-1",
+
+	// A task that was pending (not yet started) must be re-queued on restart.
+	pendingTask := &Task{
+		ID:        "task-pending",
 		Name:      "Memory Burn-in",
 		Target:    "memory-stress",
 		Priority:  2,
-		Status:    TaskRunning,
+		Status:    TaskPending,
 		CreatedAt: time.Now().Add(-2 * time.Minute),
-		StartedAt: &started,
-		params: taskParams{
-			Duration:    300,
-			BurnProfile: "smoke",
-		},
+		params:    taskParams{Duration: 300, BurnProfile: "smoke"},
+	}
+	// A task that was running when bee-web crashed must NOT be re-queued —
+	// its child processes (e.g. gpu-burn-worker) survive the restart in
+	// their own process groups and can't be cancelled retroactively.
+	runningTask := &Task{
+		ID:        "task-running",
+		Name:      "NVIDIA GPU Stress",
+		Target:    "nvidia-stress",
+		Priority:  1,
+		Status:    TaskRunning,
+		CreatedAt: time.Now().Add(-3 * time.Minute),
+		StartedAt: &started,
+		params:    taskParams{Duration: 86400},
+	}
+	for _, task := range []*Task{pendingTask, runningTask} {
+		q.tasks = append(q.tasks, task)
+		q.assignTaskLogPathLocked(task)
 	}
-	q.tasks = append(q.tasks, task)
-	q.assignTaskLogPathLocked(task)
 	q.persistLocked()
 
 	recovered := &taskQueue{
@@ -46,18 +63,47 @@ func TestTaskQueuePersistsAndRecoversPendingTasks(t *testing.T) {
 	}
 	recovered.loadLocked()
 
-	if len(recovered.tasks) != 1 {
-		t.Fatalf("tasks=%d want 1", len(recovered.tasks))
+	if len(recovered.tasks) != 2 {
+		t.Fatalf("tasks=%d want 2", len(recovered.tasks))
 	}
-	got := recovered.tasks[0]
-	if got.Status != TaskPending {
-		t.Fatalf("status=%q want %q", got.Status, TaskPending)
+
+	byID := map[string]*Task{}
+	for i := range recovered.tasks {
+		byID[recovered.tasks[i].ID] = recovered.tasks[i]
 	}
-	if got.params.Duration != 300 || got.params.BurnProfile != "smoke" {
-		t.Fatalf("params=%+v", got.params)
+
+	// Pending task must be re-queued as pending with params intact.
+	p := byID["task-pending"]
+	if p == nil {
+		t.Fatal("task-pending not found")
 	}
-	if got.LogPath == "" {
-		t.Fatal("expected log path")
+	if p.Status != TaskPending {
+		t.Fatalf("pending task: status=%q want %q", p.Status, TaskPending)
+	}
+	if p.StartedAt != nil {
+		t.Fatalf("pending task: started_at=%v want nil", p.StartedAt)
+	}
+	if p.params.Duration != 300 || p.params.BurnProfile != "smoke" {
+		t.Fatalf("pending task: params=%+v", p.params)
+	}
+	if p.LogPath == "" {
+		t.Fatal("pending task: expected log path")
+	}
+
+	// Running task must be marked failed, not re-queued, to prevent
+	// launching duplicate workers (e.g. a second set of gpu-burn-workers).
+	r := byID["task-running"]
+	if r == nil {
+		t.Fatal("task-running not found")
+	}
+	if r.Status != TaskFailed {
+		t.Fatalf("running task: status=%q want %q", r.Status, TaskFailed)
+	}
+	if r.ErrMsg == "" {
+		t.Fatal("running task: expected non-empty error message")
+	}
+	if r.DoneAt == nil {
+		t.Fatal("running task: expected done_at to be set")
 	}
 }
 
@@ -78,6 +124,130 @@ func TestNewTaskJobStateLoadsExistingLog(t *testing.T) {
 	}
 }
 
+func TestTaskQueueSnapshotSortsNewestFirst(t *testing.T) {
+	now := time.Date(2026, 4, 2, 12, 0, 0, 0, time.UTC)
+	q := &taskQueue{
+		tasks: []*Task{
+			{
+				ID:        "old-running",
+				Name:      "Old Running",
+				Status:    TaskRunning,
+				Priority:  10,
+				CreatedAt: now.Add(-3 * time.Minute),
+			},
+			{
+				ID:        "new-done",
+				Name:      "New Done",
+				Status:    TaskDone,
+				Priority:  0,
+				CreatedAt: now.Add(-1 * time.Minute),
+			},
+			{
+				ID:        "mid-pending",
+				Name:      "Mid Pending",
+				Status:    TaskPending,
+				Priority:  1,
+				CreatedAt: now.Add(-2 * time.Minute),
+			},
+		},
+	}
+
+	got := q.snapshot()
+	if len(got) != 3 {
+		t.Fatalf("snapshot len=%d want 3", len(got))
+	}
+	if got[0].ID != "new-done" || got[1].ID != "mid-pending" || got[2].ID != "old-running" {
+		t.Fatalf("snapshot order=%q,%q,%q", got[0].ID, got[1].ID, got[2].ID)
+	}
+}
+
+func TestHandleAPITasksStreamReplaysPersistedLogWithoutLiveJob(t *testing.T) {
+	dir := t.TempDir()
+	logPath := filepath.Join(dir, "task.log")
+	if err := os.WriteFile(logPath, []byte("line1\nline2\n"), 0644); err != nil {
+		t.Fatal(err)
+	}
+
+	globalQueue.mu.Lock()
+	origTasks := globalQueue.tasks
+	globalQueue.tasks = []*Task{{
+		ID:        "done-1",
+		Name:      "Done Task",
+		Status:    TaskDone,
+		CreatedAt: time.Now(),
+		LogPath:   logPath,
+	}}
+	globalQueue.mu.Unlock()
+	t.Cleanup(func() {
+		globalQueue.mu.Lock()
+		globalQueue.tasks = origTasks
+		globalQueue.mu.Unlock()
+	})
+
+	req := httptest.NewRequest(http.MethodGet, "/api/tasks/done-1/stream", nil)
+	req.SetPathValue("id", "done-1")
+	rec := httptest.NewRecorder()
+
+	h := &handler{}
+	h.handleAPITasksStream(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: line1\n\n") || !strings.Contains(body, "data: line2\n\n") {
+		t.Fatalf("body=%q", body)
+	}
+	if !strings.Contains(body, "event: done\n") {
+		t.Fatalf("missing done event: %q", body)
+	}
+}
+
+func TestHandleAPITasksStreamPendingTaskStartsSSEImmediately(t *testing.T) {
+	globalQueue.mu.Lock()
+	origTasks := globalQueue.tasks
+	globalQueue.tasks = []*Task{{
+		ID:        "pending-1",
+		Name:      "Pending Task",
+		Status:    TaskPending,
+		CreatedAt: time.Now(),
+	}}
+	globalQueue.mu.Unlock()
+	t.Cleanup(func() {
+		globalQueue.mu.Lock()
+		globalQueue.tasks = origTasks
+		globalQueue.mu.Unlock()
+	})
+
+	ctx, cancel := context.WithCancel(context.Background())
+	req := httptest.NewRequest(http.MethodGet, "/api/tasks/pending-1/stream", nil).WithContext(ctx)
+	req.SetPathValue("id", "pending-1")
+	rec := httptest.NewRecorder()
+
+	done := make(chan struct{})
+	go func() {
+		h := &handler{}
+		h.handleAPITasksStream(rec, req)
+		close(done)
+	}()
+
+	deadline := time.Now().Add(2 * time.Second)
+	for time.Now().Before(deadline) {
+		if strings.Contains(rec.Body.String(), "Task is queued. Waiting for worker...") {
+			cancel()
+			<-done
+			if rec.Code != http.StatusOK {
+				t.Fatalf("status=%d body=%s", rec.Code, rec.Body.String())
+			}
+			return
+		}
+		time.Sleep(20 * time.Millisecond)
+	}
+	cancel()
+	<-done
+	t.Fatalf("stream did not emit queued status promptly, body=%q", rec.Body.String())
+}
+
 func TestResolveBurnPreset(t *testing.T) {
 	tests := []struct {
 		profile string
@@ -95,9 +265,24 @@ func TestResolveBurnPreset(t *testing.T) {
 	}
 }
 
-func TestRunTaskHonorsCancel(t *testing.T) {
-	t.Parallel()
+func TestTaskDisplayNameUsesNvidiaStressLoader(t *testing.T) {
+	tests := []struct {
+		loader string
+		want   string
+	}{
+		{loader: "", want: "NVIDIA GPU Stress (bee-gpu-burn)"},
+		{loader: "builtin", want: "NVIDIA GPU Stress (bee-gpu-burn)"},
+		{loader: "john", want: "NVIDIA GPU Stress (John/OpenCL)"},
+		{loader: "nccl", want: "NVIDIA GPU Stress (NCCL)"},
+	}
+	for _, tc := range tests {
+		if got := taskDisplayName("nvidia-stress", "acceptance", tc.loader); got != tc.want {
+			t.Fatalf("taskDisplayName(loader=%q)=%q want %q", tc.loader, got, tc.want)
+		}
+	}
+}
 
+func TestRunTaskHonorsCancel(t *testing.T) {
 	blocked := make(chan struct{})
 	released := make(chan struct{})
 	aRun := func(_ any, ctx context.Context, _ string, _ int, _ func(string)) (string, error) {
@@ -154,3 +339,131 @@ func TestRunTaskHonorsCancel(t *testing.T) {
 		t.Fatal("runTask did not return after cancel")
 	}
 }
+
+func TestRunTaskUsesBurnProfileDurationForCPU(t *testing.T) {
+	var gotDuration int
+	q := &taskQueue{
+		opts: &HandlerOptions{App: &app.App{}},
+	}
+	tk := &Task{
+		ID:        "cpu-burn-1",
+		Name:      "CPU Burn-in",
+		Target:    "cpu",
+		Status:    TaskRunning,
+		CreatedAt: time.Now(),
+		params:    taskParams{BurnProfile: "smoke"},
+	}
+	j := &jobState{}
+
+	orig := runCPUAcceptancePackCtx
+	runCPUAcceptancePackCtx = func(_ *app.App, _ context.Context, _ string, durationSec int, _ func(string)) (string, error) {
+		gotDuration = durationSec
+		return "/tmp/cpu-burn.tar.gz", nil
+	}
+	defer func() { runCPUAcceptancePackCtx = orig }()
+
+	q.runTask(tk, j, context.Background())
+
+	if gotDuration != 5*60 {
+		t.Fatalf("duration=%d want %d", gotDuration, 5*60)
+	}
+}
+
+func TestRunTaskBuildsSupportBundleWithoutApp(t *testing.T) {
+	dir := t.TempDir()
+	q := &taskQueue{
+		opts: &HandlerOptions{ExportDir: dir},
+	}
+	tk := &Task{
+		ID:        "support-bundle-1",
+		Name:      "Support Bundle",
+		Target:    "support-bundle",
+		Status:    TaskRunning,
+		CreatedAt: time.Now(),
+	}
+	j := &jobState{}
+
+	var gotExportDir string
+	orig := buildSupportBundle
+	buildSupportBundle = func(exportDir string) (string, error) {
+		gotExportDir = exportDir
+		return filepath.Join(exportDir, "bundle.tar.gz"), nil
+	}
+	defer func() { buildSupportBundle = orig }()
+
+	q.runTask(tk, j, context.Background())
+
+	if gotExportDir != dir {
+		t.Fatalf("exportDir=%q want %q", gotExportDir, dir)
+	}
+	if j.err != "" {
+		t.Fatalf("unexpected error: %q", j.err)
+	}
+	if !strings.Contains(strings.Join(j.lines, "\n"), "Archive: "+filepath.Join(dir, "bundle.tar.gz")) {
+		t.Fatalf("lines=%v", j.lines)
+	}
+}
+
+func TestTaskElapsedSecClampsInvalidStartedAt(t *testing.T) {
+	now := time.Date(2026, 4, 1, 19, 10, 0, 0, time.UTC)
+	created := time.Date(2026, 4, 1, 19, 4, 5, 0, time.UTC)
+	started := time.Time{}
+	task := &Task{
+		Status:    TaskRunning,
+		CreatedAt: created,
+		StartedAt: &started,
+	}
+	if got := taskElapsedSec(task, now); got != 0 {
+		t.Fatalf("taskElapsedSec(zero start)=%d want 0", got)
+	}
+
+	stale := created.Add(-24 * time.Hour)
+	task.StartedAt = &stale
+	if got := taskElapsedSec(task, now); got != int(now.Sub(created).Seconds()) {
+		t.Fatalf("taskElapsedSec(stale start)=%d want %d", got, int(now.Sub(created).Seconds()))
+	}
+}
+
+func TestRunTaskInstallUsesSharedCommandStreaming(t *testing.T) {
+	q := &taskQueue{
+		opts: &HandlerOptions{},
+	}
+	tk := &Task{
+		ID:        "install-1",
+		Name:      "Install to Disk",
+		Target:    "install",
+		Status:    TaskRunning,
+		CreatedAt: time.Now(),
+		params:    taskParams{Device: "/dev/sda"},
+	}
+	j := &jobState{}
+
+	var gotDevice string
+	var gotLogPath string
+	orig := installCommand
+	installCommand = func(ctx context.Context, device string, logPath string) *exec.Cmd {
+		gotDevice = device
+		gotLogPath = logPath
+		return exec.CommandContext(ctx, "sh", "-c", "printf 'line1\nline2\n'")
+	}
+	defer func() { installCommand = orig }()
+
+	q.runTask(tk, j, context.Background())
+
+	if gotDevice != "/dev/sda" {
+		t.Fatalf("device=%q want /dev/sda", gotDevice)
+	}
+	if gotLogPath == "" {
+		t.Fatal("expected install log path")
+	}
+	logs := strings.Join(j.lines, "\n")
+	if !strings.Contains(logs, "Install log: ") {
+		t.Fatalf("missing install log line: %v", j.lines)
+	}
+	if !strings.Contains(logs, "line1") || !strings.Contains(logs, "line2") {
+		t.Fatalf("missing streamed output: %v", j.lines)
+	}
+	if j.err != "" {
+		t.Fatalf("unexpected error: %q", j.err)
+	}
+}

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/architecture/charting.md

@@ -9,6 +9,34 @@ All live metrics charts in the web UI are server-side SVG images served by Go
 and polled by the browser every 2 seconds via `<img src="...?t=now">`.
 There is no client-side canvas or JS chart library.
 
+## Rule: live charts must be visually uniform
+
+Live charts are a single UI family, not a set of one-off widgets. New charts and
+changes to existing charts must keep the same rendering model and presentation
+rules unless there is an explicit architectural decision to diverge.
+
+Default expectations:
+
+- same server-side SVG pipeline for all live metrics charts
+- same refresh behaviour and failure handling in the browser
+- same canvas size class and card layout
+- same legend placement policy across charts
+- same axis, title, and summary conventions
+- no chart-specific visual exceptions added as a quick fix
+
+Current default for live charts:
+
+- legend below the plot area when a chart has 8 series or fewer
+- legend hidden when a chart has more than 8 series
+- 10 equal Y-axis steps across the chart height
+- 1400 x 360 SVG canvas with legend
+- 1400 x 288 SVG canvas without legend
+- full-width card rendering in a single-column stack
+
+If one chart needs a different layout or legend behaviour, treat that as a
+design-level decision affecting the whole chart family, not as a local tweak to
+just one endpoint.
+
 ### Why go-analyze/charts
 
 - Pure Go, no CGO — builds cleanly inside the live-build container
@@ -29,7 +57,8 @@ self-contained SVG renderer used **only** for completed SAT run reports
 | `GET /api/metrics/chart/server.svg` | CPU temp, CPU load %, mem load %, power W, fan RPMs |
 | `GET /api/metrics/chart/gpu/{idx}.svg` | GPU temp °C, load %, mem %, power W |
 
-Charts are 1400 × 280 px SVG. The page renders them at `width: 100%` in a
+Charts are 1400 × 360 px SVG when the legend is shown, and 1400 × 288 px when
+the legend is hidden. The page renders them at `width: 100%` in a
 single-column layout so they always fill the viewport width.
 
 ### Ring buffers

bible-local/architecture/runtime-flows.md

@@ -60,6 +60,8 @@ Rules:
 - Chromium opens `http://localhost/` — the full interactive web UI
 - SSH is independent from the desktop path
 - serial console support is enabled for VM boot debugging
+- Default boot keeps the server-safe graphics path (`nomodeset` + forced `fbdev`) for IPMI/BMC consoles
+- Higher-resolution mode selection is expected only when booting through an explicit `bee.display=kms` menu entry, which disables the forced `fbdev` Xorg config before `lightdm`
 
 ## ISO build sequence
 
@@ -81,9 +83,9 @@ build-in-container.sh [--authorized-keys /path/to/keys]
   7. `build-cublas.sh`:
        a. download `libcublas`, `libcublasLt`, `libcudart` runtime + dev packages from the NVIDIA CUDA Debian repo
        b. verify packages against repo `Packages.gz`
-       c. extract headers for `bee-gpu-stress` build
+       c. extract headers for `bee-gpu-burn` worker build
        d. cache userspace libs in `dist/cublas-<version>+cuda<series>/`
-  8. build `bee-gpu-stress` against extracted cuBLASLt/cudart headers
+  8. build `bee-gpu-burn` worker against extracted cuBLASLt/cudart headers
   9. inject NVIDIA `.ko` → staged `/usr/local/lib/nvidia/`
   10. inject `nvidia-smi` → staged `/usr/local/bin/nvidia-smi`
   11. inject `libnvidia-ml` + `libcuda` + `libcublas` + `libcublasLt` + `libcudart` → staged `/usr/lib/`
@@ -104,7 +106,7 @@ Build host notes:
   1. `build-in-container.sh` / `build-nvidia-module.sh` — Debian kernel headers for module build
   2. `auto/config` — `linux-image-${DEBIAN_KERNEL_ABI}` in the ISO
 - NVIDIA modules go to staged `usr/local/lib/nvidia/` — NOT to `/lib/modules/<kver>/extra/`.
-- `bee-gpu-stress` must be built against cached CUDA userspace headers from `build-cublas.sh`, not against random host-installed CUDA headers.
+- `bee-gpu-burn` worker must be built against cached CUDA userspace headers from `build-cublas.sh`, not against random host-installed CUDA headers.
 - The live ISO must ship `libcublas`, `libcublasLt`, and `libcudart` together with `libcuda` so tensor-core stress works without internet or package installs at boot.
 - The source overlay in `iso/overlay/` is treated as immutable source. Build-time files are injected only into the staged overlay.
 - The live-build workdir under `dist/` is disposable; source files under `iso/builder/` stay clean.
@@ -153,18 +155,17 @@ Current validation state:
 Every collector returns `nil, nil` on tool-not-found. Errors are logged, never fatal.
 
 Acceptance flows:
-- `bee sat nvidia` → diagnostic archive with `nvidia-smi -q` + `nvidia-bug-report` + mixed-precision `bee-gpu-stress`
+- `bee sat nvidia` → diagnostic archive with `nvidia-smi -q` + `nvidia-bug-report` + lightweight `bee-gpu-burn`
+- NVIDIA GPU burn-in can use either `bee-gpu-burn` or `bee-john-gpu-stress` (John the Ripper jumbo via OpenCL)
 - `bee sat memory` → `memtester` archive
 - `bee sat storage` → SMART/NVMe diagnostic archive and short self-test trigger where supported
 - SAT `summary.txt` now includes `overall_status` and per-job `*_status` values (`OK`, `FAILED`, `UNSUPPORTED`)
-- `bee-gpu-stress` should prefer cuBLASLt GEMM load over the old integer/PTX burn path:
+- `bee-gpu-burn` should prefer cuBLASLt GEMM load over the old integer/PTX burn path:
   - Ampere: `fp16` + `fp32`/TF32 tensor-core load
   - Ada / Hopper: add `fp8`
   - Blackwell+: add `fp4`
   - PTX fallback is only for missing cuBLASLt/userspace or unsupported narrow datatypes
 - Runtime overrides:
-  - `BEE_GPU_STRESS_SECONDS`
-  - `BEE_GPU_STRESS_SIZE_MB`
   - `BEE_MEMTESTER_SIZE_MB`
   - `BEE_MEMTESTER_PASSES`
 
@@ -179,6 +180,6 @@ Web UI: Acceptance Tests page → Run Test button
 ```
 
 **Critical invariants:**
-- `bee-gpu-stress` uses `exec.CommandContext` — killed on job context cancel.
+- `bee-gpu-burn` / `bee-john-gpu-stress` use `exec.CommandContext` — killed on job context cancel.
 - Metric goroutine uses stopCh/doneCh pattern; main goroutine waits `<-doneCh` before reading rows (no mutex needed).
 - SVG chart is fully offline: no JS, no external CSS, pure inline SVG.

bible-local/architecture/system-overview.md

@@ -21,8 +21,8 @@ Fills gaps where Redfish/logpile is blind:
 - Read-only hardware inventory: board, CPU, memory, storage, PCIe, PSU, GPU, NIC, RAID
 - Machine-readable health summary derived from collector verdicts
 - Operator-triggered acceptance tests for NVIDIA, memory, and storage
-- NVIDIA SAT includes both diagnostic collection and mixed-precision GPU stress via `bee-gpu-stress`
-- `bee-gpu-stress` should exercise tensor/inference paths (`fp16`, `fp32`/TF32, `fp8`, `fp4` when supported by the GPU/userspace stack) and fall back to Driver API PTX burn only if cuBLASLt is unavailable
+- NVIDIA SAT includes diagnostic collection plus a lightweight in-image GPU stress step via `bee-gpu-burn`
+- `bee-gpu-burn` should exercise tensor/inference paths (`fp16`, `fp32`/TF32, `fp8`, `fp4` when supported by the GPU/userspace stack) and fall back to Driver API PTX burn only if cuBLASLt is unavailable
 - Automatic boot audit with operator-facing local console and SSH access
 - NVIDIA proprietary driver loaded at boot for GPU enrichment via `nvidia-smi`
 - SSH access (OpenSSH) always available for inspection and debugging
@@ -70,7 +70,7 @@ Fills gaps where Redfish/logpile is blind:
 | SSH | OpenSSH server |
 | NVIDIA driver | Proprietary `.run` installer, built against Debian kernel headers |
 | NVIDIA modules | Loaded via `insmod` from `/usr/local/lib/nvidia/` |
-| GPU stress backend | `bee-gpu-stress` + cuBLASLt/cuBLAS/cudart mixed-precision GEMM, with Driver API PTX fallback |
+| GPU stress backend | `bee-gpu-burn` + cuBLASLt/cuBLAS/cudart mixed-precision GEMM, with Driver API PTX fallback |
 | Builder | Debian 12 host/VM or Debian 12 container image |
 
 ## Operator UX

bible-local/decisions/2026-03-05-nvidia-proprietary-driver.md

@@ -18,6 +18,8 @@ Use the official proprietary NVIDIA `.run` installer for both kernel modules and
 - Kernel modules and nvidia-smi come from a single verified source.
 - NVIDIA publishes `.sha256sum` alongside each installer — download and verify before use.
 - Driver version pinned in `iso/builder/VERSIONS` as `NVIDIA_DRIVER_VERSION`.
+- DCGM must track the CUDA user-mode driver major version exposed by `nvidia-smi`.
+- For NVIDIA driver branch `590` with CUDA `13.x`, use DCGM 4 package family `datacenter-gpu-manager-4-cuda13`; legacy `datacenter-gpu-manager` 3.x does not provide a working path for this stack.
 - Build process: download `.run`, extract, compile `kernel/` sources against `linux-lts-dev`.
 - Modules cached in `dist/nvidia-<version>-<kver>/` — rebuild only on version or kernel change.
 - ISO size increases by ~50MB for .ko files + nvidia-smi.

bible-local/decisions/2026-04-01-memtest-build-strategy.md

@@ -0,0 +1,224 @@
+# Decision: Treat memtest as explicit ISO content, not as trusted live-build magic
+
+**Date:** 2026-04-01
+**Status:** resolved
+
+## Context
+
+We have already iterated on `memtest` multiple times and kept cycling between the same ideas.
+The commit history shows several distinct attempts:
+
+- `f91bce8` — fixed Bookworm memtest file names to `memtest86+x64.bin` / `memtest86+x64.efi`
+- `5857805` — added a binary hook to copy memtest files from the build tree into the ISO root
+- `f96b149` — added fallback extraction from the cached `.deb` when `chroot/boot/` stayed empty
+- `d43a9ae` — removed the custom hook and switched back to live-build built-in memtest integration
+- `60cb8f8` — restored explicit memtest menu entries and added ISO validation
+- `3dbc218` / `3869788` — added archived build logs and better memtest diagnostics
+
+Current evidence from the archived `easy-bee-nvidia-v3.14-amd64` logs dated 2026-04-01:
+
+- `lb binary_memtest` does run and installs `memtest86+`
+- but the final ISO still does **not** contain `boot/memtest86+x64.bin`
+- the final ISO also does **not** contain memtest menu entries in `boot/grub/grub.cfg` or `isolinux/live.cfg`
+
+So the assumption "live-build built-in memtest integration is enough on this stack" is currently false for this project until proven otherwise by a real built ISO.
+
+Additional evidence from the archived `easy-bee-nvidia-v3.17-dirty-amd64` logs dated 2026-04-01:
+
+- the build now completes successfully because memtest is non-blocking by default
+- `lb binary_memtest` still runs and installs `memtest86+`
+- the project-owned hook `config/hooks/normal/9100-memtest.hook.binary` does execute
+- but it executes too early for its current target paths:
+  - `binary/boot/grub/grub.cfg` is still missing at hook time
+  - `binary/isolinux/live.cfg` is still missing at hook time
+  - memtest binaries are also still absent in `binary/boot/`
+- later in the build, live-build does create intermediate bootloader configs with memtest lines in the workdir
+- but the final ISO still lacks memtest binaries and still lacks memtest lines in extracted ISO `boot/grub/grub.cfg` and `isolinux/live.cfg`
+
+So the assumption "the current normal binary hook path is late enough to patch final memtest artifacts" is also false.
+
+Correction after inspecting the real `easy-bee-nvidia-v3.20-5-g76a9100-amd64.iso`
+artifact dated 2026-04-01:
+
+- the final ISO does contain `boot/memtest86+x64.bin`
+- the final ISO does contain `boot/memtest86+x64.efi`
+- the final ISO does contain memtest menu entries in both `boot/grub/grub.cfg`
+  and `isolinux/live.cfg`
+- so `v3.20-5-g76a9100` was **not** another real memtest regression in the
+  shipped ISO
+- the regression was in the build-time validator/debug path in `build.sh`
+
+Root cause of the false alarm:
+
+- `build.sh` treated "ISO reader command exists" as equivalent to "ISO reader
+  successfully listed/extracted members"
+- `iso_list_files` / `iso_extract_file` failures were collapsed into the same
+  observable output as "memtest content missing"
+- this made a reader failure look identical to a missing memtest payload
+- as a result, we re-entered the same memtest investigation loop even though
+  the real ISO was already correct
+
+Additional correction from the subsequent `v3.21` build logs dated 2026-04-01:
+
+- once ISO reading was fixed, the post-build debug correctly showed the raw ISO
+  still carried live-build's default memtest layout (`live/memtest.bin`,
+  `live/memtest.efi`, `boot/grub/memtest.cfg`, `isolinux/memtest.cfg`)
+- that mismatch is expected to trigger project recovery, because `bee` requires
+  `boot/memtest86+x64.bin` / `boot/memtest86+x64.efi` plus matching menu paths
+- however, `build.sh` exited before recovery because `set -e` treated a direct
+  `iso_memtest_present` return code of `1` as fatal
+- so the next repeated loop was caused by shell control flow, not by proof that
+  the recovery design itself was wrong
+
+## Known Failed Attempts
+
+These approaches were already tried and should not be repeated blindly:
+
+1. Built-in live-build memtest only.
+Reason it failed:
+- `lb binary_memtest` runs, but the final ISO still misses memtest binaries and menu entries.
+
+2. Fixing only the memtest file names for Debian Bookworm.
+Reason it failed:
+- correct file names alone do not make the files appear in the final ISO.
+
+3. Copying memtest from `chroot/boot/` into `binary/boot/` via a binary hook.
+Reason it failed:
+- in this stack `chroot/boot/` is often empty for memtest payloads at the relevant time.
+
+4. Fallback extraction from cached `memtest86+` `.deb`.
+Reason it failed:
+- this was explored already and was not enough to stabilize the final ISO path end-to-end.
+
+5. Restoring explicit memtest menu entries in source bootloader templates only.
+Reason it failed:
+- memtest lines in source templates or intermediate workdir configs do not guarantee the final ISO contains them.
+
+6. Patching `binary/boot/grub/grub.cfg` and `binary/isolinux/live.cfg` from the current `config/hooks/normal/9100-memtest.hook.binary`.
+Reason it failed:
+- the hook runs before those files exist, so the hook cannot patch them there.
+
+## What This Means
+
+When revisiting memtest later, start from the constraints above rather than retrying the same patterns:
+
+- do not assume the built-in memtest stage is sufficient
+- do not assume `chroot/boot/` will contain memtest payloads
+- do not assume source bootloader templates are the last writer of final ISO configs
+- do not assume the current normal binary hook timing is late enough for final patching
+
+Any future memtest fix must explicitly identify:
+
+- where the memtest binaries are reliably available at build time
+- which exact build stage writes the final bootloader configs that land in the ISO
+- and a post-build proof from a real ISO, not only from intermediate workdir files
+- whether the ISO inspection step itself succeeded, rather than merely whether
+  the validator printed a memtest warning
+- whether a non-zero probe is intentionally handled inside an `if` / `case`
+  context rather than accidentally tripping `set -e`
+
+## Decision
+
+For `bee`, memtest must be treated as an explicit ISO artifact with explicit post-build validation.
+
+Project rules from now on:
+
+- Do **not** trust `--memtest memtest86+` by itself.
+- A memtest implementation is considered valid only if the produced ISO actually contains:
+  - `boot/memtest86+x64.bin`
+  - `boot/memtest86+x64.efi`
+  - a GRUB menu entry
+  - an isolinux menu entry
+- If live-build built-in integration does not produce those artifacts, use an explicit project-owned mechanism such as:
+  - a binary hook copying files into `binary/boot/`
+  - extraction from the cached `memtest86+` `.deb`
+  - another deterministic build-time copy step
+- Do **not** remove such explicit logic later unless a fresh real ISO build proves that built-in integration alone produces all required files and menu entries.
+
+Current implementation direction:
+
+- keep the live-build memtest stage enabled if it helps package acquisition
+- do not rely on the current early `binary_hooks` timing for final patching
+- prefer a post-`lb build` recovery step in `build.sh` that:
+  - patches the fully materialized `LB_DIR/binary` tree
+  - injects memtest binaries there
+  - ensures final bootloader entries there
+  - reruns late binary stages (`binary_checksums`, `binary_iso`, `binary_zsync`) after the patch
+- also treat ISO validation tooling as part of the critical path:
+  - install a stable ISO reader in the builder image
+  - fail with an explicit reader error if ISO listing/extraction fails
+  - do not treat reader failure as evidence that memtest is missing
+  - do not call a probe that may return "needs recovery" as a bare command under
+    `set -e`; wrap it in explicit control flow
+
+## Consequences
+
+- Future memtest changes must begin by reading this ADR and the commits listed above.
+- Future memtest changes must also begin by reading the failed-attempt list above.
+- We should stop re-introducing "prefer built-in live-build memtest" as a default assumption without new evidence.
+- Memtest validation in `build.sh` is not optional; it is the acceptance gate that prevents another silent regression.
+- But validation output is only trustworthy if ISO reading itself succeeded. A
+  "missing memtest" warning without a successful ISO read is not evidence.
+- If we change memtest strategy again, we must update this ADR with the exact build evidence that justified the change.
+
+## Working Solution (confirmed 2026-04-01, commits 76a9100 → 2baf3be)
+
+This approach was confirmed working in ISO `easy-bee-nvidia-v3.20-5-g76a9100-amd64.iso`
+and validated again in subsequent builds. The final ISO contains all required memtest artifacts.
+
+### Components
+
+**1. Binary hook `config/hooks/normal/9100-memtest.hook.binary`**
+
+Runs inside the live-build binary phase. Does not patch bootloader files at hook time —
+those files may not exist yet. Instead:
+
+- Tries to copy `memtest86+x64.bin` / `memtest86+x64.efi` from `chroot/boot/` first.
+- Falls back to extracting from the cached `.deb` (via `dpkg-deb -x`) if `chroot/boot/` is empty.
+- Appends GRUB and isolinux menu entries only if the respective cfg files already exist at hook time.
+  If they do not exist, the hook warns and continues (does not fail).
+
+Controlled by `BEE_REQUIRE_MEMTEST=1` env var to turn warnings into hard errors when needed.
+
+**2. Post-`lb build` recovery step in `build.sh`**
+
+After `lb build` completes, `build.sh` checks whether the fully materialized `binary/` tree
+contains all required memtest artifacts. If not:
+
+- Copies/extracts memtest binaries into `binary/boot/`.
+- Patches `binary/boot/grub/grub.cfg` and `binary/isolinux/live.cfg` directly.
+- Reruns the late binary stages (`binary_checksums`, `binary_iso`, `binary_zsync`) to rebuild
+  the ISO with the patched tree.
+
+This is the deterministic safety net: even if the hook runs at the wrong time, the recovery
+step handles the final `binary/` tree after live-build has written all bootloader configs.
+
+**3. ISO validation hardening**
+
+The memtest probe in `build.sh` is wrapped in explicit `if` / `case` control flow, not called
+as a bare command under `set -e`. A non-zero probe return (needs recovery) is intentional and
+handled — it does not abort the build prematurely.
+
+ISO reading (`xorriso -indev -ls` / extraction) is treated as a separate prerequisite.
+If the reader fails, the validator reports a reader error explicitly, not a memtest warning.
+This prevents the false-negative loop that burned 2026-04-01 v3.14–v3.19.
+
+### Why this works when earlier attempts did not
+
+The earlier patterns all shared a single flaw: they assumed a single build-time point
+(hook or source template) would be the last writer of bootloader configs and memtest payloads.
+In live-build on Debian Bookworm that assumption is false — live-build continues writing
+bootloader files after custom hooks run, and `chroot/boot/` does not reliably hold memtest payloads.
+
+The recovery step sidesteps the ordering problem entirely: it acts on the fully materialized
+`binary/` tree after `lb build` finishes, then rebuilds the ISO from that patched tree.
+There is no ordering dependency to get wrong.
+
+### Do not revert
+
+Do not remove the recovery step or the hook without a fresh real ISO build proving
+live-build alone produces all four required artifacts:
+- `boot/memtest86+x64.bin`
+- `boot/memtest86+x64.efi`
+- memtest entry in `boot/grub/grub.cfg`
+- memtest entry in `isolinux/live.cfg`

bible-local/decisions/README.md

@@ -5,3 +5,4 @@ One file per decision, named `YYYY-MM-DD-short-topic.md`.
 | Date | Decision | Status |
 |---|---|---|
 | 2026-03-05 | Use NVIDIA proprietary driver | active |
+| 2026-04-01 | Treat memtest as explicit ISO content | active |

bible-local/docs/iso-build-rules.md

@@ -13,9 +13,50 @@ Use one of:
 
 This applies to:
 - `iso/builder/config/package-lists/*.list.chroot`
-- Any package referenced in `grub.cfg`, hooks, or overlay scripts (e.g. file paths like `/boot/memtest86+x64.bin`)
+- Any package referenced in bootloader configs, hooks, or overlay scripts
 
-## Example of what goes wrong without this
+## Memtest rule
 
-`memtest86+` in Debian bookworm installs `/boot/memtest86+x64.bin`, not `/boot/memtest86+.bin`.
-Guessing the filename caused a broken GRUB entry that only surfaced at boot time, after a full rebuild.
+Do not assume live-build's built-in memtest integration is sufficient for `bee`.
+We already tried that path and regressed again on 2026-04-01: `lb binary_memtest`
+ran, but the final ISO still lacked memtest binaries and menu entries.
+
+For this project, memtest is accepted only when the produced ISO actually
+contains all of the following:
+
+- `boot/memtest86+x64.bin`
+- `boot/memtest86+x64.efi`
+- a memtest entry in `boot/grub/grub.cfg`
+- a memtest entry in `isolinux/live.cfg`
+
+Rules:
+
+- Keep explicit post-build memtest validation in `build.sh`.
+- Treat ISO reader success as a separate prerequisite from memtest content.
+  If the reader cannot list or extract from the ISO, that is a validator
+  failure, not proof that memtest is missing.
+- If built-in integration does not produce the artifacts above, use a
+  deterministic project-owned copy/extract step instead of hoping live-build
+  will "start working".
+- Do not switch back to built-in-only memtest without fresh build evidence from
+  a real ISO.
+- If you reference memtest files manually, verify the exact package file list
+  first for the target Debian release.
+
+Known bad loops for this repository:
+
+- Do not retry built-in-only memtest without new evidence. We already proved
+  that `lb binary_memtest` can run while the final ISO still has no memtest.
+- Do not assume fixing memtest file names is enough. Correct names did not fix
+  the final artifact path.
+- Do not assume `chroot/boot/` contains memtest payloads at the time hooks run.
+- Do not assume source `grub.cfg` / `live.cfg.in` are the final writers of ISO
+  bootloader configs.
+- Do not assume the current `config/hooks/normal/9100-memtest.hook.binary`
+  timing is late enough to patch final `binary/boot/grub/grub.cfg` or
+  `binary/isolinux/live.cfg`; logs from 2026-04-01 showed those files were not
+  present yet when the hook executed.
+- Do not treat a validator warning as ground truth until you have confirmed the
+  ISO reader actually succeeded. On 2026-04-01 we misdiagnosed another memtest
+  regression because the final ISO was correct but the validator produced a
+  false negative.

iso/README.md

@@ -48,6 +48,7 @@ sh iso/builder/build-in-container.sh --cache-dir /path/to/cache
 - The builder image is automatically rebuilt if the local tag exists for the wrong architecture.
 - The live ISO boots with Debian `live-boot` `toram`, so the read-only medium is copied into RAM during boot and the runtime no longer depends on the original USB/BMC virtual media staying present.
 - Target systems need enough RAM for the full compressed live medium plus normal runtime overhead, or boot may fail before reaching the TUI.
+- The NVIDIA variant installs DCGM 4 packages matched to the CUDA user-mode driver major version. For driver branch `590` / CUDA `13.x`, the package family is `datacenter-gpu-manager-4-cuda13` rather than legacy `datacenter-gpu-manager`.
 - Override the container platform only if you know why:
 
 ```sh

iso/builder/Dockerfile

@@ -17,12 +17,23 @@ RUN apt-get update -qq && apt-get install -y \
     wget \
     curl \
     tar \
+    libarchive-tools \
     xz-utils \
     rsync \
     build-essential \
     gcc \
     make \
     perl \
+    pkg-config \
+    yasm \
+    libssl-dev \
+    zlib1g-dev \
+    libbz2-dev \
+    libgmp-dev \
+    libpcap-dev \
+    libsqlite3-dev \
+    libcurl4-openssl-dev \
+    ocl-icd-opencl-dev \
     linux-headers-amd64 \
     && rm -rf /var/lib/apt/lists/*
 

iso/builder/VERSIONS

@@ -8,7 +8,8 @@ NCCL_TESTS_VERSION=2.13.10
 NVCC_VERSION=12.8
 CUBLAS_VERSION=13.0.2.14-1
 CUDA_USERSPACE_VERSION=13.0.96-1
-DCGM_VERSION=3.3.9
+DCGM_VERSION=4.5.3-1
+JOHN_JUMBO_COMMIT=67fcf9fe5a
 ROCM_VERSION=6.3.4
 ROCM_SMI_VERSION=7.4.0.60304-76~22.04
 ROCM_BANDWIDTH_TEST_VERSION=1.4.0.60304-76~22.04

iso/builder/auto/config

@@ -29,10 +29,10 @@ lb config noauto \
     --security true \
     --linux-flavours "amd64" \
     --linux-packages "${LB_LINUX_PACKAGES}" \
-    --memtest none \
-    --iso-volume "EASY-BEE-${BEE_GPU_VENDOR_UPPER:-NVIDIA}" \
+    --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=7 username=bee user-fullname=Bee modprobe.blacklist=nouveau" \
+    --bootappend-live "boot=live components video=1920x1080 console=tty0 console=ttyS0,115200n8 loglevel=6 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

@@ -29,8 +29,14 @@ typedef void *CUfunction;
 typedef void *CUstream;
 
 #define CU_SUCCESS 0
+#define CU_DEVICE_ATTRIBUTE_MULTIPROCESSOR_COUNT 16
 #define CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MAJOR 75
 #define CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MINOR 76
+#define MAX_STRESS_STREAMS 16
+#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"
@@ -97,6 +103,9 @@ typedef CUresult (*cuLaunchKernel_fn)(CUfunction,
                                       CUstream,
                                       void **,
                                       void **);
+typedef CUresult (*cuMemGetInfo_fn)(size_t *, size_t *);
+typedef CUresult (*cuStreamCreate_fn)(CUstream *, unsigned int);
+typedef CUresult (*cuStreamDestroy_fn)(CUstream);
 typedef CUresult (*cuGetErrorName_fn)(CUresult, const char **);
 typedef CUresult (*cuGetErrorString_fn)(CUresult, const char **);
 
@@ -118,6 +127,9 @@ struct cuda_api {
     cuModuleLoadDataEx_fn cuModuleLoadDataEx;
     cuModuleGetFunction_fn cuModuleGetFunction;
     cuLaunchKernel_fn cuLaunchKernel;
+    cuMemGetInfo_fn cuMemGetInfo;
+    cuStreamCreate_fn cuStreamCreate;
+    cuStreamDestroy_fn cuStreamDestroy;
     cuGetErrorName_fn cuGetErrorName;
     cuGetErrorString_fn cuGetErrorString;
 };
@@ -128,9 +140,10 @@ struct stress_report {
     int cc_major;
     int cc_minor;
     int buffer_mb;
+    int stream_count;
     unsigned long iterations;
     uint64_t checksum;
-    char details[1024];
+    char details[16384];
 };
 
 static int load_symbol(void *lib, const char *name, void **out) {
@@ -144,7 +157,7 @@ static int load_cuda(struct cuda_api *api) {
     if (!api->lib) {
         return 0;
     }
-    return
+    if (!(
         load_symbol(api->lib, "cuInit", (void **)&api->cuInit) &&
         load_symbol(api->lib, "cuDeviceGetCount", (void **)&api->cuDeviceGetCount) &&
         load_symbol(api->lib, "cuDeviceGet", (void **)&api->cuDeviceGet) &&
@@ -160,7 +173,17 @@ static int load_cuda(struct cuda_api *api) {
         load_symbol(api->lib, "cuMemcpyDtoH_v2", (void **)&api->cuMemcpyDtoH) &&
         load_symbol(api->lib, "cuModuleLoadDataEx", (void **)&api->cuModuleLoadDataEx) &&
         load_symbol(api->lib, "cuModuleGetFunction", (void **)&api->cuModuleGetFunction) &&
-        load_symbol(api->lib, "cuLaunchKernel", (void **)&api->cuLaunchKernel);
+        load_symbol(api->lib, "cuLaunchKernel", (void **)&api->cuLaunchKernel))) {
+        dlclose(api->lib);
+        memset(api, 0, sizeof(*api));
+        return 0;
+    }
+    load_symbol(api->lib, "cuMemGetInfo_v2", (void **)&api->cuMemGetInfo);
+    load_symbol(api->lib, "cuStreamCreate", (void **)&api->cuStreamCreate);
+    if (!load_symbol(api->lib, "cuStreamDestroy_v2", (void **)&api->cuStreamDestroy)) {
+        load_symbol(api->lib, "cuStreamDestroy", (void **)&api->cuStreamDestroy);
+    }
+    return 1;
 }
 
 static const char *cu_error_name(struct cuda_api *api, CUresult rc) {
@@ -193,14 +216,12 @@ static double now_seconds(void) {
     return (double)ts.tv_sec + ((double)ts.tv_nsec / 1000000000.0);
 }
 
-#if HAVE_CUBLASLT_HEADERS
 static size_t round_down_size(size_t value, size_t multiple) {
     if (multiple == 0 || value < multiple) {
         return value;
     }
     return value - (value % multiple);
 }
-#endif
 
 static int query_compute_capability(struct cuda_api *api, CUdevice dev, int *major, int *minor) {
     int cc_major = 0;
@@ -220,6 +241,75 @@ static int query_compute_capability(struct cuda_api *api, CUdevice dev, int *maj
     return 1;
 }
 
+static int query_multiprocessor_count(struct cuda_api *api, CUdevice dev, int *count) {
+    int mp_count = 0;
+    if (!check_rc(api,
+                  "cuDeviceGetAttribute(multiprocessors)",
+                  api->cuDeviceGetAttribute(&mp_count, CU_DEVICE_ATTRIBUTE_MULTIPROCESSOR_COUNT, dev))) {
+        return 0;
+    }
+    *count = mp_count;
+    return 1;
+}
+
+static size_t clamp_budget_to_free_memory(struct cuda_api *api, size_t requested_bytes) {
+    size_t free_bytes = 0;
+    size_t total_bytes = 0;
+    size_t max_bytes = requested_bytes;
+
+    if (!api->cuMemGetInfo) {
+        return requested_bytes;
+    }
+    if (api->cuMemGetInfo(&free_bytes, &total_bytes) != CU_SUCCESS || free_bytes == 0) {
+        return requested_bytes;
+    }
+
+    max_bytes = (free_bytes * 9u) / 10u;
+    if (max_bytes < (size_t)4u * 1024u * 1024u) {
+        max_bytes = (size_t)4u * 1024u * 1024u;
+    }
+    if (requested_bytes > max_bytes) {
+        return max_bytes;
+    }
+    return requested_bytes;
+}
+
+static int choose_stream_count(int mp_count, int planned_profiles, size_t total_budget, int have_streams) {
+    int stream_count = 1;
+    if (!have_streams || mp_count <= 0 || planned_profiles <= 0) {
+        return 1;
+    }
+
+    stream_count = mp_count / 8;
+    if (stream_count < 2) {
+        stream_count = 2;
+    }
+    if (stream_count > MAX_STRESS_STREAMS) {
+        stream_count = MAX_STRESS_STREAMS;
+    }
+
+    while (stream_count > 1) {
+        size_t per_stream_budget = total_budget / ((size_t)planned_profiles * (size_t)stream_count);
+        if (per_stream_budget >= MIN_STREAM_BUDGET_BYTES) {
+            break;
+        }
+        stream_count--;
+    }
+    return stream_count;
+}
+
+static void destroy_streams(struct cuda_api *api, CUstream *streams, int count) {
+    if (!api->cuStreamDestroy) {
+        return;
+    }
+    for (int i = 0; i < count; i++) {
+        if (streams[i]) {
+            api->cuStreamDestroy(streams[i]);
+            streams[i] = NULL;
+        }
+    }
+}
+
 #if HAVE_CUBLASLT_HEADERS
 static void append_detail(char *buf, size_t cap, const char *fmt, ...) {
     size_t len = strlen(buf);
@@ -242,12 +332,19 @@ static int run_ptx_fallback(struct cuda_api *api,
                             int size_mb,
                             struct stress_report *report) {
     CUcontext ctx = NULL;
-    CUdeviceptr device_mem = 0;
     CUmodule module = NULL;
     CUfunction kernel = NULL;
     uint32_t sample[256];
-    uint32_t words = 0;
+    CUdeviceptr device_mem[MAX_STRESS_STREAMS] = {0};
+    CUstream streams[MAX_STRESS_STREAMS] = {0};
+    uint32_t words[MAX_STRESS_STREAMS] = {0};
+    uint32_t rounds[MAX_STRESS_STREAMS] = {0};
+    void *params[MAX_STRESS_STREAMS][3];
+    size_t bytes_per_stream[MAX_STRESS_STREAMS] = {0};
     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");
@@ -260,64 +357,109 @@ static int run_ptx_fallback(struct cuda_api *api,
         return 0;
     }
 
-    size_t bytes = (size_t)size_mb * 1024u * 1024u;
-    if (bytes < 4u * 1024u * 1024u) {
-        bytes = 4u * 1024u * 1024u;
+    size_t requested_bytes = (size_t)size_mb * 1024u * 1024u;
+    if (requested_bytes < MIN_PROFILE_BUDGET_BYTES) {
+        requested_bytes = MIN_PROFILE_BUDGET_BYTES;
     }
-    if (bytes > (size_t)1024u * 1024u * 1024u) {
-        bytes = (size_t)1024u * 1024u * 1024u;
+    size_t total_bytes = clamp_budget_to_free_memory(api, requested_bytes);
+    if (total_bytes < MIN_PROFILE_BUDGET_BYTES) {
+        total_bytes = MIN_PROFILE_BUDGET_BYTES;
     }
-    words = (uint32_t)(bytes / sizeof(uint32_t));
+    report->buffer_mb = (int)(total_bytes / (1024u * 1024u));
 
-    if (!check_rc(api, "cuMemAlloc", api->cuMemAlloc(&device_mem, bytes))) {
-        api->cuCtxDestroy(ctx);
-        return 0;
+    if (query_multiprocessor_count(api, dev, &mp_count) &&
+        api->cuStreamCreate &&
+        api->cuStreamDestroy) {
+        stream_count = choose_stream_count(mp_count, 1, total_bytes, 1);
     }
-    if (!check_rc(api, "cuMemsetD8", api->cuMemsetD8(device_mem, 0, bytes))) {
-        api->cuMemFree(device_mem);
-        api->cuCtxDestroy(ctx);
-        return 0;
+    if (stream_count > 1) {
+        int created = 0;
+        for (; created < stream_count; created++) {
+            if (!check_rc(api, "cuStreamCreate", api->cuStreamCreate(&streams[created], 0))) {
+                destroy_streams(api, streams, created);
+                stream_count = 1;
+                break;
+            }
+        }
     }
+    report->stream_count = stream_count;
+
+    for (int lane = 0; lane < stream_count; lane++) {
+        size_t slice = total_bytes / (size_t)stream_count;
+        if (lane == stream_count - 1) {
+            slice = total_bytes - ((size_t)lane * (total_bytes / (size_t)stream_count));
+        }
+        slice = round_down_size(slice, sizeof(uint32_t));
+        if (slice < MIN_PROFILE_BUDGET_BYTES) {
+            slice = MIN_PROFILE_BUDGET_BYTES;
+        }
+        bytes_per_stream[lane] = slice;
+        words[lane] = (uint32_t)(slice / sizeof(uint32_t));
+
+        if (!check_rc(api, "cuMemAlloc", api->cuMemAlloc(&device_mem[lane], slice))) {
+            goto fail;
+        }
+        if (!check_rc(api, "cuMemsetD8", api->cuMemsetD8(device_mem[lane], 0, slice))) {
+            goto fail;
+        }
+        rounds[lane] = 2048;
+        params[lane][0] = &device_mem[lane];
+        params[lane][1] = &words[lane];
+        params[lane][2] = &rounds[lane];
+    }
+
     if (!check_rc(api,
                   "cuModuleLoadDataEx",
                   api->cuModuleLoadDataEx(&module, ptx_source, 0, NULL, NULL))) {
-        api->cuMemFree(device_mem);
-        api->cuCtxDestroy(ctx);
-        return 0;
+        goto fail;
     }
     if (!check_rc(api, "cuModuleGetFunction", api->cuModuleGetFunction(&kernel, module, "burn"))) {
-        api->cuMemFree(device_mem);
-        api->cuCtxDestroy(ctx);
-        return 0;
+        goto fail;
     }
 
     unsigned int threads = 256;
-    unsigned int blocks = (unsigned int)((words + threads - 1) / threads);
-    uint32_t rounds = 1024;
-    void *params[] = {&device_mem, &words, &rounds};
 
     double start = now_seconds();
     double deadline = start + (double)seconds;
     while (now_seconds() < deadline) {
-        if (!check_rc(api,
-                      "cuLaunchKernel",
-                      api->cuLaunchKernel(kernel, blocks, 1, 1, threads, 1, 1, 0, NULL, params, NULL))) {
-            api->cuMemFree(device_mem);
-            api->cuCtxDestroy(ctx);
-            return 0;
+        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;
+            }
         }
-        iterations++;
+        if (launches_per_wave <= 0) {
+            goto fail;
+        }
+        if (!check_rc(api, "cuCtxSynchronize", api->cuCtxSynchronize())) {
+            goto fail;
+        }
+        iterations += (unsigned long)launches_per_wave;
     }
 
-    if (!check_rc(api, "cuCtxSynchronize", api->cuCtxSynchronize())) {
-        api->cuMemFree(device_mem);
-        api->cuCtxDestroy(ctx);
-        return 0;
-    }
-    if (!check_rc(api, "cuMemcpyDtoH", api->cuMemcpyDtoH(sample, device_mem, sizeof(sample)))) {
-        api->cuMemFree(device_mem);
-        api->cuCtxDestroy(ctx);
-        return 0;
+    if (!check_rc(api, "cuMemcpyDtoH", api->cuMemcpyDtoH(sample, device_mem[0], sizeof(sample)))) {
+        goto fail;
     }
 
     for (size_t i = 0; i < sizeof(sample) / sizeof(sample[0]); i++) {
@@ -326,12 +468,34 @@ static int run_ptx_fallback(struct cuda_api *api,
     report->iterations = iterations;
     snprintf(report->details,
              sizeof(report->details),
-             "profile_int32_fallback=OK iterations=%lu\n",
+             "fallback_int32=OK requested_mb=%d actual_mb=%d streams=%d queue_depth=%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);
 
-    api->cuMemFree(device_mem);
+    for (int lane = 0; lane < stream_count; lane++) {
+        if (device_mem[lane]) {
+            api->cuMemFree(device_mem[lane]);
+        }
+    }
+    destroy_streams(api, streams, stream_count);
     api->cuCtxDestroy(ctx);
     return 1;
+
+fail:
+    for (int lane = 0; lane < MAX_STRESS_STREAMS; lane++) {
+        if (device_mem[lane]) {
+            api->cuMemFree(device_mem[lane]);
+        }
+    }
+    destroy_streams(api, streams, MAX_STRESS_STREAMS);
+    if (ctx) {
+        api->cuCtxDestroy(ctx);
+    }
+    return 0;
 }
 
 #if HAVE_CUBLASLT_HEADERS
@@ -418,6 +582,7 @@ struct profile_desc {
 
 struct prepared_profile {
     struct profile_desc desc;
+    CUstream stream;
     cublasLtMatmulDesc_t op_desc;
     cublasLtMatrixLayout_t a_layout;
     cublasLtMatrixLayout_t b_layout;
@@ -617,8 +782,8 @@ static uint64_t choose_square_dim(size_t budget_bytes, size_t bytes_per_cell, in
     if (dim < (uint64_t)multiple) {
         dim = (uint64_t)multiple;
     }
-    if (dim > 8192u) {
-        dim = 8192u;
+    if (dim > 65536u) {
+        dim = 65536u;
     }
     return dim;
 }
@@ -704,10 +869,12 @@ static int prepare_profile(struct cublaslt_api *cublas,
                            cublasLtHandle_t handle,
                            struct cuda_api *cuda,
                            const struct profile_desc *desc,
+                           CUstream stream,
                            size_t profile_budget_bytes,
                            struct prepared_profile *out) {
     memset(out, 0, sizeof(*out));
     out->desc = *desc;
+    out->stream = stream;
 
     size_t bytes_per_cell = 0;
     bytes_per_cell += bytes_for_elements(desc->a_type, 1);
@@ -935,7 +1102,7 @@ static int run_cublas_profile(cublasLtHandle_t handle,
                                                &profile->heuristic.algo,
                                                (void *)(uintptr_t)profile->workspace_dev,
                                                profile->workspace_size,
-                                               (cudaStream_t)0));
+                                               profile->stream));
 }
 
 static int run_cublaslt_stress(struct cuda_api *cuda,
@@ -947,13 +1114,22 @@ static int run_cublaslt_stress(struct cuda_api *cuda,
                                int size_mb,
                                struct stress_report *report) {
     struct cublaslt_api cublas;
-    struct prepared_profile prepared[sizeof(k_profiles) / sizeof(k_profiles[0])];
+    struct prepared_profile prepared[MAX_STRESS_STREAMS * MAX_CUBLAS_PROFILES];
     cublasLtHandle_t handle = NULL;
     CUcontext ctx = NULL;
+    CUstream streams[MAX_STRESS_STREAMS] = {0};
     uint16_t sample[256];
     int cc = cc_major * 10 + cc_minor;
     int planned = 0;
     int active = 0;
+    int mp_count = 0;
+    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;
 
     memset(report, 0, sizeof(*report));
     snprintf(report->backend, sizeof(report->backend), "cublasLt");
@@ -986,16 +1162,46 @@ static int run_cublaslt_stress(struct cuda_api *cuda,
         return 0;
     }
 
-    size_t total_budget = (size_t)size_mb * 1024u * 1024u;
-    if (total_budget < (size_t)planned * 4u * 1024u * 1024u) {
-        total_budget = (size_t)planned * 4u * 1024u * 1024u;
+    requested_budget = (size_t)size_mb * 1024u * 1024u;
+    if (requested_budget < (size_t)planned * MIN_PROFILE_BUDGET_BYTES) {
+        requested_budget = (size_t)planned * MIN_PROFILE_BUDGET_BYTES;
     }
-    size_t per_profile_budget = total_budget / (size_t)planned;
-    if (per_profile_budget < 4u * 1024u * 1024u) {
-        per_profile_budget = 4u * 1024u * 1024u;
+    total_budget = clamp_budget_to_free_memory(cuda, requested_budget);
+    if (total_budget < (size_t)planned * MIN_PROFILE_BUDGET_BYTES) {
+        total_budget = (size_t)planned * MIN_PROFILE_BUDGET_BYTES;
     }
+    if (query_multiprocessor_count(cuda, dev, &mp_count) &&
+        cuda->cuStreamCreate &&
+        cuda->cuStreamDestroy) {
+        stream_count = choose_stream_count(mp_count, planned, total_budget, 1);
+    }
+    if (stream_count > 1) {
+        int created = 0;
+        for (; created < stream_count; created++) {
+            if (!check_rc(cuda, "cuStreamCreate", cuda->cuStreamCreate(&streams[created], 0))) {
+                destroy_streams(cuda, streams, created);
+                stream_count = 1;
+                break;
+            }
+        }
+    }
+    report->stream_count = stream_count;
+    per_profile_budget = total_budget / ((size_t)planned * (size_t)stream_count);
+    if (per_profile_budget < MIN_PROFILE_BUDGET_BYTES) {
+        per_profile_budget = MIN_PROFILE_BUDGET_BYTES;
+    }
+    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",
+                  size_mb,
+                  report->buffer_mb,
+                  report->stream_count,
+                  STRESS_LAUNCH_DEPTH,
+                  mp_count,
+                  per_profile_budget / (1024u * 1024u));
 
-    for (size_t i = 0; i < sizeof(k_profiles) / sizeof(k_profiles[0]); i++) {
+    for (int i = 0; i < profile_count; i++) {
         const struct profile_desc *desc = &k_profiles[i];
         if (!(desc->enabled && cc >= desc->min_cc)) {
             append_detail(report->details,
@@ -1005,63 +1211,87 @@ static int run_cublaslt_stress(struct cuda_api *cuda,
                           desc->min_cc);
             continue;
         }
-        if (prepare_profile(&cublas, handle, cuda, desc, per_profile_budget, &prepared[i])) {
-            active++;
-            append_detail(report->details,
-                          sizeof(report->details),
-                          "%s=READY dim=%llux%llux%llu block=%s\n",
-                          desc->name,
-                          (unsigned long long)prepared[i].m,
-                          (unsigned long long)prepared[i].n,
-                          (unsigned long long)prepared[i].k,
-                          desc->block_label);
-        } else {
-            append_detail(report->details, sizeof(report->details), "%s=SKIPPED unsupported\n", desc->name);
+        for (int lane = 0; lane < stream_count; lane++) {
+            CUstream stream = streams[lane];
+            if (prepared_count >= (int)(sizeof(prepared) / sizeof(prepared[0]))) {
+                break;
+            }
+            if (prepare_profile(&cublas, handle, cuda, desc, stream, per_profile_budget, &prepared[prepared_count])) {
+                active++;
+                append_detail(report->details,
+                              sizeof(report->details),
+                              "%s[%d]=READY dim=%llux%llux%llu block=%s stream=%d\n",
+                              desc->name,
+                              lane,
+                              (unsigned long long)prepared[prepared_count].m,
+                              (unsigned long long)prepared[prepared_count].n,
+                              (unsigned long long)prepared[prepared_count].k,
+                              desc->block_label,
+                              lane);
+                prepared_count++;
+            } else {
+                append_detail(report->details,
+                              sizeof(report->details),
+                              "%s[%d]=SKIPPED unsupported\n",
+                              desc->name,
+                              lane);
+            }
         }
     }
 
     if (active <= 0) {
         cublas.cublasLtDestroy(handle);
+        destroy_streams(cuda, streams, stream_count);
         cuda->cuCtxDestroy(ctx);
         return 0;
     }
 
     double deadline = now_seconds() + (double)seconds;
     while (now_seconds() < deadline) {
-        for (size_t i = 0; i < sizeof(prepared) / sizeof(prepared[0]); 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 (size_t j = 0; j < sizeof(prepared) / sizeof(prepared[0]); j++) {
-                    destroy_profile(&cublas, cuda, &prepared[j]);
+        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;
                 }
-                cublas.cublasLtDestroy(handle);
-                cuda->cuCtxDestroy(ctx);
-                return 0;
+                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++;
             }
-            prepared[i].iterations++;
-            report->iterations++;
-            if (now_seconds() >= deadline) {
+            if (launched_this_batch <= 0) {
                 break;
             }
         }
-    }
-
-    if (!check_rc(cuda, "cuCtxSynchronize", cuda->cuCtxSynchronize())) {
-        for (size_t i = 0; i < sizeof(prepared) / sizeof(prepared[0]); i++) {
-            destroy_profile(&cublas, cuda, &prepared[i]);
+        if (wave_launches <= 0) {
+            break;
+        }
+        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);
-        cuda->cuCtxDestroy(ctx);
-        return 0;
     }
 
-    for (size_t i = 0; i < sizeof(prepared) / sizeof(prepared[0]); i++) {
+    for (int i = 0; i < prepared_count; i++) {
         if (!prepared[i].ready) {
             continue;
         }
@@ -1072,7 +1302,7 @@ static int run_cublaslt_stress(struct cuda_api *cuda,
                       prepared[i].iterations);
     }
 
-    for (size_t i = 0; i < sizeof(prepared) / sizeof(prepared[0]); i++) {
+    for (int i = 0; i < prepared_count; i++) {
         if (prepared[i].ready) {
             if (check_rc(cuda, "cuMemcpyDtoH", cuda->cuMemcpyDtoH(sample, prepared[i].d_dev, sizeof(sample)))) {
                 for (size_t j = 0; j < sizeof(sample) / sizeof(sample[0]); j++) {
@@ -1083,10 +1313,11 @@ static int run_cublaslt_stress(struct cuda_api *cuda,
         }
     }
 
-    for (size_t i = 0; i < sizeof(prepared) / sizeof(prepared[0]); i++) {
+    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 1;
 }
@@ -1095,13 +1326,16 @@ static int run_cublaslt_stress(struct cuda_api *cuda,
 int main(int argc, char **argv) {
     int seconds = 5;
     int size_mb = 64;
+    int device_index = 0;
     for (int i = 1; i < argc; i++) {
         if ((strcmp(argv[i], "--seconds") == 0 || strcmp(argv[i], "-t") == 0) && i + 1 < argc) {
             seconds = atoi(argv[++i]);
         } else if ((strcmp(argv[i], "--size-mb") == 0 || strcmp(argv[i], "-m") == 0) && i + 1 < argc) {
             size_mb = atoi(argv[++i]);
+        } else if ((strcmp(argv[i], "--device") == 0 || strcmp(argv[i], "-d") == 0) && i + 1 < argc) {
+            device_index = atoi(argv[++i]);
         } else {
-            fprintf(stderr, "usage: %s [--seconds N] [--size-mb N]\n", argv[0]);
+            fprintf(stderr, "usage: %s [--seconds N] [--size-mb N] [--device N]\n", argv[0]);
             return 2;
         }
     }
@@ -1111,6 +1345,9 @@ int main(int argc, char **argv) {
     if (size_mb <= 0) {
         size_mb = 64;
     }
+    if (device_index < 0) {
+        device_index = 0;
+    }
 
     struct cuda_api cuda;
     if (!load_cuda(&cuda)) {
@@ -1133,8 +1370,13 @@ int main(int argc, char **argv) {
         return 1;
     }
 
+    if (device_index >= count) {
+        fprintf(stderr, "device index %d out of range (found %d CUDA device(s))\n", device_index, count);
+        return 1;
+    }
+
     CUdevice dev = 0;
-    if (!check_rc(&cuda, "cuDeviceGet", cuda.cuDeviceGet(&dev, 0))) {
+    if (!check_rc(&cuda, "cuDeviceGet", cuda.cuDeviceGet(&dev, device_index))) {
         return 1;
     }
 
@@ -1162,10 +1404,12 @@ int main(int argc, char **argv) {
     }
 
     printf("device=%s\n", report.device);
+    printf("device_index=%d\n", device_index);
     printf("compute_capability=%d.%d\n", report.cc_major, report.cc_minor);
     printf("backend=%s\n", report.backend);
     printf("duration_s=%d\n", seconds);
     printf("buffer_mb=%d\n", report.buffer_mb);
+    printf("streams=%d\n", report.stream_count);
     printf("iterations=%lu\n", report.iterations);
     printf("checksum=%llu\n", (unsigned long long)report.checksum);
     if (report.details[0] != '\0') {

iso/builder/build-cublas.sh

@@ -1,9 +1,9 @@
 #!/bin/sh
-# build-cublas.sh — download cuBLASLt/cuBLAS/cudart runtime + headers for bee-gpu-stress.
+# build-cublas.sh — download cuBLASLt/cuBLAS/cudart runtime + headers for bee-gpu-burn worker.
 #
 # Downloads .deb packages from NVIDIA's CUDA apt repository (Debian 12, x86_64),
 # verifies them against Packages.gz, and extracts the small subset we need:
-#   - headers for compiling bee-gpu-stress against cuBLASLt
+#   - headers for compiling bee-gpu-burn worker against cuBLASLt
 #   - runtime libs for libcublas, libcublasLt, libcudart inside the ISO
 
 set -e

iso/builder/build-in-container.sh

@@ -48,8 +48,8 @@ while [ $# -gt 0 ]; do
 done
 
 case "$VARIANT" in
-    nvidia|amd|all) ;;
-    *) echo "unknown variant: $VARIANT (expected nvidia, amd, or all)" >&2; exit 1 ;;
+    nvidia|amd|nogpu|all) ;;
+    *) echo "unknown variant: $VARIANT (expected nvidia, amd, nogpu, or all)" >&2; exit 1 ;;
 esac
 
 if [ "$CLEAN_CACHE" = "1" ]; then
@@ -62,6 +62,7 @@ if [ "$CLEAN_CACHE" = "1" ]; then
     echo "=== cleaning live-build work dirs ==="
     rm -rf "${REPO_ROOT}/dist/live-build-work-nvidia"
     rm -rf "${REPO_ROOT}/dist/live-build-work-amd"
+    rm -rf "${REPO_ROOT}/dist/live-build-work-nogpu"
     echo "=== caches cleared, proceeding with build ==="
 fi
 
@@ -182,8 +183,12 @@ case "$VARIANT" in
     amd)
         run_variant amd
         ;;
+    nogpu)
+        run_variant nogpu
+        ;;
     all)
         run_variant nvidia
         run_variant amd
+        run_variant nogpu
         ;;
 esac

iso/builder/build-john.sh

@@ -0,0 +1,55 @@
+#!/bin/sh
+# build-john.sh — build John the Ripper jumbo with OpenCL support for the LiveCD.
+#
+# Downloads a pinned source snapshot from the official openwall/john repository,
+# builds it inside the builder container, and caches the resulting run/ tree.
+
+set -e
+
+JOHN_COMMIT="$1"
+DIST_DIR="$2"
+
+[ -n "$JOHN_COMMIT" ] || { echo "usage: $0 <john-commit> <dist-dir>"; exit 1; }
+[ -n "$DIST_DIR" ] || { echo "usage: $0 <john-commit> <dist-dir>"; exit 1; }
+
+echo "=== John the Ripper jumbo ${JOHN_COMMIT} ==="
+
+CACHE_DIR="${DIST_DIR}/john-${JOHN_COMMIT}"
+CACHE_ROOT="${BEE_CACHE_DIR:-${DIST_DIR}/cache}"
+DOWNLOAD_CACHE_DIR="${CACHE_ROOT}/john-downloads"
+SRC_TAR="${DOWNLOAD_CACHE_DIR}/john-${JOHN_COMMIT}.tar.gz"
+SRC_URL="https://github.com/openwall/john/archive/${JOHN_COMMIT}.tar.gz"
+
+if [ -x "${CACHE_DIR}/run/john" ] && [ -f "${CACHE_DIR}/run/john.conf" ]; then
+    echo "=== john cached, skipping build ==="
+    echo "run dir: ${CACHE_DIR}/run"
+    exit 0
+fi
+
+mkdir -p "${DOWNLOAD_CACHE_DIR}"
+if [ ! -f "${SRC_TAR}" ]; then
+    echo "=== downloading john source snapshot ==="
+    wget --show-progress -O "${SRC_TAR}" "${SRC_URL}"
+fi
+
+BUILD_TMP=$(mktemp -d)
+trap 'rm -rf "${BUILD_TMP}"' EXIT INT TERM
+
+cd "${BUILD_TMP}"
+tar xf "${SRC_TAR}"
+SRC_DIR=$(find . -maxdepth 1 -type d -name 'john-*' | head -1)
+[ -n "${SRC_DIR}" ] || { echo "ERROR: john source directory not found"; exit 1; }
+
+cd "${SRC_DIR}/src"
+echo "=== configuring john ==="
+./configure
+echo "=== building john ==="
+make clean >/dev/null 2>&1 || true
+make -j"$(nproc)"
+
+mkdir -p "${CACHE_DIR}"
+cp -a "../run" "${CACHE_DIR}/run"
+chmod +x "${CACHE_DIR}/run/john"
+
+echo "=== john build complete ==="
+echo "run dir: ${CACHE_DIR}/run"

iso/builder/build-nccl-tests.sh

@@ -9,6 +9,7 @@
 #
 # Output layout:
 #   $CACHE_DIR/bin/all_reduce_perf
+#   $CACHE_DIR/lib/libcudart.so* copied from the nvcc toolchain used to build nccl-tests
 
 set -e
 
@@ -30,7 +31,7 @@ CACHE_DIR="${DIST_DIR}/nccl-tests-${NCCL_TESTS_VERSION}"
 CACHE_ROOT="${BEE_CACHE_DIR:-${DIST_DIR}/cache}"
 DOWNLOAD_CACHE_DIR="${CACHE_ROOT}/nccl-tests-downloads"
 
-if [ -f "${CACHE_DIR}/bin/all_reduce_perf" ]; then
+if [ -f "${CACHE_DIR}/bin/all_reduce_perf" ] && [ "$(find "${CACHE_DIR}/lib" -maxdepth 1 -name 'libcudart.so*' 2>/dev/null | wc -l)" -gt 0 ]; then
     echo "=== nccl-tests cached, skipping build ==="
     echo "binary: ${CACHE_DIR}/bin/all_reduce_perf"
     exit 0
@@ -52,6 +53,23 @@ echo "nvcc: $NVCC"
 CUDA_HOME="$(dirname "$(dirname "$NVCC")")"
 echo "CUDA_HOME: $CUDA_HOME"
 
+find_cudart_dir() {
+    for dir in \
+        "${CUDA_HOME}/targets/x86_64-linux/lib" \
+        "${CUDA_HOME}/targets/x86_64-linux/lib/stubs" \
+        "${CUDA_HOME}/lib64" \
+        "${CUDA_HOME}/lib"; do
+        if [ -d "$dir" ] && find "$dir" -maxdepth 1 -name 'libcudart.so*' -type f | grep -q .; then
+            printf '%s\n' "$dir"
+            return 0
+        fi
+    done
+    return 1
+}
+
+CUDART_DIR="$(find_cudart_dir)" || { echo "ERROR: libcudart.so* not found under ${CUDA_HOME}"; exit 1; }
+echo "cudart dir: $CUDART_DIR"
+
 # Download libnccl-dev for nccl.h
 REPO_BASE="https://developer.download.nvidia.com/compute/cuda/repos/debian${DEBIAN_VERSION}/x86_64"
 DEV_PKG="libnccl-dev_${NCCL_VERSION}+cuda${NCCL_CUDA_VERSION}_amd64.deb"
@@ -136,6 +154,11 @@ mkdir -p "${CACHE_DIR}/bin"
 cp "./build/all_reduce_perf" "${CACHE_DIR}/bin/all_reduce_perf"
 chmod +x "${CACHE_DIR}/bin/all_reduce_perf"
 
+mkdir -p "${CACHE_DIR}/lib"
+find "${CUDART_DIR}" -maxdepth 1 -name 'libcudart.so*' -type f -exec cp -a {} "${CACHE_DIR}/lib/" \;
+[ "$(find "${CACHE_DIR}/lib" -maxdepth 1 -name 'libcudart.so*' -type f | wc -l)" -gt 0 ] || { echo "ERROR: libcudart runtime copy failed"; exit 1; }
+
 echo "=== nccl-tests build complete ==="
 echo "binary: ${CACHE_DIR}/bin/all_reduce_perf"
 ls -lh "${CACHE_DIR}/bin/all_reduce_perf"
+ls -lh "${CACHE_DIR}/lib/"libcudart.so* 2>/dev/null || true

iso/builder/build-nvidia-module.sh

@@ -10,7 +10,7 @@
 # Output layout:
 #   $CACHE_DIR/modules/   — nvidia*.ko files
 #   $CACHE_DIR/bin/       — nvidia-smi, nvidia-debugdump
-#   $CACHE_DIR/lib/       — libnvidia-ml.so*, libcuda.so* (for nvidia-smi)
+#   $CACHE_DIR/lib/       — libnvidia-ml.so*, libcuda.so*, OpenCL-related libs
 
 set -e
 
@@ -46,7 +46,10 @@ CACHE_DIR="${DIST_DIR}/nvidia-${NVIDIA_VERSION}-${KVER}"
 CACHE_ROOT="${BEE_CACHE_DIR:-${DIST_DIR}/cache}"
 DOWNLOAD_CACHE_DIR="${CACHE_ROOT}/nvidia-downloads"
 EXTRACT_CACHE_DIR="${CACHE_ROOT}/nvidia-extract"
+CACHE_LAYOUT_VERSION="2"
+CACHE_LAYOUT_MARKER="${CACHE_DIR}/.cache-layout-v${CACHE_LAYOUT_VERSION}"
 if [ -d "$CACHE_DIR/modules" ] && [ -f "$CACHE_DIR/bin/nvidia-smi" ] \
+        && [ -f "$CACHE_LAYOUT_MARKER" ] \
         && [ "$(ls "$CACHE_DIR/lib/libnvidia-ptxjitcompiler.so."* 2>/dev/null | wc -l)" -gt 0 ]; then
     echo "=== NVIDIA cached, skipping build ==="
     echo "cache: $CACHE_DIR"
@@ -130,17 +133,30 @@ else
     echo "WARNING: no firmware/ dir found in installer (may be needed for Hopper GPUs)"
 fi
 
-# Copy ALL userspace library files.
-# libnvidia-ptxjitcompiler is required by libcuda for PTX JIT compilation
-# (cuModuleLoadDataEx with PTX source) — without it CUDA_ERROR_JIT_COMPILER_NOT_FOUND.
-for lib in libnvidia-ml libcuda libnvidia-ptxjitcompiler; do
-    count=0
-    for f in $(find "$EXTRACT_DIR" -maxdepth 1 -name "${lib}.so.*" 2>/dev/null); do
-        cp "$f" "$CACHE_DIR/lib/" && count=$((count+1))
-    done
-    if [ "$count" -eq 0 ]; then
-        echo "ERROR: ${lib}.so.* not found in $EXTRACT_DIR"
-        ls "$EXTRACT_DIR/"*.so* 2>/dev/null | head -20 || true
+# Copy NVIDIA userspace libraries broadly instead of whitelisting a few names.
+# Newer driver branches add extra runtime deps (for example OpenCL/compiler side
+# libraries). If we only copy a narrow allowlist, clinfo/John can see nvidia.icd
+# but still fail with "no OpenCL platforms" because one dependent .so is absent.
+copied_libs=0
+for f in $(find "$EXTRACT_DIR" -maxdepth 1 \( -name 'libnvidia*.so.*' -o -name 'libcuda.so.*' \) -type f 2>/dev/null | sort); do
+    cp "$f" "$CACHE_DIR/lib/"
+    copied_libs=$((copied_libs+1))
+done
+
+if [ "$copied_libs" -eq 0 ]; then
+    echo "ERROR: no NVIDIA userspace libraries found in $EXTRACT_DIR"
+    ls "$EXTRACT_DIR/"*.so* 2>/dev/null | head -40 || true
+    exit 1
+fi
+
+for lib in \
+    libnvidia-ml \
+    libcuda \
+    libnvidia-ptxjitcompiler \
+    libnvidia-opencl; do
+    if ! ls "$CACHE_DIR/lib/${lib}.so."* >/dev/null 2>&1; then
+        echo "ERROR: required ${lib}.so.* not found in extracted userspace libs"
+        ls "$CACHE_DIR/lib/" | sort >&2 || true
         exit 1
     fi
 done
@@ -149,16 +165,17 @@ done
 ko_count=$(ls "$CACHE_DIR/modules/"*.ko 2>/dev/null | wc -l)
 [ "$ko_count" -gt 0 ] || { echo "ERROR: no .ko files built in $CACHE_DIR/modules/"; exit 1; }
 
-# Create soname symlinks: use [0-9][0-9]* to avoid circular symlink (.so.1 has single digit)
-for lib in libnvidia-ml libcuda libnvidia-ptxjitcompiler; do
-    versioned=$(ls "$CACHE_DIR/lib/${lib}.so."[0-9][0-9]* 2>/dev/null | head -1)
-    [ -n "$versioned" ] || continue
+# Create soname symlinks for every copied versioned library.
+for versioned in "$CACHE_DIR"/lib/*.so.*; do
+    [ -f "$versioned" ] || continue
     base=$(basename "$versioned")
-    ln -sf "$base" "$CACHE_DIR/lib/${lib}.so.1"
-    ln -sf "${lib}.so.1" "$CACHE_DIR/lib/${lib}.so" 2>/dev/null || true
-    echo "${lib}: .so.1 -> $base"
+    stem=${base%%.so.*}
+    ln -sf "$base" "$CACHE_DIR/lib/${stem}.so.1"
+    ln -sf "${stem}.so.1" "$CACHE_DIR/lib/${stem}.so" 2>/dev/null || true
 done
 
+touch "$CACHE_LAYOUT_MARKER"
+
 echo "=== NVIDIA build complete ==="
 echo "cache: $CACHE_DIR"
 echo "modules: $ko_count .ko files"

iso/builder/build.sh

@@ -27,8 +27,8 @@ while [ $# -gt 0 ]; do
 done
 
 case "$BEE_GPU_VENDOR" in
-    nvidia|amd) ;;
-    *) echo "unknown variant: $BEE_GPU_VENDOR (expected nvidia or amd)" >&2; exit 1 ;;
+    nvidia|amd|nogpu) ;;
+    *) echo "unknown variant: $BEE_GPU_VENDOR (expected nvidia, amd, or nogpu)" >&2; exit 1 ;;
 esac
 
 BUILD_WORK_DIR="${DIST_DIR}/live-build-work-${BEE_GPU_VENDOR}"
@@ -38,6 +38,7 @@ export BEE_GPU_VENDOR
 
 . "${BUILDER_DIR}/VERSIONS"
 export PATH="$PATH:/usr/local/go/bin"
+: "${BEE_REQUIRE_MEMTEST:=0}"
 
 # Allow git to read the bind-mounted repo (different UID inside container).
 git config --global safe.directory "${REPO_ROOT}"
@@ -53,15 +54,8 @@ 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
@@ -111,8 +105,629 @@ resolve_iso_version() {
     resolve_audit_version
 }
 
+iso_list_files() {
+    iso_path="$1"
+
+    if command -v bsdtar >/dev/null 2>&1; then
+        bsdtar -tf "$iso_path"
+        return $?
+    fi
+
+    if command -v xorriso >/dev/null 2>&1; then
+        xorriso -indev "$iso_path" -find / -type f -print 2>/dev/null | sed 's#^/##'
+        return $?
+    fi
+
+    return 127
+}
+
+iso_extract_file() {
+    iso_path="$1"
+    iso_member="$2"
+
+    if command -v bsdtar >/dev/null 2>&1; then
+        bsdtar -xOf "$iso_path" "$iso_member"
+        return $?
+    fi
+
+    if command -v xorriso >/dev/null 2>&1; then
+        xorriso -osirrox on -indev "$iso_path" -cat "/$iso_member" 2>/dev/null
+        return $?
+    fi
+
+    return 127
+}
+
+iso_read_file_list() {
+    iso_path="$1"
+    out_path="$2"
+
+    iso_list_files "$iso_path" > "$out_path" || return 1
+    [ -s "$out_path" ] || return 1
+    return 0
+}
+
+iso_read_member() {
+    iso_path="$1"
+    iso_member="$2"
+    out_path="$3"
+
+    iso_extract_file "$iso_path" "$iso_member" > "$out_path" || return 1
+    [ -s "$out_path" ] || return 1
+    return 0
+}
+
+require_iso_reader() {
+    command -v bsdtar >/dev/null 2>&1 && return 0
+    command -v xorriso >/dev/null 2>&1 && return 0
+    memtest_fail "ISO reader is required for validation/debug (expected bsdtar or xorriso)" "${1:-}"
+}
+
+dump_memtest_debug() {
+    phase="$1"
+    lb_dir="${2:-}"
+    iso_path="${3:-}"
+    phase_slug="$(printf '%s' "${phase}" | tr ' /' '__')"
+    memtest_log="${LOG_DIR:-}/memtest-${phase_slug}.log"
+
+    (
+        echo "=== memtest debug: ${phase} ==="
+
+        echo "-- auto/config --"
+        if [ -f "${BUILDER_DIR}/auto/config" ]; then
+            grep -n -- '--memtest' "${BUILDER_DIR}/auto/config" || echo "  (no --memtest line found)"
+        else
+            echo "  (missing ${BUILDER_DIR}/auto/config)"
+        fi
+
+        echo "-- source bootloader templates --"
+        for cfg in \
+            "${BUILDER_DIR}/config/bootloaders/grub-pc/grub.cfg" \
+            "${BUILDER_DIR}/config/bootloaders/isolinux/live.cfg.in"; do
+            if [ -f "$cfg" ]; then
+                echo "  file: $cfg"
+                grep -n 'Memory Test\|memtest' "$cfg" || echo "    (no memtest lines)"
+            fi
+        done
+
+        echo "-- source binary hooks --"
+        for hook in \
+            "${BUILDER_DIR}/config/hooks/normal/9100-memtest.hook.binary"; do
+            if [ -f "$hook" ]; then
+                echo "  hook: $hook"
+            else
+                echo "  (missing $hook)"
+            fi
+        done
+
+        if [ -n "$lb_dir" ] && [ -d "$lb_dir" ]; then
+            echo "-- live-build workdir package lists --"
+            for pkg in \
+                "$lb_dir/config/package-lists/bee.list.chroot" \
+                "$lb_dir/config/package-lists/bee-gpu.list.chroot" \
+                "$lb_dir/config/package-lists/bee-nvidia.list.chroot"; do
+                if [ -f "$pkg" ]; then
+                    echo "  file: $pkg"
+                    grep -n 'memtest' "$pkg" || echo "    (no memtest lines)"
+                fi
+            done
+
+            echo "-- live-build chroot/boot --"
+            if [ -d "$lb_dir/chroot/boot" ]; then
+                find "$lb_dir/chroot/boot" -maxdepth 1 -name 'memtest*' -print | sed 's/^/  /' || true
+            else
+                echo "  (missing $lb_dir/chroot/boot)"
+            fi
+
+            echo "-- live-build binary/boot --"
+            if [ -d "$lb_dir/binary/boot" ]; then
+                find "$lb_dir/binary/boot" -maxdepth 1 -name 'memtest*' -print | sed 's/^/  /' || true
+            else
+                echo "  (missing $lb_dir/binary/boot)"
+            fi
+
+            echo "-- live-build binary grub cfg --"
+            if [ -f "$lb_dir/binary/boot/grub/grub.cfg" ]; then
+                grep -n 'Memory Test\|memtest' "$lb_dir/binary/boot/grub/grub.cfg" || echo "  (no memtest lines)"
+            else
+                echo "  (missing $lb_dir/binary/boot/grub/grub.cfg)"
+            fi
+
+            echo "-- live-build binary isolinux cfg --"
+            if [ -f "$lb_dir/binary/isolinux/live.cfg" ]; then
+                grep -n 'Memory Test\|memtest' "$lb_dir/binary/isolinux/live.cfg" || echo "  (no memtest lines)"
+            else
+                echo "  (missing $lb_dir/binary/isolinux/live.cfg)"
+            fi
+
+            echo "-- live-build package cache --"
+            if [ -d "$lb_dir/cache/packages.chroot" ]; then
+                find "$lb_dir/cache/packages.chroot" -maxdepth 1 -name 'memtest86+*.deb' -print | sed 's/^/  /' || true
+            else
+                echo "  (missing $lb_dir/cache/packages.chroot)"
+            fi
+        fi
+
+        if [ -n "$iso_path" ] && [ -f "$iso_path" ]; then
+            iso_files="$(mktemp)"
+            iso_grub_cfg="$(mktemp)"
+            iso_isolinux_cfg="$(mktemp)"
+
+            echo "-- ISO memtest files --"
+            if iso_read_file_list "$iso_path" "$iso_files"; then
+                grep 'memtest' "$iso_files" | sed 's/^/  /' || echo "  (no memtest files in ISO)"
+            else
+                echo "  (failed to list ISO contents)"
+            fi
+
+            echo "-- ISO GRUB memtest lines --"
+            if iso_read_member "$iso_path" boot/grub/grub.cfg "$iso_grub_cfg"; then
+                grep -n 'Memory Test\|memtest' "$iso_grub_cfg" || echo "  (no memtest lines in boot/grub/grub.cfg)"
+            else
+                echo "  (failed to read boot/grub/grub.cfg from ISO)"
+            fi
+
+            echo "-- ISO isolinux memtest lines --"
+            if iso_read_member "$iso_path" isolinux/live.cfg "$iso_isolinux_cfg"; then
+                grep -n 'Memory Test\|memtest' "$iso_isolinux_cfg" || echo "  (no memtest lines in isolinux/live.cfg)"
+            else
+                echo "  (failed to read isolinux/live.cfg from ISO)"
+            fi
+
+            rm -f "$iso_files" "$iso_grub_cfg" "$iso_isolinux_cfg"
+        fi
+
+        echo "=== end memtest debug: ${phase} ==="
+    ) | {
+        if [ -n "${LOG_DIR:-}" ] && [ -d "${LOG_DIR}" ]; then
+            tee "${memtest_log}"
+        else
+            cat
+        fi
+    }
+}
+
+memtest_fail() {
+    msg="$1"
+    iso_path="${2:-}"
+    level="WARNING"
+    if [ "${BEE_REQUIRE_MEMTEST:-0}" = "1" ]; then
+        level="ERROR"
+    fi
+    echo "${level}: ${msg}" >&2
+    dump_memtest_debug "failure" "${LB_DIR:-}" "$iso_path" >&2
+    if [ "${BEE_REQUIRE_MEMTEST:-0}" = "1" ]; then
+        exit 1
+    fi
+    return 0
+}
+
+iso_memtest_present() {
+    iso_path="$1"
+    iso_files="$(mktemp)"
+
+    [ -f "$iso_path" ] || return 1
+
+    if command -v bsdtar >/dev/null 2>&1; then
+        :
+    elif command -v xorriso >/dev/null 2>&1; then
+        :
+    else
+        return 2
+    fi
+
+    iso_read_file_list "$iso_path" "$iso_files" || {
+        rm -f "$iso_files"
+        return 2
+    }
+
+    grep -q '^boot/memtest86+x64\.bin$' "$iso_files" || {
+        rm -f "$iso_files"
+        return 1
+    }
+    grep -q '^boot/memtest86+x64\.efi$' "$iso_files" || {
+        rm -f "$iso_files"
+        return 1
+    }
+
+    grub_cfg="$(mktemp)"
+    isolinux_cfg="$(mktemp)"
+
+    iso_read_member "$iso_path" boot/grub/grub.cfg "$grub_cfg" || {
+        rm -f "$iso_files" "$grub_cfg" "$isolinux_cfg"
+        return 2
+    }
+    iso_read_member "$iso_path" isolinux/live.cfg "$isolinux_cfg" || {
+        rm -f "$iso_files" "$grub_cfg" "$isolinux_cfg"
+        return 2
+    }
+
+    grep -q 'Memory Test (memtest86+)' "$grub_cfg" || {
+        rm -f "$iso_files" "$grub_cfg" "$isolinux_cfg"
+        return 1
+    }
+    grep -q '/boot/memtest86+x64\.efi' "$grub_cfg" || {
+        rm -f "$iso_files" "$grub_cfg" "$isolinux_cfg"
+        return 1
+    }
+    grep -q '/boot/memtest86+x64\.bin' "$grub_cfg" || {
+        rm -f "$iso_files" "$grub_cfg" "$isolinux_cfg"
+        return 1
+    }
+    grep -q 'Memory Test (memtest86+)' "$isolinux_cfg" || {
+        rm -f "$iso_files" "$grub_cfg" "$isolinux_cfg"
+        return 1
+    }
+    grep -q '/boot/memtest86+x64\.bin' "$isolinux_cfg" || {
+        rm -f "$iso_files" "$grub_cfg" "$isolinux_cfg"
+        return 1
+    }
+
+    rm -f "$iso_files" "$grub_cfg" "$isolinux_cfg"
+    return 0
+}
+
+validate_iso_memtest() {
+    iso_path="$1"
+    echo "=== validating memtest in ISO ==="
+
+    [ -f "$iso_path" ] || {
+        memtest_fail "ISO not found for validation: $iso_path" "$iso_path"
+        return 0
+    }
+    require_iso_reader "$iso_path" || return 0
+
+    iso_files="$(mktemp)"
+    iso_read_file_list "$iso_path" "$iso_files" || {
+        memtest_fail "failed to list ISO contents while validating memtest" "$iso_path"
+        rm -f "$iso_files"
+        return 0
+    }
+
+    grep -q '^boot/memtest86+x64\.bin$' "$iso_files" || {
+        memtest_fail "memtest BIOS binary missing in ISO: boot/memtest86+x64.bin" "$iso_path"
+        rm -f "$iso_files"
+        return 0
+    }
+    grep -q '^boot/memtest86+x64\.efi$' "$iso_files" || {
+        memtest_fail "memtest EFI binary missing in ISO: boot/memtest86+x64.efi" "$iso_path"
+        rm -f "$iso_files"
+        return 0
+    }
+
+    grub_cfg="$(mktemp)"
+    isolinux_cfg="$(mktemp)"
+
+    iso_read_member "$iso_path" boot/grub/grub.cfg "$grub_cfg" || {
+        memtest_fail "failed to read boot/grub/grub.cfg from ISO" "$iso_path"
+        rm -f "$iso_files" "$grub_cfg" "$isolinux_cfg"
+        return 0
+    }
+    iso_read_member "$iso_path" isolinux/live.cfg "$isolinux_cfg" || {
+        memtest_fail "failed to read isolinux/live.cfg from ISO" "$iso_path"
+        rm -f "$iso_files" "$grub_cfg" "$isolinux_cfg"
+        return 0
+    }
+
+    grep -q 'Memory Test (memtest86+)' "$grub_cfg" || {
+        memtest_fail "GRUB menu entry for memtest is missing" "$iso_path"
+        rm -f "$iso_files" "$grub_cfg" "$isolinux_cfg"
+        return 0
+    }
+    grep -q '/boot/memtest86+x64\.efi' "$grub_cfg" || {
+        memtest_fail "GRUB memtest EFI path is missing" "$iso_path"
+        rm -f "$iso_files" "$grub_cfg" "$isolinux_cfg"
+        return 0
+    }
+    grep -q '/boot/memtest86+x64\.bin' "$grub_cfg" || {
+        memtest_fail "GRUB memtest BIOS path is missing" "$iso_path"
+        rm -f "$iso_files" "$grub_cfg" "$isolinux_cfg"
+        return 0
+    }
+    grep -q 'Memory Test (memtest86+)' "$isolinux_cfg" || {
+        memtest_fail "isolinux menu entry for memtest is missing" "$iso_path"
+        rm -f "$iso_files" "$grub_cfg" "$isolinux_cfg"
+        return 0
+    }
+    grep -q '/boot/memtest86+x64\.bin' "$isolinux_cfg" || {
+        memtest_fail "isolinux memtest path is missing" "$iso_path"
+        rm -f "$iso_files" "$grub_cfg" "$isolinux_cfg"
+        return 0
+    }
+
+    rm -f "$iso_files" "$grub_cfg" "$isolinux_cfg"
+    echo "=== memtest validation OK ==="
+}
+
+append_memtest_grub_entry() {
+    grub_cfg="$1"
+    [ -f "$grub_cfg" ] || return 1
+    grep -q 'Memory Test (memtest86+)' "$grub_cfg" && return 0
+    grep -q '### BEE MEMTEST ###' "$grub_cfg" && return 0
+
+    cat >> "$grub_cfg" <<'EOF'
+
+### BEE MEMTEST ###
+if [ "${grub_platform}" = "efi" ]; then
+    menuentry "Memory Test (memtest86+)" {
+        chainloader /boot/memtest86+x64.efi
+    }
+else
+    menuentry "Memory Test (memtest86+)" {
+        linux16 /boot/memtest86+x64.bin
+    }
+fi
+### /BEE MEMTEST ###
+EOF
+}
+
+append_memtest_isolinux_entry() {
+    isolinux_cfg="$1"
+    [ -f "$isolinux_cfg" ] || return 1
+    grep -q 'Memory Test (memtest86+)' "$isolinux_cfg" && return 0
+    grep -q '### BEE MEMTEST ###' "$isolinux_cfg" && return 0
+
+    cat >> "$isolinux_cfg" <<'EOF'
+
+# ### BEE MEMTEST ###
+label memtest
+    menu label ^Memory Test (memtest86+)
+    linux /boot/memtest86+x64.bin
+# ### /BEE MEMTEST ###
+EOF
+}
+
+copy_memtest_from_deb() {
+    deb="$1"
+    dst_boot="$2"
+    tmpdir="$(mktemp -d)"
+
+    dpkg-deb -x "$deb" "$tmpdir"
+    for f in memtest86+x64.bin memtest86+x64.efi; do
+        if [ -f "$tmpdir/boot/$f" ]; then
+            cp "$tmpdir/boot/$f" "$dst_boot/$f"
+        fi
+    done
+    rm -rf "$tmpdir"
+}
+
+reset_live_build_stage() {
+    lb_dir="$1"
+    stage="$2"
+
+    for root in \
+        "$lb_dir/.build" \
+        "$lb_dir/.stage" \
+        "$lb_dir/auto"; do
+        [ -d "$root" ] || continue
+        find "$root" -maxdepth 1 \( -name "${stage}" -o -name "${stage}.*" -o -name "*${stage}*" \) -exec rm -rf {} + 2>/dev/null || true
+    done
+}
+
+recover_iso_memtest() {
+    lb_dir="$1"
+    iso_path="$2"
+    binary_boot="$lb_dir/binary/boot"
+    grub_cfg="$lb_dir/binary/boot/grub/grub.cfg"
+    isolinux_cfg="$lb_dir/binary/isolinux/live.cfg"
+
+    echo "=== attempting memtest recovery in binary tree ==="
+
+    mkdir -p "$binary_boot"
+
+    for root in \
+        "$lb_dir/chroot/boot" \
+        "/boot"; do
+        for f in memtest86+x64.bin memtest86+x64.efi; do
+            if [ ! -f "$binary_boot/$f" ] && [ -f "$root/$f" ]; then
+                cp "$root/$f" "$binary_boot/$f"
+                echo "memtest recovery: copied $f from $root"
+            fi
+        done
+    done
+
+    if [ ! -f "$binary_boot/memtest86+x64.bin" ] || [ ! -f "$binary_boot/memtest86+x64.efi" ]; then
+        for dir in \
+            "$lb_dir/cache/packages.binary" \
+            "$lb_dir/cache/packages.chroot" \
+            "$lb_dir/chroot/var/cache/apt/archives" \
+            "${BEE_CACHE_DIR:-${DIST_DIR}/cache}/lb-packages" \
+            "/var/cache/apt/archives"; do
+            [ -d "$dir" ] || continue
+            deb="$(find "$dir" -maxdepth 1 -type f -name 'memtest86+*.deb' 2>/dev/null | head -1)"
+            [ -n "$deb" ] || continue
+            echo "memtest recovery: extracting payload from $deb"
+            copy_memtest_from_deb "$deb" "$binary_boot"
+            break
+        done
+    fi
+
+    if [ ! -f "$binary_boot/memtest86+x64.bin" ] || [ ! -f "$binary_boot/memtest86+x64.efi" ]; then
+        tmpdl="$(mktemp -d)"
+        if (
+            cd "$tmpdl" && apt-get download memtest86+ >/dev/null 2>&1
+        ); then
+            deb="$(find "$tmpdl" -maxdepth 1 -type f -name 'memtest86+*.deb' 2>/dev/null | head -1)"
+            if [ -n "$deb" ]; then
+                echo "memtest recovery: downloaded $deb"
+                copy_memtest_from_deb "$deb" "$binary_boot"
+            fi
+        fi
+        rm -rf "$tmpdl"
+    fi
+
+    if [ -f "$grub_cfg" ]; then
+        append_memtest_grub_entry "$grub_cfg" && echo "memtest recovery: ensured GRUB entry"
+    else
+        echo "memtest recovery: WARNING: missing $grub_cfg"
+    fi
+
+    if [ -f "$isolinux_cfg" ]; then
+        append_memtest_isolinux_entry "$isolinux_cfg" && echo "memtest recovery: ensured isolinux entry"
+    else
+        echo "memtest recovery: WARNING: missing $isolinux_cfg"
+    fi
+
+    reset_live_build_stage "$lb_dir" "binary_checksums"
+    reset_live_build_stage "$lb_dir" "binary_iso"
+    reset_live_build_stage "$lb_dir" "binary_zsync"
+
+    run_optional_step_sh "rebuild live-build checksums after memtest recovery" "91-lb-checksums" "lb binary_checksums 2>&1"
+    run_optional_step_sh "rebuild ISO after memtest recovery" "92-lb-binary-iso" "rm -f '$iso_path' && lb binary_iso 2>&1"
+    run_optional_step_sh "rebuild zsync after memtest recovery" "93-lb-zsync" "lb binary_zsync 2>&1"
+
+    if [ ! -f "$iso_path" ]; then
+        memtest_fail "ISO rebuild was skipped or failed after memtest recovery: $iso_path" "$iso_path"
+    fi
+}
+
 AUDIT_VERSION_EFFECTIVE="$(resolve_audit_version)"
 ISO_VERSION_EFFECTIVE="$(resolve_iso_version)"
+ISO_BASENAME="easy-bee-${BEE_GPU_VENDOR}-v${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 "${OUT_DIR}"
+LOG_DIR="${OUT_DIR}/${ISO_BASENAME}.logs"
+LOG_ARCHIVE="${OUT_DIR}/${ISO_BASENAME}.logs.tar.gz"
+ISO_OUT="${OUT_DIR}/${ISO_BASENAME}.iso"
+LOG_OUT="${LOG_DIR}/build.log"
+
+cleanup_build_log() {
+    status="${1:-$?}"
+    trap - EXIT INT TERM HUP
+
+    if [ "${STEP_LOG_ACTIVE:-0}" = "1" ]; then
+        cleanup_step_log "${status}" || true
+    fi
+
+    if [ "${BUILD_LOG_ACTIVE:-0}" = "1" ]; then
+        BUILD_LOG_ACTIVE=0
+        exec 1>&3 2>&4
+        exec 3>&- 4>&-
+        if [ -n "${BUILD_TEE_PID:-}" ]; then
+            wait "${BUILD_TEE_PID}" 2>/dev/null || true
+        fi
+        rm -f "${BUILD_LOG_PIPE}"
+    fi
+
+    if [ -n "${LOG_DIR:-}" ] && [ -d "${LOG_DIR}" ] && command -v tar >/dev/null 2>&1; then
+        rm -f "${LOG_ARCHIVE}"
+        tar -czf "${LOG_ARCHIVE}" -C "$(dirname "${LOG_DIR}")" "$(basename "${LOG_DIR}")" 2>/dev/null || true
+        rm -rf "${LOG_DIR}"
+    fi
+
+    exit "${status}"
+}
+
+start_build_log() {
+    command -v tee >/dev/null 2>&1 || {
+        echo "ERROR: tee is required for build logging" >&2
+        exit 1
+    }
+
+    rm -rf "${LOG_DIR}"
+    rm -f "${LOG_ARCHIVE}"
+    mkdir -p "${LOG_DIR}"
+    BUILD_LOG_PIPE="$(mktemp -u "${TMPDIR:-/tmp}/bee-build-log.XXXXXX")"
+    mkfifo "${BUILD_LOG_PIPE}"
+
+    exec 3>&1 4>&2
+    tee "${LOG_OUT}" < "${BUILD_LOG_PIPE}" &
+    BUILD_TEE_PID=$!
+    exec > "${BUILD_LOG_PIPE}" 2>&1
+    BUILD_LOG_ACTIVE=1
+
+    trap 'cleanup_build_log "$?"' EXIT INT TERM HUP
+
+    echo "=== build log dir: ${LOG_DIR} ==="
+    echo "=== build log: ${LOG_OUT} ==="
+    echo "=== build log archive: ${LOG_ARCHIVE} ==="
+}
+
+cleanup_step_log() {
+    status="${1:-$?}"
+
+    if [ "${STEP_LOG_ACTIVE:-0}" = "1" ]; then
+        STEP_LOG_ACTIVE=0
+        exec 1>&5 2>&6
+        exec 5>&- 6>&-
+        if [ -n "${STEP_TEE_PID:-}" ]; then
+            wait "${STEP_TEE_PID}" 2>/dev/null || true
+        fi
+        rm -f "${STEP_LOG_PIPE}"
+    fi
+
+    return "${status}"
+}
+
+run_step() {
+    step_name="$1"
+    step_slug="$2"
+    shift 2
+
+    step_log="${LOG_DIR}/${step_slug}.log"
+    echo ""
+    echo "=== step: ${step_name} ==="
+    echo "=== step log: ${step_log} ==="
+
+    STEP_LOG_PIPE="$(mktemp -u "${TMPDIR:-/tmp}/bee-step-log.XXXXXX")"
+    mkfifo "${STEP_LOG_PIPE}"
+
+    exec 5>&1 6>&2
+    tee "${step_log}" < "${STEP_LOG_PIPE}" >&5 &
+    STEP_TEE_PID=$!
+    exec > "${STEP_LOG_PIPE}" 2>&1
+    STEP_LOG_ACTIVE=1
+
+    set +e
+    "$@"
+    step_status=$?
+    set -e
+
+    cleanup_step_log "${step_status}"
+    if [ "${step_status}" -ne 0 ]; then
+        echo "ERROR: step failed: ${step_name} (see ${step_log})" >&2
+        exit "${step_status}"
+    fi
+
+    echo "=== step OK: ${step_name} ==="
+}
+
+run_step_sh() {
+    step_name="$1"
+    step_slug="$2"
+    step_script="$3"
+
+    run_step "${step_name}" "${step_slug}" sh -c "${step_script}"
+}
+
+run_optional_step_sh() {
+    step_name="$1"
+    step_slug="$2"
+    step_script="$3"
+
+    if [ "${BEE_REQUIRE_MEMTEST:-0}" = "1" ]; then
+        run_step_sh "${step_name}" "${step_slug}" "${step_script}"
+        return 0
+    fi
+
+    step_log="${LOG_DIR}/${step_slug}.log"
+    echo ""
+    echo "=== optional step: ${step_name} ==="
+    echo "=== optional step log: ${step_log} ==="
+    set +e
+    sh -c "${step_script}" > "${step_log}" 2>&1
+    step_status=$?
+    set -e
+    cat "${step_log}"
+    if [ "${step_status}" -ne 0 ]; then
+        echo "WARNING: optional step failed: ${step_name} (see ${step_log})" >&2
+    else
+        echo "=== optional step OK: ${step_name} ==="
+    fi
+}
+
+start_build_log
 
 # Auto-detect kernel ABI: refresh apt index, then query current linux-image-amd64 dependency.
 # If headers for the detected ABI are not yet installed (kernel updated since image build),
@@ -147,8 +762,8 @@ echo "Debian: ${DEBIAN_VERSION}, Kernel ABI: ${DEBIAN_KERNEL_ABI}, Go: ${GO_VERS
 echo "Audit version: ${AUDIT_VERSION_EFFECTIVE}, ISO version: ${ISO_VERSION_EFFECTIVE}"
 echo ""
 
-echo "=== syncing git submodules ==="
-git -C "${REPO_ROOT}" submodule update --init --recursive
+run_step "sync git submodules" "05-git-submodules" \
+    git -C "${REPO_ROOT}" submodule update --init --recursive
 
 # --- compile bee binary (static, Linux amd64) ---
 # Shared between variants — built once, reused on second pass.
@@ -160,13 +775,13 @@ if [ -f "$BEE_BIN" ]; then
 fi
 
 if [ "$NEED_BUILD" = "1" ]; then
-    echo "=== building bee binary ==="
-    cd "${REPO_ROOT}/audit"
-    GOOS=linux GOARCH=amd64 CGO_ENABLED=0 \
-        go build \
-        -ldflags "-s -w -X main.Version=${AUDIT_VERSION_EFFECTIVE}" \
-        -o "$BEE_BIN" \
-        ./cmd/bee
+    run_step_sh "build bee binary" "10-build-bee" \
+        "cd '${REPO_ROOT}/audit' && \
+        env GOOS=linux GOARCH=amd64 CGO_ENABLED=0 \
+            go build \
+            -ldflags '-s -w -X main.Version=${AUDIT_VERSION_EFFECTIVE}' \
+            -o '${BEE_BIN}' \
+            ./cmd/bee"
     echo "binary: $BEE_BIN"
     if command -v stat >/dev/null 2>&1; then
         BEE_SIZE_BYTES="$(stat -c '%s' "$BEE_BIN" 2>/dev/null || stat -f '%z' "$BEE_BIN")"
@@ -183,11 +798,10 @@ else
 fi
 
 # --- NVIDIA-only build steps ---
-GPU_STRESS_BIN="${DIST_DIR}/bee-gpu-stress-linux-amd64"
+GPU_BURN_WORKER_BIN="${DIST_DIR}/bee-gpu-burn-worker-linux-amd64"
 if [ "$BEE_GPU_VENDOR" = "nvidia" ]; then
-    echo ""
-    echo "=== downloading cuBLAS/cuBLASLt/cudart ${NCCL_CUDA_VERSION} userspace ==="
-    sh "${BUILDER_DIR}/build-cublas.sh" \
+    run_step "download cuBLAS/cuBLASLt/cudart ${NCCL_CUDA_VERSION} userspace" "20-cublas" \
+        sh "${BUILDER_DIR}/build-cublas.sh" \
         "${CUBLAS_VERSION}" \
         "${CUDA_USERSPACE_VERSION}" \
         "${NCCL_CUDA_VERSION}" \
@@ -196,20 +810,20 @@ if [ "$BEE_GPU_VENDOR" = "nvidia" ]; then
     CUBLAS_CACHE="${DIST_DIR}/cublas-${CUBLAS_VERSION}+cuda${NCCL_CUDA_VERSION}"
 
     GPU_STRESS_NEED_BUILD=1
-    if [ -f "$GPU_STRESS_BIN" ] && [ "${BUILDER_DIR}/bee-gpu-stress.c" -ot "$GPU_STRESS_BIN" ]; then
+    if [ -f "$GPU_BURN_WORKER_BIN" ] && [ "${BUILDER_DIR}/bee-gpu-stress.c" -ot "$GPU_BURN_WORKER_BIN" ]; then
         GPU_STRESS_NEED_BUILD=0
     fi
 
     if [ "$GPU_STRESS_NEED_BUILD" = "1" ]; then
-        echo "=== building bee-gpu-stress ==="
-        gcc -O2 -s -Wall -Wextra \
+        run_step "build bee-gpu-burn worker" "21-gpu-burn-worker" \
+            gcc -O2 -s -Wall -Wextra \
             -I"${CUBLAS_CACHE}/include" \
-            -o "$GPU_STRESS_BIN" \
+            -o "$GPU_BURN_WORKER_BIN" \
             "${BUILDER_DIR}/bee-gpu-stress.c" \
             -ldl -lm
-        echo "binary: $GPU_STRESS_BIN"
+        echo "binary: $GPU_BURN_WORKER_BIN"
     else
-        echo "=== bee-gpu-stress up to date, skipping build ==="
+        echo "=== bee-gpu-burn worker up to date, skipping build ==="
     fi
 fi
 
@@ -245,12 +859,15 @@ rm -f \
     "${OVERLAY_STAGE_DIR}/etc/bee-release" \
     "${OVERLAY_STAGE_DIR}/root/.ssh/authorized_keys" \
     "${OVERLAY_STAGE_DIR}/usr/local/bin/bee" \
-    "${OVERLAY_STAGE_DIR}/usr/local/bin/bee-gpu-stress" \
+    "${OVERLAY_STAGE_DIR}/usr/local/bin/john" \
+    "${OVERLAY_STAGE_DIR}/usr/local/lib/bee/bee-gpu-burn-worker" \
     "${OVERLAY_STAGE_DIR}/usr/local/bin/bee-smoketest" \
     "${OVERLAY_STAGE_DIR}/usr/local/bin/all_reduce_perf"
+rm -rf \
+    "${OVERLAY_STAGE_DIR}/usr/local/lib/bee/john"
 
-# Remove NVIDIA-specific overlay files for AMD variant
-if [ "$BEE_GPU_VENDOR" = "amd" ]; then
+# Remove NVIDIA-specific overlay files for non-nvidia variants
+if [ "$BEE_GPU_VENDOR" != "nvidia" ]; then
     rm -f "${OVERLAY_STAGE_DIR}/usr/local/bin/bee-nvidia-load"
     rm -f "${OVERLAY_STAGE_DIR}/etc/systemd/system/bee-nvidia.service"
 fi
@@ -293,9 +910,13 @@ mkdir -p "${OVERLAY_STAGE_DIR}/usr/local/bin"
 cp "${DIST_DIR}/bee-linux-amd64" "${OVERLAY_STAGE_DIR}/usr/local/bin/bee"
 chmod +x "${OVERLAY_STAGE_DIR}/usr/local/bin/bee"
 
-if [ "$BEE_GPU_VENDOR" = "nvidia" ] && [ -f "$GPU_STRESS_BIN" ]; then
-    cp "${GPU_STRESS_BIN}" "${OVERLAY_STAGE_DIR}/usr/local/bin/bee-gpu-stress"
-    chmod +x "${OVERLAY_STAGE_DIR}/usr/local/bin/bee-gpu-stress"
+if [ "$BEE_GPU_VENDOR" = "nvidia" ] && [ -f "$GPU_BURN_WORKER_BIN" ]; then
+    mkdir -p "${OVERLAY_STAGE_DIR}/usr/local/lib/bee" "${OVERLAY_STAGE_DIR}/usr/local/bin"
+    cp "${GPU_BURN_WORKER_BIN}" "${OVERLAY_STAGE_DIR}/usr/local/lib/bee/bee-gpu-burn-worker"
+    chmod +x "${OVERLAY_STAGE_DIR}/usr/local/lib/bee/bee-gpu-burn-worker"
+    chmod +x "${OVERLAY_STAGE_DIR}/usr/local/bin/bee-gpu-burn" 2>/dev/null || true
+    chmod +x "${OVERLAY_STAGE_DIR}/usr/local/bin/bee-john-gpu-stress" 2>/dev/null || true
+    chmod +x "${OVERLAY_STAGE_DIR}/usr/local/bin/bee-nccl-gpu-stress" 2>/dev/null || true
 fi
 
 # --- inject smoketest into overlay so it runs directly on the live CD ---
@@ -315,9 +936,8 @@ done
 
 # --- NVIDIA kernel modules and userspace libs ---
 if [ "$BEE_GPU_VENDOR" = "nvidia" ]; then
-    echo ""
-    echo "=== building NVIDIA ${NVIDIA_DRIVER_VERSION} modules ==="
-    sh "${BUILDER_DIR}/build-nvidia-module.sh" "${NVIDIA_DRIVER_VERSION}" "${DIST_DIR}" "${DEBIAN_KERNEL_ABI}"
+    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}"
 
     KVER="${DEBIAN_KERNEL_ABI}-amd64"
     NVIDIA_CACHE="${DIST_DIR}/nvidia-${NVIDIA_DRIVER_VERSION}-${KVER}"
@@ -334,6 +954,8 @@ if [ "$BEE_GPU_VENDOR" = "nvidia" ]; then
     cp "${NVIDIA_CACHE}/bin/nvidia-bug-report.sh" "${OVERLAY_STAGE_DIR}/usr/local/bin/" 2>/dev/null || true
     chmod +x "${OVERLAY_STAGE_DIR}/usr/local/bin/nvidia-bug-report.sh" 2>/dev/null || true
     cp "${NVIDIA_CACHE}/lib/"* "${OVERLAY_STAGE_DIR}/usr/lib/" 2>/dev/null || true
+    mkdir -p "${OVERLAY_STAGE_DIR}/etc/OpenCL/vendors"
+    printf 'libnvidia-opencl.so.1\n' > "${OVERLAY_STAGE_DIR}/etc/OpenCL/vendors/nvidia.icd"
 
     # Inject GSP firmware into /lib/firmware/nvidia/<version>/
     if [ -d "${NVIDIA_CACHE}/firmware" ] && [ "$(ls -A "${NVIDIA_CACHE}/firmware" 2>/dev/null)" ]; then
@@ -343,9 +965,8 @@ if [ "$BEE_GPU_VENDOR" = "nvidia" ]; then
     fi
 
     # --- build / download NCCL ---
-    echo ""
-    echo "=== downloading NCCL ${NCCL_VERSION}+cuda${NCCL_CUDA_VERSION} ==="
-    sh "${BUILDER_DIR}/build-nccl.sh" "${NCCL_VERSION}" "${NCCL_CUDA_VERSION}" "${DIST_DIR}" "${NCCL_SHA256:-}"
+    run_step "download NCCL ${NCCL_VERSION}+cuda${NCCL_CUDA_VERSION}" "50-nccl" \
+        sh "${BUILDER_DIR}/build-nccl.sh" "${NCCL_VERSION}" "${NCCL_CUDA_VERSION}" "${DIST_DIR}" "${NCCL_SHA256:-}"
 
     NCCL_CACHE="${DIST_DIR}/nccl-${NCCL_VERSION}+cuda${NCCL_CUDA_VERSION}"
 
@@ -353,14 +974,13 @@ if [ "$BEE_GPU_VENDOR" = "nvidia" ]; then
     cp "${NCCL_CACHE}/lib/"* "${OVERLAY_STAGE_DIR}/usr/lib/"
     echo "=== NCCL: $(ls "${NCCL_CACHE}/lib/" | wc -l) files injected into /usr/lib/ ==="
 
-    # Inject cuBLAS/cuBLASLt/cudart runtime libs used by bee-gpu-stress tensor-core GEMM path
+    # Inject cuBLAS/cuBLASLt/cudart runtime libs used by the bee-gpu-burn worker tensor-core GEMM path
     cp "${CUBLAS_CACHE}/lib/"* "${OVERLAY_STAGE_DIR}/usr/lib/"
     echo "=== cuBLAS: $(ls "${CUBLAS_CACHE}/lib/" | wc -l) files injected into /usr/lib/ ==="
 
     # --- build nccl-tests ---
-    echo ""
-    echo "=== building nccl-tests ${NCCL_TESTS_VERSION} ==="
-    sh "${BUILDER_DIR}/build-nccl-tests.sh" \
+    run_step "build nccl-tests ${NCCL_TESTS_VERSION}" "60-nccl-tests" \
+        sh "${BUILDER_DIR}/build-nccl-tests.sh" \
         "${NCCL_TESTS_VERSION}" \
         "${NCCL_VERSION}" \
         "${NCCL_CUDA_VERSION}" \
@@ -371,7 +991,17 @@ if [ "$BEE_GPU_VENDOR" = "nvidia" ]; then
     NCCL_TESTS_CACHE="${DIST_DIR}/nccl-tests-${NCCL_TESTS_VERSION}"
     cp "${NCCL_TESTS_CACHE}/bin/all_reduce_perf" "${OVERLAY_STAGE_DIR}/usr/local/bin/all_reduce_perf"
     chmod +x "${OVERLAY_STAGE_DIR}/usr/local/bin/all_reduce_perf"
+    cp "${NCCL_TESTS_CACHE}/lib/"* "${OVERLAY_STAGE_DIR}/usr/lib/" 2>/dev/null || true
     echo "=== all_reduce_perf injected ==="
+
+    run_step "build john jumbo ${JOHN_JUMBO_COMMIT}" "70-john" \
+        sh "${BUILDER_DIR}/build-john.sh" "${JOHN_JUMBO_COMMIT}" "${DIST_DIR}"
+    JOHN_CACHE="${DIST_DIR}/john-${JOHN_JUMBO_COMMIT}"
+    mkdir -p "${OVERLAY_STAGE_DIR}/usr/local/lib/bee/john"
+    rsync -a --delete "${JOHN_CACHE}/run/" "${OVERLAY_STAGE_DIR}/usr/local/lib/bee/john/run/"
+    ln -sfn ../lib/bee/john/run/john "${OVERLAY_STAGE_DIR}/usr/local/bin/john"
+    chmod +x "${OVERLAY_STAGE_DIR}/usr/local/lib/bee/john/run/john"
+    echo "=== john injected ==="
 fi
 
 # --- embed build metadata ---
@@ -385,11 +1015,15 @@ NCCL_VERSION=${NCCL_VERSION}
 NCCL_CUDA_VERSION=${NCCL_CUDA_VERSION}
 CUBLAS_VERSION=${CUBLAS_VERSION}
 CUDA_USERSPACE_VERSION=${CUDA_USERSPACE_VERSION}
-NCCL_TESTS_VERSION=${NCCL_TESTS_VERSION}"
+NCCL_TESTS_VERSION=${NCCL_TESTS_VERSION}
+JOHN_JUMBO_COMMIT=${JOHN_JUMBO_COMMIT}"
     GPU_BUILD_INFO="nvidia:${NVIDIA_DRIVER_VERSION}"
-else
+elif [ "$BEE_GPU_VENDOR" = "amd" ]; then
     GPU_VERSION_LINE="ROCM_VERSION=${ROCM_VERSION}"
     GPU_BUILD_INFO="rocm:${ROCM_VERSION}"
+else
+    GPU_VERSION_LINE=""
+    GPU_BUILD_INFO="nogpu"
 fi
 
 cat > "${OVERLAY_STAGE_DIR}/etc/bee-release" <<EOF
@@ -414,17 +1048,27 @@ if [ -f "${OVERLAY_STAGE_DIR}/etc/motd" ]; then
     mv "${OVERLAY_STAGE_DIR}/etc/motd.patched" "${OVERLAY_STAGE_DIR}/etc/motd"
 fi
 
-# --- copy variant-specific package list into work dir ---
+# --- copy variant-specific package list, remove all other variant lists ---
+# live-build picks up ALL .list.chroot files — delete other variants to avoid conflicts.
 cp "${BUILD_WORK_DIR}/config/package-lists/bee-${BEE_GPU_VENDOR}.list.chroot" \
    "${BUILD_WORK_DIR}/config/package-lists/bee-gpu.list.chroot"
+rm -f "${BUILD_WORK_DIR}/config/package-lists/bee-nvidia.list.chroot" \
+      "${BUILD_WORK_DIR}/config/package-lists/bee-amd.list.chroot" \
+      "${BUILD_WORK_DIR}/config/package-lists/bee-nogpu.list.chroot"
 
-# --- remove archives for the other vendor ---
+# --- remove archives for the other vendor(s) ---
 if [ "$BEE_GPU_VENDOR" = "nvidia" ]; then
     rm -f "${BUILD_WORK_DIR}/config/archives/rocm.list.chroot" \
           "${BUILD_WORK_DIR}/config/archives/rocm.key.chroot"
-else
+elif [ "$BEE_GPU_VENDOR" = "amd" ]; then
     rm -f "${BUILD_WORK_DIR}/config/archives/nvidia-cuda.list.chroot" \
           "${BUILD_WORK_DIR}/config/archives/nvidia-cuda.key.chroot"
+else
+    # nogpu: remove both
+    rm -f "${BUILD_WORK_DIR}/config/archives/rocm.list.chroot" \
+          "${BUILD_WORK_DIR}/config/archives/rocm.key.chroot" \
+          "${BUILD_WORK_DIR}/config/archives/nvidia-cuda.list.chroot" \
+          "${BUILD_WORK_DIR}/config/archives/nvidia-cuda.key.chroot"
 fi
 
 # --- substitute version placeholders in package list and archive ---
@@ -432,7 +1076,7 @@ if [ "$BEE_GPU_VENDOR" = "nvidia" ]; then
     sed -i \
         -e "s/%%DCGM_VERSION%%/${DCGM_VERSION}/g" \
         "${BUILD_WORK_DIR}/config/package-lists/bee-gpu.list.chroot"
-else
+elif [ "$BEE_GPU_VENDOR" = "amd" ]; then
     sed -i \
         -e "s/%%ROCM_VERSION%%/${ROCM_VERSION}/g" \
         -e "s/%%ROCM_SMI_VERSION%%/${ROCM_SMI_VERSION}/g" \
@@ -472,9 +1116,10 @@ BEE_GPU_VENDOR_UPPER="$(echo "${BEE_GPU_VENDOR}" | tr 'a-z' 'A-Z')"
 export BEE_GPU_VENDOR_UPPER
 
 cd "${LB_DIR}"
-lb clean 2>&1 | tail -3
-lb config 2>&1 | tail -5
-lb build 2>&1
+run_step_sh "live-build clean" "80-lb-clean" "lb clean 2>&1 | tail -3"
+run_step_sh "live-build config" "81-lb-config" "lb config 2>&1 | tail -5"
+dump_memtest_debug "pre-build" "${LB_DIR}"
+run_step_sh "live-build build" "90-lb-build" "lb build 2>&1"
 
 # --- persist deb package cache back to shared location ---
 # This allows the second variant to reuse all downloaded packages.
@@ -485,8 +1130,20 @@ fi
 
 # live-build outputs live-image-amd64.hybrid.iso in LB_DIR
 ISO_RAW="${LB_DIR}/live-image-amd64.hybrid.iso"
-ISO_OUT="${DIST_DIR}/easy-bee-${BEE_GPU_VENDOR}-v${ISO_VERSION_EFFECTIVE}-amd64.iso"
 if [ -f "$ISO_RAW" ]; then
+    dump_memtest_debug "post-build" "${LB_DIR}" "$ISO_RAW"
+    if iso_memtest_present "$ISO_RAW"; then
+        :
+    else
+        memtest_status=$?
+        if [ "$memtest_status" -eq 1 ]; then
+            recover_iso_memtest "${LB_DIR}" "$ISO_RAW"
+            dump_memtest_debug "post-recovery" "${LB_DIR}" "$ISO_RAW"
+        elif [ "$memtest_status" -eq 2 ]; then
+            memtest_fail "failed to inspect ISO for memtest before recovery" "$ISO_RAW"
+        fi
+    fi
+    validate_iso_memtest "$ISO_RAW"
     cp "$ISO_RAW" "$ISO_OUT"
     echo ""
     echo "=== done (${BEE_GPU_VENDOR}) ==="

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

@@ -14,6 +14,11 @@ 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@
+}
+
 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@
@@ -24,6 +29,11 @@ menuentry "EASY-BEE (NVIDIA GSP=off)" {
     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@

iso/builder/config/bootloaders/isolinux/live.cfg.in

@@ -5,6 +5,12 @@ label live-@FLAVOUR@-normal
     initrd @INITRD@
     append @APPEND_LIVE@ bee.nvidia.mode=normal
 
+label live-@FLAVOUR@-kms
+    menu label EASY-BEE (^graphics/KMS)
+    linux @LINUX@
+    initrd @INITRD@
+    append @APPEND_LIVE@ bee.display=kms bee.nvidia.mode=normal
+
 label live-@FLAVOUR@-toram
     menu label EASY-BEE (^load to RAM)
     linux @LINUX@
@@ -17,8 +23,18 @@ label live-@FLAVOUR@-gsp-off
     initrd @INITRD@
     append @APPEND_LIVE@ nomodeset bee.nvidia.mode=gsp-off
 
+label live-@FLAVOUR@-kms-gsp-off
+    menu label EASY-BEE (g^raphics/KMS, GSP=off)
+    linux @LINUX@
+    initrd @INITRD@
+    append @APPEND_LIVE@ bee.display=kms bee.nvidia.mode=gsp-off
+
 label live-@FLAVOUR@-failsafe
     menu label EASY-BEE (^fail-safe)
     linux @LINUX@
     initrd @INITRD@
     append @APPEND_LIVE@ bee.nvidia.mode=gsp-off memtest noapic noapm nodma nomce nolapic nosmp vga=normal
+
+label memtest
+    menu label ^Memory Test (memtest86+)
+    linux /boot/memtest86+x64.bin

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

@@ -30,6 +30,7 @@ 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 ssh.service
 systemctl enable lightdm.service 2>/dev/null || true
 systemctl enable qemu-guest-agent.service 2>/dev/null || true
@@ -50,6 +51,7 @@ elif [ "$GPU_VENDOR" = "amd" ]; then
         fi
     done
 fi
+# nogpu: no GPU services needed
 
 # Ensure scripts are executable
 chmod +x /usr/local/bin/bee-network.sh  2>/dev/null || true
@@ -57,8 +59,12 @@ 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
 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
+    chmod +x /usr/local/bin/bee-john-gpu-stress 2>/dev/null || true
+    chmod +x /usr/local/bin/bee-nccl-gpu-stress 2>/dev/null || true
 fi
 
 # Reload udev rules

iso/builder/config/hooks/normal/9100-memtest.hook.binary

@@ -1,16 +1,139 @@
 #!/bin/sh
-# Copy memtest86+ binaries from chroot /boot into the ISO boot directory
-# so GRUB can chainload them directly (they must be on the ISO filesystem,
-# not inside the squashfs).
+# Ensure memtest is present in the final ISO even if live-build's built-in
+# memtest stage does not copy the binaries or expose menu entries.
 set -e
 
-echo "memtest: scanning chroot/boot/ for memtest files:"
-ls chroot/boot/memtest* 2>/dev/null || echo "memtest: WARNING: no memtest files found in chroot/boot/"
+: "${BEE_REQUIRE_MEMTEST:=0}"
 
-for f in memtest86+x64.bin memtest86+x64.efi memtest86+ia32.bin memtest86+ia32.efi; do
-    src="chroot/boot/${f}"
-    if [ -f "${src}" ]; then
-        cp "${src}" "binary/boot/${f}"
-        echo "memtest: copied ${f} to binary/boot/"
+MEMTEST_FILES="memtest86+x64.bin memtest86+x64.efi"
+BINARY_BOOT_DIR="binary/boot"
+GRUB_CFG="binary/boot/grub/grub.cfg"
+ISOLINUX_CFG="binary/isolinux/live.cfg"
+
+log() {
+    echo "memtest hook: $*"
+}
+
+fail_or_warn() {
+    msg="$1"
+    if [ "${BEE_REQUIRE_MEMTEST}" = "1" ]; then
+        log "ERROR: ${msg}"
+        exit 1
     fi
-done
+    log "WARNING: ${msg}"
+    return 0
+}
+
+copy_memtest_file() {
+    src="$1"
+    base="$(basename "$src")"
+    dst="${BINARY_BOOT_DIR}/${base}"
+
+    [ -f "$src" ] || return 1
+    mkdir -p "${BINARY_BOOT_DIR}"
+    cp "$src" "$dst"
+    log "copied ${base} from ${src}"
+}
+
+extract_memtest_from_deb() {
+    deb="$1"
+    tmpdir="$(mktemp -d)"
+
+    log "extracting memtest payload from ${deb}"
+    dpkg-deb -x "$deb" "$tmpdir"
+    for f in ${MEMTEST_FILES}; do
+        if [ -f "${tmpdir}/boot/${f}" ]; then
+            copy_memtest_file "${tmpdir}/boot/${f}"
+        fi
+    done
+    rm -rf "$tmpdir"
+}
+
+ensure_memtest_binaries() {
+    missing=0
+    for f in ${MEMTEST_FILES}; do
+        [ -f "${BINARY_BOOT_DIR}/${f}" ] || missing=1
+    done
+    [ "$missing" -eq 1 ] || return 0
+
+    for root in chroot/boot /boot; do
+        for f in ${MEMTEST_FILES}; do
+            [ -f "${BINARY_BOOT_DIR}/${f}" ] || copy_memtest_file "${root}/${f}" || true
+        done
+    done
+
+    missing=0
+    for f in ${MEMTEST_FILES}; do
+        [ -f "${BINARY_BOOT_DIR}/${f}" ] || missing=1
+    done
+    [ "$missing" -eq 1 ] || return 0
+
+    for root in cache chroot/var/cache/apt/archives /var/cache/apt/archives; do
+        [ -d "$root" ] || continue
+        deb="$(find "$root" -type f \( -name 'memtest86+_*.deb' -o -name 'memtest86+*.deb' \) 2>/dev/null | head -1)"
+        [ -n "$deb" ] || continue
+        extract_memtest_from_deb "$deb"
+        break
+    done
+
+    missing=0
+    for f in ${MEMTEST_FILES}; do
+        if [ ! -f "${BINARY_BOOT_DIR}/${f}" ]; then
+            fail_or_warn "missing ${BINARY_BOOT_DIR}/${f}"
+            missing=1
+        fi
+    done
+    [ "$missing" -eq 0 ] || return 0
+}
+
+ensure_grub_entry() {
+    [ -f "$GRUB_CFG" ] || {
+        fail_or_warn "missing ${GRUB_CFG}"
+        return 0
+    }
+
+    grep -q '### BEE MEMTEST ###' "$GRUB_CFG" && return 0
+
+    cat >> "$GRUB_CFG" <<'EOF'
+
+### BEE MEMTEST ###
+if [ "${grub_platform}" = "efi" ]; then
+    menuentry "Memory Test (memtest86+)" {
+        chainloader /boot/memtest86+x64.efi
+    }
+else
+    menuentry "Memory Test (memtest86+)" {
+        linux16 /boot/memtest86+x64.bin
+    }
+fi
+### /BEE MEMTEST ###
+EOF
+
+    log "appended memtest entry to ${GRUB_CFG}"
+}
+
+ensure_isolinux_entry() {
+    [ -f "$ISOLINUX_CFG" ] || {
+        fail_or_warn "missing ${ISOLINUX_CFG}"
+        return 0
+    }
+
+    grep -q '### BEE MEMTEST ###' "$ISOLINUX_CFG" && return 0
+
+    cat >> "$ISOLINUX_CFG" <<'EOF'
+
+# ### BEE MEMTEST ###
+label memtest
+    menu label ^Memory Test (memtest86+)
+    linux /boot/memtest86+x64.bin
+# ### /BEE MEMTEST ###
+EOF
+
+    log "appended memtest entry to ${ISOLINUX_CFG}"
+}
+
+log "ensuring memtest binaries and menu entries in binary image"
+ensure_memtest_binaries
+ensure_grub_entry
+ensure_isolinux_entry
+log "memtest assets ready"

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

@@ -1,3 +1,6 @@
+# AMD GPU firmware
+firmware-amd-graphics
+
 # AMD ROCm — GPU monitoring, bandwidth test, and compute stress (RVS GST)
 rocm-smi-lib=%%ROCM_SMI_VERSION%%
 rocm-bandwidth-test=%%ROCM_BANDWIDTH_TEST_VERSION%%

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

@@ -0,0 +1 @@
+# No GPU variant — no NVIDIA, no AMD/ROCm packages

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

@@ -1,2 +1,8 @@
-# NVIDIA DCGM (Data Center GPU Manager) — dcgmi diag for acceptance testing
-datacenter-gpu-manager=1:%%DCGM_VERSION%%
+# 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.
+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%%
+ocl-icd-libopencl1
+clinfo

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

@@ -21,8 +21,15 @@ openssh-server
 # Disk installer
 squashfs-tools
 parted
+# Keep GRUB install tools without selecting a single active platform package.
+# grub-pc and grub-efi-amd64 conflict with each other, but grub2-common
+# provides grub-install/update-grub and the *-bin packages provide BIOS/UEFI modules.
+grub2-common
 grub-pc-bin
 grub-efi-amd64-bin
+grub-efi-amd64-signed
+shim-signed
+efibootmgr
 
 # Filesystem support for USB export targets
 exfatprogs
@@ -39,11 +46,11 @@ vim-tiny
 mc
 htop
 nvtop
+btop
 sudo
 zstd
 mstflint
 memtester
-memtest86+
 stress-ng
 stressapptest
 
@@ -64,9 +71,7 @@ lightdm
 firmware-linux-free
 firmware-linux-nonfree
 firmware-misc-nonfree
-firmware-amd-graphics
 firmware-realtek
-firmware-intel-sound
 firmware-bnx2
 firmware-bnx2x
 firmware-cavium

iso/builder/smoketest.sh

@@ -52,6 +52,14 @@ else
     fail "nvidia-smi: 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: $p"
+    else
+        fail "$tool: NOT FOUND"
+    fi
+done
+
 echo ""
 echo "-- NVIDIA modules --"
 KO_DIR="/usr/local/lib/nvidia"
@@ -109,6 +117,40 @@ else
     fail "nvidia-smi: not found in PATH"
 fi
 
+echo ""
+echo "-- OpenCL / John --"
+if [ -f /etc/OpenCL/vendors/nvidia.icd ]; then
+    ok "OpenCL ICD present: /etc/OpenCL/vendors/nvidia.icd"
+else
+    fail "OpenCL ICD missing: /etc/OpenCL/vendors/nvidia.icd"
+fi
+
+if ldconfig -p 2>/dev/null | grep -q "libnvidia-opencl.so.1"; then
+    ok "libnvidia-opencl.so.1 present in linker cache"
+else
+    fail "libnvidia-opencl.so.1 missing from linker cache"
+fi
+
+if command -v clinfo >/dev/null 2>&1; then
+    if clinfo -l 2>/dev/null | grep -q "Platform"; then
+        ok "clinfo: OpenCL platform detected"
+    else
+        fail "clinfo: no OpenCL platform detected"
+    fi
+else
+    fail "clinfo: not found in PATH"
+fi
+
+if command -v john >/dev/null 2>&1; then
+    if john --list=opencl-devices 2>/dev/null | grep -q "Device #"; then
+        ok "john: OpenCL devices detected"
+    else
+        fail "john: no OpenCL devices detected"
+    fi
+else
+    fail "john: not found in PATH"
+fi
+
 echo ""
 echo "-- lib symlinks --"
 for lib in libnvidia-ml libcuda; do
@@ -129,6 +171,12 @@ 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

@@ -1,14 +1,13 @@
 [Unit]
-Description=Bee: run hardware audit
-After=bee-network.service bee-nvidia.service bee-preflight.service
-Before=bee-web.service
+Description=Bee: hardware audit
+After=bee-preflight.service bee-network.service bee-nvidia.service
 
 [Service]
 Type=oneshot
-ExecStart=/usr/local/bin/bee-log-run /appdata/bee/export/bee-audit.log /bin/sh -c '/usr/local/bin/bee audit --runtime livecd --output file:/appdata/bee/export/bee-audit.json; rc=$?; if [ "$rc" -ne 0 ]; then echo "[bee-audit] WARN: audit exited with rc=$rc"; fi; exit 0'
+RemainAfterExit=yes
+ExecStart=/usr/local/bin/bee-log-run /appdata/bee/export/bee-audit.log /usr/local/bin/bee audit --runtime auto --output file:/appdata/bee/export/bee-audit.json
 StandardOutput=journal
 StandardError=journal
-RemainAfterExit=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

@@ -1,16 +1,18 @@
 [Unit]
 Description=Bee: hardware audit web viewer
-After=bee-network.service
-Wants=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
+# Keep the web server responsive during GPU/CPU stress (children inherit nice+10
+# via Setpriority in runCmdJob, but the bee-web parent stays at 0).
+Nice=0
 
 [Install]
 WantedBy=multi-user.target

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

@@ -0,0 +1,6 @@
+[Unit]
+Wants=bee-preflight.service
+After=bee-preflight.service
+
+[Service]
+ExecStartPre=/usr/local/bin/bee-display-mode

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

@@ -4,3 +4,6 @@
 RestartSec=10
 StartLimitIntervalSec=60
 StartLimitBurst=3
+# Raise scheduling priority of the X server so the graphical console (KVM/IPMI)
+# stays responsive during GPU/CPU stress tests running at nice+10.
+Nice=-5

iso/overlay/usr/local/bin/bee-display-mode

@@ -0,0 +1,54 @@
+#!/bin/sh
+# Select Xorg display mode based on kernel cmdline.
+# Default is the current server-safe path: keep forced fbdev.
+set -eu
+
+cmdline_param() {
+    key="$1"
+    for token in $(cat /proc/cmdline 2>/dev/null); do
+        case "$token" in
+            "$key"=*)
+                echo "${token#*=}"
+                return 0
+                ;;
+        esac
+    done
+    return 1
+}
+
+log() {
+    echo "bee-display-mode: $*"
+}
+
+mode="$(cmdline_param bee.display || true)"
+if [ -z "$mode" ]; then
+    mode="safe"
+fi
+
+xorg_dir="/etc/X11/xorg.conf.d"
+fbdev_conf="${xorg_dir}/10-fbdev.conf"
+fbdev_park="${xorg_dir}/10-fbdev.conf.disabled"
+
+mkdir -p "$xorg_dir"
+
+case "$mode" in
+    kms|auto)
+        if [ -f "$fbdev_conf" ]; then
+            mv "$fbdev_conf" "$fbdev_park"
+            log "mode=${mode}; disabled forced fbdev config"
+        else
+            log "mode=${mode}; fbdev config already disabled"
+        fi
+        ;;
+    safe|fbdev|"")
+        if [ -f "$fbdev_park" ] && [ ! -f "$fbdev_conf" ]; then
+            mv "$fbdev_park" "$fbdev_conf"
+            log "mode=${mode}; restored forced fbdev config"
+        else
+            log "mode=${mode}; keeping forced fbdev config"
+        fi
+        ;;
+    *)
+        log "unknown bee.display=${mode}; keeping forced fbdev config"
+        ;;
+esac

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

@@ -0,0 +1,102 @@
+#!/bin/sh
+set -eu
+
+SECONDS=5
+SIZE_MB=0
+DEVICES=""
+EXCLUDE=""
+WORKER="/usr/local/lib/bee/bee-gpu-burn-worker"
+
+usage() {
+    echo "usage: $0 [--seconds N] [--size-mb N] [--devices 0,1] [--exclude 2,3]" >&2
+    exit 2
+}
+
+normalize_list() {
+    echo "${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},"
+}
+
+while [ "$#" -gt 0 ]; do
+    case "$1" in
+        --seconds|-t) [ "$#" -ge 2 ] || usage; 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="$2"; shift 2 ;;
+        *) usage ;;
+    esac
+done
+
+[ -x "${WORKER}" ] || { echo "bee-gpu-burn worker not found: ${WORKER}" >&2; exit 1; }
+
+ALL_DEVICES=$(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=bee-gpu-burn"
+echo "selected_gpus=${FINAL}"
+
+TMP_DIR=$(mktemp -d)
+trap 'rm -rf "${TMP_DIR}"' EXIT INT TERM
+
+WORKERS=""
+for id in $(echo "${FINAL}" | tr ',' ' '); do
+    log="${TMP_DIR}/gpu-${id}.log"
+    gpu_size_mb="${SIZE_MB}"
+    if [ "${gpu_size_mb}" -le 0 ] 2>/dev/null; then
+        total_mb=$(nvidia-smi --id="${id}" --query-gpu=memory.total --format=csv,noheader,nounits 2>/dev/null | tr -d '[:space:]')
+        if [ -n "${total_mb}" ] && [ "${total_mb}" -gt 0 ] 2>/dev/null; then
+            gpu_size_mb=$(( total_mb * 95 / 100 ))
+        else
+            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}" >"${log}" 2>&1 &
+    pid=$!
+    WORKERS="${WORKERS} ${pid}:${id}:${log}"
+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-install

@@ -12,17 +12,55 @@
 
 set -euo pipefail
 
+usage() {
+    cat >&2 <<'EOF'
+Usage: bee-install <device> [logfile]
+
+  Installs the live system to a local disk (WIPES the target).
+
+  device   Target block device, e.g. /dev/sda or /dev/nvme0n1
+           Must be a hard disk or NVMe — NOT a CD-ROM (/dev/sr*)
+  logfile  Optional path for progress log (default: /tmp/bee-install.log)
+
+Examples:
+  bee-install /dev/sda
+  bee-install /dev/nvme0n1
+  bee-install /dev/sdb /tmp/my-install.log
+
+WARNING: ALL DATA ON <device> WILL BE ERASED.
+
+Layout (UEFI):  GPT — partition 1: EFI 512MB vfat, partition 2: root ext4
+Layout (BIOS):  MBR — partition 1: root ext4
+EOF
+    exit 1
+}
+
 DEVICE="${1:-}"
 LOGFILE="${2:-/tmp/bee-install.log}"
 
-if [ -z "$DEVICE" ]; then
-    echo "Usage: bee-install <device> [logfile]" >&2
-    exit 1
+if [ -z "$DEVICE" ] || [ "$DEVICE" = "--help" ] || [ "$DEVICE" = "-h" ]; then
+    usage
 fi
 if [ ! -b "$DEVICE" ]; then
     echo "ERROR: $DEVICE is not a block device" >&2
+    echo "Run 'lsblk' to list available disks." >&2
     exit 1
 fi
+# Block CD-ROM devices
+case "$DEVICE" in
+    /dev/sr*|/dev/scd*)
+        echo "ERROR: $DEVICE is a CD-ROM/optical device — cannot install to it." >&2
+        echo "Run 'lsblk' to find the target disk (e.g. /dev/sda, /dev/nvme0n1)." >&2
+        exit 1
+        ;;
+esac
+# Check required tools
+for tool in parted mkfs.vfat mkfs.ext4 unsquashfs grub-install update-grub; do
+    if ! command -v "$tool" >/dev/null 2>&1; then
+        echo "ERROR: required tool not found: $tool" >&2
+        exit 1
+    fi
+done
 
 SQUASHFS="/run/live/medium/live/filesystem.squashfs"
 if [ ! -f "$SQUASHFS" ]; then
@@ -158,20 +196,56 @@ mount --bind /sys  "${MOUNT_ROOT}/sys"
 
 # ------------------------------------------------------------------
 log "--- Step 7/7: Installing GRUB bootloader ---"
+
+# Helper: run a chroot command, log all output, return its exit code.
+# Needed because "cmd | while" pipelines hide the exit code of cmd.
+chroot_log() {
+    local rc=0
+    local out
+    out=$(chroot "$MOUNT_ROOT" "$@" 2>&1) || rc=$?
+    echo "$out" | while IFS= read -r line; do log "  $line"; done
+    return $rc
+}
+
 if [ "$UEFI" = "1" ]; then
-    chroot "$MOUNT_ROOT" grub-install \
-        --target=x86_64-efi \
-        --efi-directory=/boot/efi \
-        --bootloader-id=bee \
-        --recheck 2>&1 | while read -r line; do log "  $line"; done || true
+    # Primary attempt: write EFI NVRAM entry (requires writable efivars)
+    if ! chroot_log grub-install \
+            --target=x86_64-efi \
+            --efi-directory=/boot/efi \
+            --bootloader-id=bee \
+            --recheck; then
+        log "  WARNING: grub-install (with NVRAM) failed — retrying with --no-nvram"
+        # --no-nvram: write grubx64.efi but skip EFI variable update.
+        # Needed on headless servers where efivars is read-only or unavailable.
+        chroot_log grub-install \
+            --target=x86_64-efi \
+            --efi-directory=/boot/efi \
+            --bootloader-id=bee \
+            --no-nvram \
+            --recheck || log "  WARNING: grub-install --no-nvram also failed — check logs"
+    fi
+
+    # Always install the UEFI fallback path EFI/BOOT/BOOTX64.EFI.
+    # Many UEFI implementations (especially server BMCs and some firmware)
+    # ignore the NVRAM boot entry and only look for this path.
+    GRUB_EFI="${MOUNT_ROOT}/boot/efi/EFI/bee/grubx64.efi"
+    FALLBACK_DIR="${MOUNT_ROOT}/boot/efi/EFI/BOOT"
+    if [ -f "$GRUB_EFI" ]; then
+        mkdir -p "$FALLBACK_DIR"
+        cp "$GRUB_EFI" "${FALLBACK_DIR}/BOOTX64.EFI"
+        log "  Fallback EFI binary installed: EFI/BOOT/BOOTX64.EFI"
+    else
+        log "  WARNING: grubx64.efi not found at $GRUB_EFI — UEFI fallback path not set"
+    fi
 else
-    chroot "$MOUNT_ROOT" grub-install \
+    chroot_log grub-install \
         --target=i386-pc \
         --recheck \
-        "$DEVICE" 2>&1 | while read -r line; do log "  $line"; done || true
+        "$DEVICE" || log "  WARNING: grub-install (BIOS) failed — check logs"
 fi
-chroot "$MOUNT_ROOT" update-grub 2>&1 | while read -r line; do log "  $line"; done || true
-log "  GRUB installed."
+
+chroot_log update-grub || log "  WARNING: update-grub failed — check logs"
+log "  GRUB step complete."
 
 # ------------------------------------------------------------------
 # Cleanup

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

@@ -0,0 +1,243 @@
+#!/bin/sh
+set -eu
+
+DURATION_SEC=300
+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: $0 [--seconds N] [--devices 0,1] [--exclude 2,3] [--format name]" >&2
+    exit 2
+}
+
+normalize_list() {
+    echo "${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},"
+}
+
+show_opencl_diagnostics() {
+    echo "-- OpenCL ICD vendors --" >&2
+    if [ -d /etc/OpenCL/vendors ]; then
+        ls -l /etc/OpenCL/vendors >&2 || true
+        for icd in /etc/OpenCL/vendors/*.icd; do
+            [ -f "${icd}" ] || continue
+            echo "  file: ${icd}" >&2
+            sed 's/^/    /' "${icd}" >&2 || true
+        done
+    else
+        echo "  /etc/OpenCL/vendors is missing" >&2
+    fi
+    echo "-- NVIDIA device nodes --" >&2
+    ls -l /dev/nvidia* >&2 || true
+    echo "-- ldconfig OpenCL/NVIDIA --" >&2
+    ldconfig -p 2>/dev/null | grep 'libOpenCL\|libcuda\|libnvidia-opencl' >&2 || true
+    if command -v clinfo >/dev/null 2>&1; then
+        echo "-- clinfo -l --" >&2
+        clinfo -l >&2 || true
+    fi
+    echo "-- john --list=opencl-devices --" >&2
+    ./john --list=opencl-devices >&2 || true
+}
+
+refresh_nvidia_runtime() {
+    if [ "$(id -u)" != "0" ]; then
+        return 1
+    fi
+    if command -v bee-nvidia-load >/dev/null 2>&1; then
+        bee-nvidia-load >/dev/null 2>&1 || true
+    fi
+    ldconfig >/dev/null 2>&1 || true
+    return 0
+}
+
+ensure_nvidia_uvm() {
+    if lsmod 2>/dev/null | grep -q '^nvidia_uvm '; then
+        return 0
+    fi
+    if [ "$(id -u)" != "0" ]; then
+        return 1
+    fi
+
+    ko="/usr/local/lib/nvidia/nvidia-uvm.ko"
+    [ -f "${ko}" ] || return 1
+
+    if ! insmod "${ko}" >/dev/null 2>&1; then
+        return 1
+    fi
+
+    uvm_major=$(grep -m1 ' nvidia-uvm$' /proc/devices | awk '{print $1}')
+    if [ -n "${uvm_major}" ]; then
+        mknod -m 666 /dev/nvidia-uvm c "${uvm_major}" 0 2>/dev/null || true
+        mknod -m 666 /dev/nvidia-uvm-tools c "${uvm_major}" 1 2>/dev/null || true
+    fi
+    return 0
+}
+
+ensure_opencl_ready() {
+    out=$(./john --list=opencl-devices 2>&1 || true)
+    if echo "${out}" | grep -q "Device #"; then
+        return 0
+    fi
+
+    if refresh_nvidia_runtime; then
+        out=$(./john --list=opencl-devices 2>&1 || true)
+        if echo "${out}" | grep -q "Device #"; then
+            return 0
+        fi
+    fi
+
+    if ensure_nvidia_uvm; then
+        out=$(./john --list=opencl-devices 2>&1 || true)
+        if echo "${out}" | grep -q "Device #"; then
+            return 0
+        fi
+    fi
+
+    echo "OpenCL devices are not available for John." >&2
+    if ! lsmod 2>/dev/null | grep -q '^nvidia_uvm '; then
+        echo "nvidia_uvm is not loaded." >&2
+    fi
+    if [ ! -e /dev/nvidia-uvm ]; then
+        echo "/dev/nvidia-uvm is missing." >&2
+    fi
+    show_opencl_diagnostics
+    return 1
+}
+
+while [ "$#" -gt 0 ]; do
+    case "$1" in
+        --seconds|-t) [ "$#" -ge 2 ] || usage; DURATION_SEC="$2"; shift 2 ;;
+        --devices) [ "$#" -ge 2 ] || usage; DEVICES="$2"; shift 2 ;;
+        --exclude) [ "$#" -ge 2 ] || usage; EXCLUDE="$2"; shift 2 ;;
+        --format) [ "$#" -ge 2 ] || usage; FORMAT="$2"; shift 2 ;;
+        *) usage ;;
+    esac
+done
+
+[ -x "${JOHN_BIN}" ] || { echo "john binary not found: ${JOHN_BIN}" >&2; exit 1; }
+
+ALL_DEVICES=$(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; }
+
+JOHN_DEVICES=""
+for id in $(echo "${FINAL}" | tr ',' ' '); do
+    opencl_id=$((id + 1))
+    if [ -z "${JOHN_DEVICES}" ]; then
+        JOHN_DEVICES="${opencl_id}"
+    else
+        JOHN_DEVICES="${JOHN_DEVICES},${opencl_id}"
+    fi
+done
+
+echo "loader=john"
+echo "selected_gpus=${FINAL}"
+echo "john_devices=${JOHN_DEVICES}"
+
+cd "${JOHN_DIR}"
+
+ensure_opencl_ready || exit 1
+
+choose_format() {
+    if [ -n "${FORMAT}" ]; then
+        echo "${FORMAT}"
+        return 0
+    fi
+    for candidate in sha512crypt-opencl pbkdf2-hmac-sha512-opencl 7z-opencl sha256crypt-opencl md5crypt-opencl; do
+        if ./john --test=1 --format="${candidate}" --devices="${JOHN_DEVICES}" >/dev/null 2>&1; then
+            echo "${candidate}"
+            return 0
+        fi
+    done
+    return 1
+}
+
+CHOSEN_FORMAT=$(choose_format) || {
+    echo "no suitable john OpenCL format found" >&2
+    ./john --list=opencl-devices >&2 || true
+    exit 1
+}
+
+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}"
+DEADLINE=$(( $(date +%s) + DURATION_SEC ))
+_first=1
+for opencl_id in $(echo "${JOHN_DEVICES}" | tr ',' ' '); do
+    [ "${_first}" = "1" ] || sleep 3
+    _first=0
+    run_john_loop "${opencl_id}" "${DEADLINE}" &
+    pid=$!
+    PIDS="${PIDS} ${pid}"
+done
+FAIL=0
+for pid in ${PIDS}; do
+    wait "${pid}" || FAIL=$((FAIL+1))
+done
+[ "${FAIL}" -eq 0 ] || { echo "john: ${FAIL} device(s) failed" >&2; exit 1; }

iso/overlay/usr/local/bin/bee-log-run

@@ -17,7 +17,7 @@ mkdir -p "$(dirname "$log_file")"
 serial_sink() {
     local tty="$1"
     if [ -w "$tty" ]; then
-        cat > "$tty"
+        cat > "$tty" 2>/dev/null || true
     else
         cat > /dev/null
     fi

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

@@ -0,0 +1,91 @@
+#!/bin/sh
+set -eu
+
+SECONDS=300
+DEVICES=""
+EXCLUDE=""
+MIN_BYTES="512M"
+MAX_BYTES="4G"
+FACTOR="2"
+ITERS="20"
+ALL_REDUCE_BIN="/usr/local/bin/all_reduce_perf"
+
+usage() {
+    echo "usage: $0 [--seconds N] [--devices 0,1] [--exclude 2,3]" >&2
+    exit 2
+}
+
+normalize_list() {
+    echo "${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},"
+}
+
+while [ "$#" -gt 0 ]; do
+    case "$1" in
+        --seconds|-t) [ "$#" -ge 2 ] || usage; SECONDS="$2"; shift 2 ;;
+        --devices) [ "$#" -ge 2 ] || usage; DEVICES="$2"; shift 2 ;;
+        --exclude) [ "$#" -ge 2 ] || usage; EXCLUDE="$2"; shift 2 ;;
+        *) usage ;;
+    esac
+done
+
+[ -x "${ALL_REDUCE_BIN}" ] || { echo "all_reduce_perf not found: ${ALL_REDUCE_BIN}" >&2; exit 1; }
+
+ALL_DEVICES=$(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; }
+
+GPU_COUNT=$(echo "${FINAL}" | tr ',' '\n' | awk 'NF' | wc -l | awk '{print $1}')
+[ "${GPU_COUNT}" -gt 0 ] || { echo "selected GPU count is zero" >&2; exit 1; }
+
+echo "loader=nccl"
+echo "selected_gpus=${FINAL}"
+echo "gpu_count=${GPU_COUNT}"
+echo "range=${MIN_BYTES}..${MAX_BYTES}"
+echo "iters=${ITERS}"
+
+deadline=$(( $(date +%s) + SECONDS ))
+round=0
+
+while :; do
+    now=$(date +%s)
+    if [ "${now}" -ge "${deadline}" ]; then
+        break
+    fi
+    round=$((round + 1))
+    remaining=$((deadline - now))
+    echo "round=${round} remaining_sec=${remaining}"
+    CUDA_VISIBLE_DEVICES="${FINAL}" \
+        "${ALL_REDUCE_BIN}" \
+        -b "${MIN_BYTES}" \
+        -e "${MAX_BYTES}" \
+        -f "${FACTOR}" \
+        -g "${GPU_COUNT}" \
+        --iters "${ITERS}"
+done

iso/overlay/usr/local/bin/bee-network.sh

@@ -6,25 +6,66 @@ LOG_PREFIX="bee-network"
 
 log() { echo "[$LOG_PREFIX] $*"; }
 
-# find physical interfaces: exclude lo and virtual (docker/virbr/veth/tun/tap)
-interfaces=$(ip -o link show \
-    | awk -F': ' '{print $2}' \
-    | grep -v '^lo$' \
-    | grep -vE '^(docker|virbr|veth|tun|tap|br-|bond|dummy)' \
-    | sort)
+list_interfaces() {
+    ip -o link show \
+        | awk -F': ' '{print $2}' \
+        | grep -v '^lo$' \
+        | grep -vE '^(docker|virbr|veth|tun|tap|br-|bond|dummy)' \
+        | sort
+}
 
-if [ -z "$interfaces" ]; then
+# Give udev a short chance to expose late NICs before the first scan.
+if command -v udevadm >/dev/null 2>&1; then
+    udevadm settle --timeout=5 >/dev/null 2>&1 || log "WARN: udevadm settle timed out"
+fi
+
+started_ifaces=""
+started_count=0
+scan_pass=1
+
+# Some server NICs appear a bit later after module/firmware init. Do a small
+# bounded rescan window without turning network bring-up into a boot blocker.
+while [ "$scan_pass" -le 3 ]; do
+    interfaces=$(list_interfaces)
+
+    if [ -n "$interfaces" ]; then
+        for iface in $interfaces; do
+            case " $started_ifaces " in
+                *" $iface "*) continue ;;
+            esac
+
+            log "bringing up $iface"
+            if ! ip link set "$iface" up; then
+                log "WARN: could not bring up $iface"
+                continue
+            fi
+
+            carrier=$(cat "/sys/class/net/$iface/carrier" 2>/dev/null || true)
+            if [ "$carrier" = "1" ]; then
+                log "carrier detected on $iface"
+            else
+                log "carrier not detected yet on $iface"
+            fi
+
+            # DHCP in background — non-blocking, keep dhclient verbose output in the service log.
+            dhclient -4 -v -nw "$iface" &
+            log "DHCP started for $iface (pid $!)"
+
+            started_ifaces="$started_ifaces $iface"
+            started_count=$((started_count + 1))
+        done
+    fi
+
+    if [ "$scan_pass" -ge 3 ]; then
+        break
+    fi
+    scan_pass=$((scan_pass + 1))
+    sleep 2
+done
+
+if [ "$started_count" -eq 0 ]; then
     log "no physical interfaces found"
     exit 0
 fi
 
-for iface in $interfaces; do
-    log "bringing up $iface"
-    ip link set "$iface" up || { log "WARN: could not bring up $iface"; continue; }
-
-    # DHCP in background — non-blocking, keep dhclient verbose output in the service log.
-    dhclient -4 -v -nw "$iface" &
-    log "DHCP started for $iface (pid $!)"
-done
-
-log "done"
+log "done (interfaces started: $started_count)"

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

@@ -59,15 +59,28 @@ load_module() {
     return 1
 }
 
+load_host_module() {
+    mod="$1"
+    if modprobe "$mod" >/dev/null 2>&1; then
+        log "host module loaded: $mod"
+        return 0
+    fi
+    return 1
+}
+
 case "$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|*)
+    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.
@@ -76,6 +89,15 @@ case "$nvidia_mode" in
         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
 
 # Create /dev/nvidia* device nodes (udev rules absent since we use .run installer)
@@ -105,4 +127,38 @@ fi
 ldconfig 2>/dev/null || true
 log "ldconfig refreshed"
 
+# Start DCGM host engine so dcgmi can discover GPUs.
+# 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
+        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
+else
+    log "WARN: nv-hostengine not found — dcgmi diagnostics will not work"
+fi
+
 log "done"

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

@@ -8,20 +8,23 @@ xset -dpms
 xset s noblank
 
 tint2 &
-# Wait for bee-web to bind (Go starts fast, usually <2s)
+
+# 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 30 ]; do
+while [ $i -lt 120 ]; do
     if curl -sf http://localhost/healthz >/dev/null 2>&1; then break; fi
     sleep 1
     i=$((i+1))
 done
+
 chromium \
     --disable-infobars \
     --disable-translate \
     --no-first-run \
     --disable-session-crashed-bubble \
     --disable-features=TranslateUI \
-    --start-fullscreen \
+    --start-maximized \
     http://localhost/ &
 
 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 -nn 2>/dev/null | grep -qi '10de:'
+}
+
+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"