Fix GPU clock lock normalization for Blackwell (clocks.max.* unsupported)

clocks.max.graphics / clocks.max.memory CSV fields return exit status 2 on
RTX PRO 6000 Blackwell (driver 98.x), causing the entire gpu inventory query
to fail and clock lock to be skipped → normalization: partial.

Fix:
- Add minimal fallback query (index,uuid,name,pci.bus_id,vbios_version,
  power.limit) that succeeds even without clock fields
- Add enrichGPUInfoWithMaxClocks: parses "Max Clocks" section of
  nvidia-smi -q verbose output to fill MaxGraphicsClockMHz /
  MaxMemoryClockMHz when CSV fields fail
- Move nvidia-smi -q execution before queryBenchmarkGPUInfo so its output
  is available for clock enrichment immediately after
- Tests: cover enrichment and skip-if-populated cases

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

audit/cmd/bee/main.go

@@ -382,9 +382,9 @@ func runSAT(args []string, stdout, stderr io.Writer) int {
 			archive, err = application.RunNvidiaAcceptancePack("", logLine)
 		}
 	case "memory":
-		archive, err = application.RunMemoryAcceptancePackCtx(context.Background(), "", logLine)
+		archive, err = application.RunMemoryAcceptancePackCtx(context.Background(), "", 256, 1, logLine)
 	case "storage":
-		archive, err = application.RunStorageAcceptancePackCtx(context.Background(), "", logLine)
+		archive, err = application.RunStorageAcceptancePackCtx(context.Background(), "", false, logLine)
 	case "cpu":
 		dur := *duration
 		if dur <= 0 {

audit/internal/app/app.go

@@ -117,13 +117,15 @@ type satRunner interface {
 	RunNvidiaAcceptancePackWithOptions(ctx context.Context, baseDir string, diagLevel int, gpuIndices []int, logFunc func(string)) (string, error)
 	RunNvidiaTargetedStressValidatePack(ctx context.Context, baseDir string, durationSec int, gpuIndices []int, logFunc func(string)) (string, error)
 	RunNvidiaBenchmark(ctx context.Context, baseDir string, opts platform.NvidiaBenchmarkOptions, logFunc func(string)) (string, error)
-	RunNvidiaOfficialComputePack(ctx context.Context, baseDir string, durationSec int, gpuIndices []int, logFunc func(string)) (string, error)
+	RunNvidiaOfficialComputePack(ctx context.Context, baseDir string, durationSec int, gpuIndices []int, staggerSec int, logFunc func(string)) (string, error)
 	RunNvidiaTargetedPowerPack(ctx context.Context, baseDir string, durationSec int, gpuIndices []int, logFunc func(string)) (string, error)
 	RunNvidiaPulseTestPack(ctx context.Context, baseDir string, durationSec int, gpuIndices []int, logFunc func(string)) (string, error)
 	RunNvidiaBandwidthPack(ctx context.Context, baseDir string, gpuIndices []int, logFunc func(string)) (string, error)
 	RunNvidiaStressPack(ctx context.Context, baseDir string, opts platform.NvidiaStressOptions, logFunc func(string)) (string, error)
-	RunMemoryAcceptancePack(ctx context.Context, baseDir string, logFunc func(string)) (string, error)
-	RunStorageAcceptancePack(ctx context.Context, baseDir string, logFunc func(string)) (string, error)
+	ListNvidiaGPUStatuses() ([]platform.NvidiaGPUStatus, error)
+	ResetNvidiaGPU(index int) (string, error)
+	RunMemoryAcceptancePack(ctx context.Context, baseDir string, sizeMB, passes int, logFunc func(string)) (string, error)
+	RunStorageAcceptancePack(ctx context.Context, baseDir string, extended bool, logFunc func(string)) (string, error)
 	RunCPUAcceptancePack(ctx context.Context, baseDir string, durationSec int, logFunc func(string)) (string, error)
 	ListNvidiaGPUs() ([]platform.NvidiaGPU, error)
 	DetectGPUVendor() string
@@ -188,6 +190,7 @@ func (a *App) RunAudit(runtimeMode runtimeenv.Mode, output string) (string, erro
 	}
 	result := collector.Run(runtimeMode)
 	applyLatestSATStatuses(&result.Hardware, DefaultSATBaseDir, a.StatusDB)
+	writePSUStatusesToDB(a.StatusDB, result.Hardware.PowerSupplies)
 	if health, err := ReadRuntimeHealth(DefaultRuntimeJSONPath); err == nil {
 		result.Runtime = &health
 	}
@@ -521,6 +524,15 @@ func (a *App) ListNvidiaGPUs() ([]platform.NvidiaGPU, error) {
 	return a.sat.ListNvidiaGPUs()
 }
 
+func (a *App) ListNvidiaGPUStatuses() ([]platform.NvidiaGPUStatus, error) {
+	return a.sat.ListNvidiaGPUStatuses()
+}
+
+func (a *App) ResetNvidiaGPU(index int) (ActionResult, error) {
+	out, err := a.sat.ResetNvidiaGPU(index)
+	return ActionResult{Title: fmt.Sprintf("Reset NVIDIA GPU %d", index), Body: strings.TrimSpace(out)}, err
+}
+
 func (a *App) RunNvidiaAcceptancePackWithOptions(ctx context.Context, baseDir string, diagLevel int, gpuIndices []int, logFunc func(string)) (ActionResult, error) {
 	if strings.TrimSpace(baseDir) == "" {
 		baseDir = DefaultSATBaseDir
@@ -555,11 +567,11 @@ func (a *App) RunNvidiaBenchmarkCtx(ctx context.Context, baseDir string, opts pl
 	return a.sat.RunNvidiaBenchmark(ctx, baseDir, opts, logFunc)
 }
 
-func (a *App) RunNvidiaOfficialComputePack(ctx context.Context, baseDir string, durationSec int, gpuIndices []int, logFunc func(string)) (string, error) {
+func (a *App) RunNvidiaOfficialComputePack(ctx context.Context, baseDir string, durationSec int, gpuIndices []int, staggerSec int, logFunc func(string)) (string, error) {
 	if strings.TrimSpace(baseDir) == "" {
 		baseDir = DefaultSATBaseDir
 	}
-	return a.sat.RunNvidiaOfficialComputePack(ctx, baseDir, durationSec, gpuIndices, logFunc)
+	return a.sat.RunNvidiaOfficialComputePack(ctx, baseDir, durationSec, gpuIndices, staggerSec, logFunc)
 }
 
 func (a *App) RunNvidiaTargetedPowerPack(ctx context.Context, baseDir string, durationSec int, gpuIndices []int, logFunc func(string)) (string, error) {
@@ -591,14 +603,14 @@ func (a *App) RunNvidiaStressPackCtx(ctx context.Context, baseDir string, opts p
 }
 
 func (a *App) RunMemoryAcceptancePack(baseDir string, logFunc func(string)) (string, error) {
-	return a.RunMemoryAcceptancePackCtx(context.Background(), baseDir, logFunc)
+	return a.RunMemoryAcceptancePackCtx(context.Background(), baseDir, 256, 1, logFunc)
 }
 
-func (a *App) RunMemoryAcceptancePackCtx(ctx context.Context, baseDir string, logFunc func(string)) (string, error) {
+func (a *App) RunMemoryAcceptancePackCtx(ctx context.Context, baseDir string, sizeMB, passes int, logFunc func(string)) (string, error) {
 	if strings.TrimSpace(baseDir) == "" {
 		baseDir = DefaultSATBaseDir
 	}
-	return a.sat.RunMemoryAcceptancePack(ctx, baseDir, logFunc)
+	return a.sat.RunMemoryAcceptancePack(ctx, baseDir, sizeMB, passes, logFunc)
 }
 
 func (a *App) RunMemoryAcceptancePackResult(baseDir string) (ActionResult, error) {
@@ -623,14 +635,14 @@ func (a *App) RunCPUAcceptancePackResult(baseDir string, durationSec int) (Actio
 }
 
 func (a *App) RunStorageAcceptancePack(baseDir string, logFunc func(string)) (string, error) {
-	return a.RunStorageAcceptancePackCtx(context.Background(), baseDir, logFunc)
+	return a.RunStorageAcceptancePackCtx(context.Background(), baseDir, false, logFunc)
 }
 
-func (a *App) RunStorageAcceptancePackCtx(ctx context.Context, baseDir string, logFunc func(string)) (string, error) {
+func (a *App) RunStorageAcceptancePackCtx(ctx context.Context, baseDir string, extended bool, logFunc func(string)) (string, error) {
 	if strings.TrimSpace(baseDir) == "" {
 		baseDir = DefaultSATBaseDir
 	}
-	return a.sat.RunStorageAcceptancePack(ctx, baseDir, logFunc)
+	return a.sat.RunStorageAcceptancePack(ctx, baseDir, extended, logFunc)
 }
 
 func (a *App) RunStorageAcceptancePackResult(baseDir string) (ActionResult, error) {
@@ -915,6 +927,41 @@ func bodyOr(body, fallback string) string {
 	return body
 }
 
+// writePSUStatusesToDB records PSU statuses collected during audit into the
+// component-status DB so they are visible in the Hardware Summary card.
+// PSU status is sourced from IPMI (ipmitool fru + sdr) during audit.
+func writePSUStatusesToDB(db *ComponentStatusDB, psus []schema.HardwarePowerSupply) {
+	if db == nil || len(psus) == 0 {
+		return
+	}
+	const source = "audit:ipmi"
+	worstStatus := "OK"
+	for _, psu := range psus {
+		if psu.Status == nil {
+			continue
+		}
+		slot := "?"
+		if psu.Slot != nil {
+			slot = *psu.Slot
+		}
+		st := *psu.Status
+		detail := ""
+		if psu.ErrorDescription != nil {
+			detail = *psu.ErrorDescription
+		}
+		db.Record("psu:"+slot, source, st, detail)
+		switch st {
+		case "Critical":
+			worstStatus = "Critical"
+		case "Warning":
+			if worstStatus != "Critical" {
+				worstStatus = "Warning"
+			}
+		}
+	}
+	db.Record("psu:all", source, worstStatus, "")
+}
+
 func ReadRuntimeHealth(path string) (schema.RuntimeHealth, error) {
 	raw, err := os.ReadFile(path)
 	if err != nil {

audit/internal/app/app_test.go

@@ -135,6 +135,8 @@ type fakeSAT struct {
 	listAMDGPUsFn             func() ([]platform.AMDGPUInfo, error)
 	runAMDPackFn              func(string) (string, error)
 	listNvidiaGPUsFn          func() ([]platform.NvidiaGPU, error)
+	listNvidiaGPUStatusesFn   func() ([]platform.NvidiaGPUStatus, error)
+	resetNvidiaGPUFn          func(int) (string, error)
 }
 
 func (f fakeSAT) RunNvidiaAcceptancePack(baseDir string, _ func(string)) (string, error) {
@@ -159,7 +161,7 @@ func (f fakeSAT) RunNvidiaTargetedStressValidatePack(_ context.Context, baseDir
 	return f.runNvidiaFn(baseDir)
 }
 
-func (f fakeSAT) RunNvidiaOfficialComputePack(_ context.Context, baseDir string, durationSec int, gpuIndices []int, _ func(string)) (string, error) {
+func (f fakeSAT) RunNvidiaOfficialComputePack(_ context.Context, baseDir string, durationSec int, gpuIndices []int, _ int, _ func(string)) (string, error) {
 	if f.runNvidiaComputeFn != nil {
 		return f.runNvidiaComputeFn(baseDir, durationSec, gpuIndices)
 	}
@@ -201,11 +203,25 @@ func (f fakeSAT) ListNvidiaGPUs() ([]platform.NvidiaGPU, error) {
 	return nil, nil
 }
 
-func (f fakeSAT) RunMemoryAcceptancePack(_ context.Context, baseDir string, _ func(string)) (string, error) {
+func (f fakeSAT) ListNvidiaGPUStatuses() ([]platform.NvidiaGPUStatus, error) {
+	if f.listNvidiaGPUStatusesFn != nil {
+		return f.listNvidiaGPUStatusesFn()
+	}
+	return nil, nil
+}
+
+func (f fakeSAT) ResetNvidiaGPU(index int) (string, error) {
+	if f.resetNvidiaGPUFn != nil {
+		return f.resetNvidiaGPUFn(index)
+	}
+	return "", nil
+}
+
+func (f fakeSAT) RunMemoryAcceptancePack(_ context.Context, baseDir string, _, _ int, _ func(string)) (string, error) {
 	return f.runMemoryFn(baseDir)
 }
 
-func (f fakeSAT) RunStorageAcceptancePack(_ context.Context, baseDir string, _ func(string)) (string, error) {
+func (f fakeSAT) RunStorageAcceptancePack(_ context.Context, baseDir string, _ bool, _ func(string)) (string, error) {
 	return f.runStorageFn(baseDir)
 }
 
@@ -526,8 +542,6 @@ func TestActionResultsUseFallbackBody(t *testing.T) {
 }
 
 func TestExportSupportBundleResultMentionsUnmountedUSB(t *testing.T) {
-	t.Parallel()
-
 	tmp := t.TempDir()
 	oldExportDir := DefaultExportDir
 	DefaultExportDir = tmp
@@ -564,8 +578,6 @@ func TestExportSupportBundleResultMentionsUnmountedUSB(t *testing.T) {
 }
 
 func TestExportSupportBundleResultDoesNotPretendSuccessOnError(t *testing.T) {
-	t.Parallel()
-
 	tmp := t.TempDir()
 	oldExportDir := DefaultExportDir
 	DefaultExportDir = tmp
@@ -627,8 +639,6 @@ func TestRunNvidiaAcceptancePackResult(t *testing.T) {
 }
 
 func TestRunSATDefaultsToExportDir(t *testing.T) {
-	t.Parallel()
-
 	oldSATBaseDir := DefaultSATBaseDir
 	DefaultSATBaseDir = "/tmp/export/bee-sat"
 	t.Cleanup(func() { DefaultSATBaseDir = oldSATBaseDir })
@@ -805,6 +815,9 @@ func TestBuildSupportBundleIncludesExportDirContents(t *testing.T) {
 	for _, want := range []string{
 		"/system/ip-link.txt",
 		"/system/ip-link-stats.txt",
+		"/system/kernel-aer-nvidia.txt",
+		"/system/lspci-nvidia-bridges-vv.txt",
+		"/system/pcie-aer-sysfs.txt",
 		"/system/ethtool-info.txt",
 		"/system/ethtool-link.txt",
 		"/system/ethtool-module.txt",

audit/internal/app/sat_overlay.go

@@ -3,6 +3,7 @@ package app
 import (
 	"os"
 	"path/filepath"
+	"strconv"
 	"sort"
 	"strings"
 
@@ -18,6 +19,7 @@ func applyLatestSATStatuses(snap *schema.HardwareSnapshot, baseDir string, db *C
 	}
 	if summary, ok := loadLatestSATSummary(baseDir, "gpu-nvidia-"); ok {
 		applyGPUVendorSAT(snap.PCIeDevices, "nvidia", summary)
+		applyNvidiaPerGPUStatus(snap.PCIeDevices, baseDir)
 	}
 	if summary, ok := loadLatestSATSummary(baseDir, "memory-"); ok {
 		applyMemorySAT(snap.Memory, summary)
@@ -32,6 +34,100 @@ func applyLatestSATStatuses(snap *schema.HardwareSnapshot, baseDir string, db *C
 	applyComponentStatusDB(snap, db)
 }
 
+type nvidiaPerGPUStatus struct {
+	runStatus string
+	reason    string
+}
+
+func applyNvidiaPerGPUStatus(devs []schema.HardwarePCIeDevice, baseDir string) {
+	statusByIndex, ts, ok := loadLatestNvidiaPerGPUStatus(baseDir)
+	if !ok {
+		return
+	}
+	for i := range devs {
+		if devs[i].Telemetry == nil {
+			continue
+		}
+		rawIdx, ok := devs[i].Telemetry["nvidia_gpu_index"]
+		if !ok {
+			continue
+		}
+		idx, ok := telemetryInt(rawIdx)
+		if !ok {
+			continue
+		}
+		st, ok := statusByIndex[idx]
+		if !ok {
+			continue
+		}
+		status, description, ok := satKeyStatus(st.runStatus, firstNonEmpty(strings.TrimSpace(st.reason), "nvidia GPU SAT"))
+		if !ok {
+			continue
+		}
+		mergeComponentStatusPreferDetail(&devs[i].HardwareComponentStatus, ts, status, description)
+	}
+}
+
+func loadLatestNvidiaPerGPUStatus(baseDir string) (map[int]nvidiaPerGPUStatus, string, bool) {
+	matches, err := filepath.Glob(filepath.Join(baseDir, "gpu-nvidia-*"))
+	if err != nil || len(matches) == 0 {
+		return nil, "", false
+	}
+	sort.Strings(matches)
+	runDir := matches[len(matches)-1]
+	summaryRaw, err := os.ReadFile(filepath.Join(runDir, "summary.txt"))
+	if err != nil {
+		return nil, "", false
+	}
+	summaryKV := parseKeyValueSummary(string(summaryRaw))
+	runAtUTC := strings.TrimSpace(summaryKV["run_at_utc"])
+	files, err := filepath.Glob(filepath.Join(runDir, "gpu-*-status.txt"))
+	if err != nil || len(files) == 0 {
+		return nil, "", false
+	}
+	out := make(map[int]nvidiaPerGPUStatus, len(files))
+	for _, file := range files {
+		raw, err := os.ReadFile(file)
+		if err != nil {
+			continue
+		}
+		kv := parseKeyValueSummary(string(raw))
+		idx, err := strconv.Atoi(strings.TrimSpace(kv["gpu_index"]))
+		if err != nil {
+			continue
+		}
+		out[idx] = nvidiaPerGPUStatus{
+			runStatus: strings.ToUpper(strings.TrimSpace(kv["run_status"])),
+			reason:    strings.TrimSpace(kv["reason"]),
+		}
+	}
+	if len(out) == 0 {
+		return nil, "", false
+	}
+	return out, runAtUTC, true
+}
+
+func telemetryInt(v any) (int, bool) {
+	switch value := v.(type) {
+	case int:
+		return value, true
+	case int32:
+		return int(value), true
+	case int64:
+		return int(value), true
+	case float64:
+		return int(value), true
+	case string:
+		n, err := strconv.Atoi(strings.TrimSpace(value))
+		if err != nil {
+			return 0, false
+		}
+		return n, true
+	default:
+		return 0, false
+	}
+}
+
 type satSummary struct {
 	runAtUTC string
 	overall  string
@@ -176,6 +272,31 @@ func mergeComponentStatus(component *schema.HardwareComponentStatus, changedAt,
 	}
 }
 
+func mergeComponentStatusPreferDetail(component *schema.HardwareComponentStatus, changedAt, satStatus, description string) {
+	if component == nil || satStatus == "" {
+		return
+	}
+	current := strings.TrimSpace(ptrString(component.Status))
+	newSeverity := statusSeverity(satStatus)
+	currentSeverity := statusSeverity(current)
+	if current == "" || current == "Unknown" || newSeverity > currentSeverity {
+		mergeComponentStatus(component, changedAt, satStatus, description)
+		return
+	}
+	if newSeverity == currentSeverity && strings.TrimSpace(description) != "" {
+		component.Status = appStringPtr(satStatus)
+		component.ErrorDescription = appStringPtr(description)
+		if strings.TrimSpace(changedAt) != "" {
+			component.StatusChangedAt = appStringPtr(changedAt)
+			component.StatusHistory = append(component.StatusHistory, schema.HardwareStatusHistory{
+				Status:    satStatus,
+				ChangedAt: changedAt,
+				Details:   appStringPtr(description),
+			})
+		}
+	}
+}
+
 func statusSeverity(status string) int {
 	switch strings.TrimSpace(status) {
 	case "Critical":

audit/internal/app/sat_overlay_test.go

@@ -59,3 +59,51 @@ func TestApplyLatestSATStatusesMarksAMDGPUs(t *testing.T) {
 		t.Fatalf("gpu status=%v want Critical", snap.PCIeDevices[0].Status)
 	}
 }
+
+func TestApplyLatestSATStatusesMarksNvidiaGPUByPerGPUStatusFile(t *testing.T) {
+	baseDir := t.TempDir()
+	runDir := filepath.Join(baseDir, "gpu-nvidia-20260407-162123")
+	if err := os.MkdirAll(runDir, 0755); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(filepath.Join(runDir, "summary.txt"), []byte("run_at_utc=2026-04-07T16:21:23Z\noverall_status=FAILED\n"), 0644); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(filepath.Join(runDir, "gpu-1-status.txt"), []byte("gpu_index=1\ngpu_name=NVIDIA H100 PCIe\nrun_status=FAILED\nreason=GPU requires reset\n"), 0644); err != nil {
+		t.Fatal(err)
+	}
+
+	class := "VideoController"
+	manufacturer := "NVIDIA Corporation"
+	bdf0 := "0000:4b:00.0"
+	bdf1 := "0000:4f:00.0"
+	snap := schema.HardwareSnapshot{
+		PCIeDevices: []schema.HardwarePCIeDevice{
+			{
+				DeviceClass:  &class,
+				Manufacturer: &manufacturer,
+				BDF:          &bdf0,
+				Telemetry:    map[string]any{"nvidia_gpu_index": 0},
+			},
+			{
+				DeviceClass:  &class,
+				Manufacturer: &manufacturer,
+				BDF:          &bdf1,
+				Telemetry:    map[string]any{"nvidia_gpu_index": 1},
+			},
+		},
+	}
+
+	applyLatestSATStatuses(&snap, baseDir, nil)
+
+	if snap.PCIeDevices[1].Status == nil || *snap.PCIeDevices[1].Status != "Critical" {
+		t.Fatalf("gpu1 status=%v want Critical", snap.PCIeDevices[1].Status)
+	}
+	if snap.PCIeDevices[1].ErrorDescription == nil || *snap.PCIeDevices[1].ErrorDescription != "GPU requires reset failed" {
+		got := "<nil>"
+		if snap.PCIeDevices[1].ErrorDescription != nil {
+			got = *snap.PCIeDevices[1].ErrorDescription
+		}
+		t.Fatalf("gpu1 error=%q want per-gpu reason", got)
+	}
+}

audit/internal/app/support_bundle.go

@@ -40,17 +40,75 @@ var supportBundleCommands = []struct {
 	{name: "system/mount.txt", cmd: []string{"mount"}},
 	{name: "system/df-h.txt", cmd: []string{"df", "-h"}},
 	{name: "system/dmesg.txt", cmd: []string{"dmesg"}},
+	{name: "system/kernel-aer-nvidia.txt", cmd: []string{"sh", "-c", `
+if command -v dmesg >/dev/null 2>&1; then
+  dmesg | grep -iE 'AER|NVRM|Xid|pcieport|nvidia' || echo "no AER/NVRM/Xid kernel messages found"
+else
+  echo "dmesg not found"
+fi
+`}},
 	{name: "system/nvidia-smi-q.txt", cmd: []string{"nvidia-smi", "-q"}},
+	{name: "system/lspci-nvidia-bridges-vv.txt", cmd: []string{"sh", "-c", `
+if ! command -v lspci >/dev/null 2>&1; then
+  echo "lspci not found"
+  exit 0
+fi
+found=0
+	for gpu in $(lspci -Dn | awk '$2 ~ /^03(00|02):$/ && $3 ~ /^10de:/ {print $1}'); do
+  found=1
+  echo "=== GPU $gpu ==="
+  lspci -s "$gpu" -vv 2>&1 || true
+  bridge=$(basename "$(readlink -f "/sys/bus/pci/devices/$gpu/.." 2>/dev/null)" 2>/dev/null)
+  if [ -n "$bridge" ] && [ "$bridge" != "$gpu" ]; then
+    echo
+    echo "=== UPSTREAM $bridge for $gpu ==="
+    lspci -s "$bridge" -vv 2>&1 || true
+  fi
+  echo
+done
+if [ "$found" -eq 0 ]; then
+  echo "no NVIDIA PCI devices found"
+fi
+`}},
 	{name: "system/pcie-nvidia-link.txt", cmd: []string{"sh", "-c", `
 for d in /sys/bus/pci/devices/*/; do
   vendor=$(cat "$d/vendor" 2>/dev/null)
-  [ "$vendor" = "0x10de" ] || continue
-  dev=$(basename "$d")
+	  [ "$vendor" = "0x10de" ] || continue
+	  class=$(cat "$d/class" 2>/dev/null)
+	  case "$class" in
+	    0x030000|0x030200) ;;
+	    *) continue ;;
+	  esac
+	  dev=$(basename "$d")
   echo "=== $dev ==="
   for f in current_link_speed current_link_width max_link_speed max_link_width; do
     printf "  %-22s %s\n" "$f" "$(cat "$d/$f" 2>/dev/null)"
   done
 done
+`}},
+	{name: "system/pcie-aer-sysfs.txt", cmd: []string{"sh", "-c", `
+found=0
+for dev in /sys/bus/pci/devices/*; do
+  [ -e "$dev" ] || continue
+  bdf=$(basename "$dev")
+  block=""
+  for f in aer_dev_correctable aer_dev_fatal aer_dev_nonfatal aer_rootport_total_err_cor aer_rootport_total_err_fatal aer_rootport_total_err_nonfatal; do
+    if [ -r "$dev/$f" ]; then
+      if [ -z "$block" ]; then
+        block=1
+        found=1
+        echo "=== $bdf ==="
+      fi
+      printf "  %-30s %s\n" "$f" "$(cat "$dev/$f" 2>/dev/null)"
+    fi
+  done
+  if [ -n "$block" ]; then
+    echo
+  fi
+done
+if [ "$found" -eq 0 ]; then
+  echo "no PCIe AER sysfs counters found"
+fi
 `}},
 	{name: "system/ethtool-info.txt", cmd: []string{"sh", "-c", `
 if ! command -v ethtool >/dev/null 2>&1; then
@@ -139,7 +197,7 @@ var supportBundleOptionalFiles = []struct {
 	{name: "system/syslog.txt", src: "/var/log/syslog"},
 }
 
-const supportBundleGlob = "bee-support-*.tar.gz"
+const supportBundleGlob = "????-??-?? (BEE-SP*)*.tar.gz"
 
 func BuildSupportBundle(exportDir string) (string, error) {
 	exportDir = strings.TrimSpace(exportDir)
@@ -153,9 +211,14 @@ func BuildSupportBundle(exportDir string) (string, error) {
 		return "", err
 	}
 
-	host := sanitizeFilename(hostnameOr("unknown"))
-	ts := time.Now().UTC().Format("20060102-150405")
-	stageRoot := filepath.Join(os.TempDir(), fmt.Sprintf("bee-support-%s-%s", host, ts))
+	now := time.Now().UTC()
+	date := now.Format("2006-01-02")
+	tod := now.Format("150405")
+	ver := bundleVersion()
+	model := serverModelForBundle()
+	sn := serverSerialForBundle()
+
+	stageRoot := filepath.Join(os.TempDir(), fmt.Sprintf("bee-support-stage-%s-%s", sanitizeFilename(hostnameOr("unknown")), now.Format("20060102-150405")))
 	if err := os.MkdirAll(stageRoot, 0755); err != nil {
 		return "", err
 	}
@@ -187,7 +250,8 @@ func BuildSupportBundle(exportDir string) (string, error) {
 		return "", err
 	}
 
-	archivePath := filepath.Join(os.TempDir(), fmt.Sprintf("bee-support-%s-%s.tar.gz", host, ts))
+	archiveName := fmt.Sprintf("%s (BEE-SP v%s) %s %s %s.tar.gz", date, ver, model, sn, tod)
+	archivePath := filepath.Join(os.TempDir(), archiveName)
 	if err := createSupportTarGz(archivePath, stageRoot); err != nil {
 		return "", err
 	}
@@ -344,6 +408,60 @@ func writeManifest(dst, exportDir, stageRoot string) error {
 	return os.WriteFile(dst, []byte(body.String()), 0644)
 }
 
+func bundleVersion() string {
+	v := buildVersion()
+	v = strings.TrimPrefix(v, "v")
+	v = strings.TrimPrefix(v, "V")
+	if v == "" || v == "unknown" {
+		return "0.0"
+	}
+	return v
+}
+
+func serverModelForBundle() string {
+	raw, err := exec.Command("dmidecode", "-t", "1").Output()
+	if err != nil {
+		return "unknown"
+	}
+	for _, line := range strings.Split(string(raw), "\n") {
+		line = strings.TrimSpace(line)
+		key, val, ok := strings.Cut(line, ": ")
+		if !ok {
+			continue
+		}
+		if strings.TrimSpace(key) == "Product Name" {
+			val = strings.TrimSpace(val)
+			if val == "" {
+				return "unknown"
+			}
+			return strings.ReplaceAll(val, " ", "_")
+		}
+	}
+	return "unknown"
+}
+
+func serverSerialForBundle() string {
+	raw, err := exec.Command("dmidecode", "-t", "1").Output()
+	if err != nil {
+		return "unknown"
+	}
+	for _, line := range strings.Split(string(raw), "\n") {
+		line = strings.TrimSpace(line)
+		key, val, ok := strings.Cut(line, ": ")
+		if !ok {
+			continue
+		}
+		if strings.TrimSpace(key) == "Serial Number" {
+			val = strings.TrimSpace(val)
+			if val == "" {
+				return "unknown"
+			}
+			return val
+		}
+	}
+	return "unknown"
+}
+
 func buildVersion() string {
 	raw, err := exec.Command("bee", "version").CombinedOutput()
 	if err != nil {

audit/internal/collector/nic_mellanox.go

@@ -179,11 +179,3 @@ func commandOutputWithTimeout(timeout time.Duration, name string, args ...string
 	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

@@ -58,12 +58,10 @@ func enrichPCIeWithNICTelemetry(devs []schema.HardwarePCIeDevice) []schema.Hardw
 			}
 		}
 
-		if interfaceHasCarrier(iface) {
-			if out, err := ethtoolModuleQuery(iface); err == nil {
-				if injectSFPDOMTelemetry(&devs[i], out) {
-					enriched++
-					continue
-				}
+		if out, err := ethtoolModuleQuery(iface); err == nil {
+			if injectSFPDOMTelemetry(&devs[i], out) {
+				enriched++
+				continue
 			}
 		}
 		if len(devs[i].MacAddresses) > 0 || devs[i].Firmware != nil {
@@ -115,8 +113,38 @@ func injectSFPDOMTelemetry(dev *schema.HardwarePCIeDevice, raw string) bool {
 		}
 		key := strings.ToLower(strings.TrimSpace(trimmed[:idx]))
 		val := strings.TrimSpace(trimmed[idx+1:])
+		if val == "" || strings.EqualFold(val, "not supported") || strings.EqualFold(val, "unknown") {
+			continue
+		}
 
 		switch {
+		case key == "identifier":
+			s := parseSFPIdentifier(val)
+			dev.SFPIdentifier = &s
+			t := true
+			dev.SFPPresent = &t
+			changed = true
+		case key == "connector":
+			s := parseSFPConnector(val)
+			dev.SFPConnector = &s
+			changed = true
+		case key == "vendor name":
+			s := strings.TrimSpace(val)
+			dev.SFPVendor = &s
+			changed = true
+		case key == "vendor pn":
+			s := strings.TrimSpace(val)
+			dev.SFPPartNumber = &s
+			changed = true
+		case key == "vendor sn":
+			s := strings.TrimSpace(val)
+			dev.SFPSerialNumber = &s
+			changed = true
+		case strings.Contains(key, "laser wavelength"):
+			if f, ok := firstFloat(val); ok {
+				dev.SFPWavelengthNM = &f
+				changed = true
+			}
 		case strings.Contains(key, "module temperature"):
 			if f, ok := firstFloat(val); ok {
 				dev.SFPTemperatureC = &f
@@ -147,12 +175,61 @@ func injectSFPDOMTelemetry(dev *schema.HardwarePCIeDevice, raw string) bool {
 	return changed
 }
 
+// parseSFPIdentifier extracts the human-readable transceiver type from the
+// raw ethtool identifier line, e.g. "0x03 (SFP)" → "SFP".
+func parseSFPIdentifier(val string) string {
+	if s := extractParens(val); s != "" {
+		return s
+	}
+	return val
+}
+
+// parseSFPConnector extracts the connector type from the raw ethtool line,
+// e.g. "0x07 (LC)" → "LC".
+func parseSFPConnector(val string) string {
+	if s := extractParens(val); s != "" {
+		return s
+	}
+	return val
+}
+
+var parenRe = regexp.MustCompile(`\(([^)]+)\)`)
+
+func extractParens(s string) string {
+	m := parenRe.FindStringSubmatch(s)
+	if len(m) < 2 {
+		return ""
+	}
+	return strings.TrimSpace(m[1])
+}
+
 func parseSFPDOM(raw string) map[string]any {
 	dev := schema.HardwarePCIeDevice{}
 	if !injectSFPDOMTelemetry(&dev, raw) {
 		return map[string]any{}
 	}
 	out := map[string]any{}
+	if dev.SFPPresent != nil {
+		out["sfp_present"] = *dev.SFPPresent
+	}
+	if dev.SFPIdentifier != nil {
+		out["sfp_identifier"] = *dev.SFPIdentifier
+	}
+	if dev.SFPConnector != nil {
+		out["sfp_connector"] = *dev.SFPConnector
+	}
+	if dev.SFPVendor != nil {
+		out["sfp_vendor"] = *dev.SFPVendor
+	}
+	if dev.SFPPartNumber != nil {
+		out["sfp_part_number"] = *dev.SFPPartNumber
+	}
+	if dev.SFPSerialNumber != nil {
+		out["sfp_serial_number"] = *dev.SFPSerialNumber
+	}
+	if dev.SFPWavelengthNM != nil {
+		out["sfp_wavelength_nm"] = *dev.SFPWavelengthNM
+	}
 	if dev.SFPTemperatureC != nil {
 		out["sfp_temperature_c"] = *dev.SFPTemperatureC
 	}

audit/internal/collector/nic_telemetry_test.go

@@ -122,10 +122,7 @@ func TestEnrichPCIeWithNICTelemetrySkipsModuleQueryWithoutCarrier(t *testing.T)
 	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
-	}
+	ethtoolModuleQuery = func(string) (string, error) { return "", fmt.Errorf("no module") }
 
 	class := "EthernetController"
 	bdf := "0000:18:00.0"

audit/internal/collector/nvidia.go

@@ -13,7 +13,9 @@ import (
 const nvidiaVendorID = 0x10de
 
 type nvidiaGPUInfo struct {
+	Index              int
 	BDF                string
+	Name               string
 	Serial             string
 	VBIOS              string
 	TemperatureC       *float64
@@ -72,6 +74,9 @@ func enrichPCIeWithNVIDIAData(devs []schema.HardwarePCIeDevice, gpuByBDF map[str
 			continue
 		}
 
+		if v := strings.TrimSpace(info.Name); v != "" {
+			devs[i].Model = &v
+		}
 		if v := strings.TrimSpace(info.Serial); v != "" {
 			devs[i].SerialNumber = &v
 		}
@@ -98,7 +103,7 @@ 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,pcie.link.gen.current,pcie.link.gen.max,pcie.link.width.current,pcie.link.width.max",
+		"--query-gpu=index,pci.bus_id,name,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 {
@@ -122,8 +127,8 @@ func parseNVIDIASMIQuery(raw string) (map[string]nvidiaGPUInfo, error) {
 		if len(rec) == 0 {
 			continue
 		}
-		if len(rec) < 13 {
-			return nil, fmt.Errorf("unexpected nvidia-smi columns: got %d, want 13", len(rec))
+		if len(rec) < 14 {
+			return nil, fmt.Errorf("unexpected nvidia-smi columns: got %d, want 14", len(rec))
 		}
 
 		bdf := normalizePCIeBDF(rec[1])
@@ -132,18 +137,20 @@ func parseNVIDIASMIQuery(raw string) (map[string]nvidiaGPUInfo, error) {
 		}
 
 		info := nvidiaGPUInfo{
+			Index:              parseRequiredInt(rec[0]),
 			BDF:                bdf,
-			Serial:             strings.TrimSpace(rec[2]),
-			VBIOS:              strings.TrimSpace(rec[3]),
-			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]),
+			Name:               strings.TrimSpace(rec[2]),
+			Serial:             strings.TrimSpace(rec[3]),
+			VBIOS:              strings.TrimSpace(rec[4]),
+			TemperatureC:       parseMaybeFloat(rec[5]),
+			PowerW:             parseMaybeFloat(rec[6]),
+			ECCUncorrected:     parseMaybeInt64(rec[7]),
+			ECCCorrected:       parseMaybeInt64(rec[8]),
+			HWSlowdown:         parseMaybeBool(rec[9]),
+			PCIeLinkGenCurrent: parseMaybeInt(rec[10]),
+			PCIeLinkGenMax:     parseMaybeInt(rec[11]),
+			PCIeLinkWidthCur:   parseMaybeInt(rec[12]),
+			PCIeLinkWidthMax:   parseMaybeInt(rec[13]),
 		}
 		result[bdf] = info
 	}
@@ -187,6 +194,14 @@ func parseMaybeInt(v string) *int {
 	return &n
 }
 
+func parseRequiredInt(v string) int {
+	n, err := strconv.Atoi(strings.TrimSpace(v))
+	if err != nil {
+		return 0
+	}
+	return n
+}
+
 func pcieLinkGenLabel(gen int) string {
 	return fmt.Sprintf("Gen%d", gen)
 }
@@ -240,6 +255,10 @@ func setPCIeFallback(dev *schema.HardwarePCIeDevice) {
 }
 
 func injectNVIDIATelemetry(dev *schema.HardwarePCIeDevice, info nvidiaGPUInfo) {
+	if dev.Telemetry == nil {
+		dev.Telemetry = map[string]any{}
+	}
+	dev.Telemetry["nvidia_gpu_index"] = info.Index
 	if info.TemperatureC != nil {
 		dev.TemperatureC = info.TemperatureC
 	}

audit/internal/collector/nvidia_test.go

@@ -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, 4, 4, 16, 16\n"
+	raw := "0, 00000000:65:00.0, NVIDIA H100 80GB HBM3, 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)
@@ -16,6 +16,9 @@ func TestParseNVIDIASMIQuery(t *testing.T) {
 	if !ok {
 		t.Fatalf("gpu by normalized bdf not found")
 	}
+	if gpu.Name != "NVIDIA H100 80GB HBM3" {
+		t.Fatalf("name: got %q", gpu.Name)
+	}
 	if gpu.Serial != "GPU-SERIAL-1" {
 		t.Fatalf("serial: got %q", gpu.Serial)
 	}
@@ -86,6 +89,9 @@ func TestEnrichPCIeWithNVIDIAData_driverLoaded(t *testing.T) {
 	if out[0].Firmware == nil || *out[0].Firmware != "96.00.1F.00.02" {
 		t.Fatalf("firmware: got %v", out[0].Firmware)
 	}
+	if out[0].Telemetry == nil || out[0].Telemetry["nvidia_gpu_index"] != 0 {
+		t.Fatalf("telemetry nvidia_gpu_index: got %#v", out[0].Telemetry)
+	}
 	if out[0].Status == nil || *out[0].Status != statusWarning {
 		t.Fatalf("status: got %v", out[0].Status)
 	}

audit/internal/collector/pcie.go

@@ -2,6 +2,7 @@ package collector
 
 import (
 	"bee/audit/internal/schema"
+	"fmt"
 	"log/slog"
 	"os/exec"
 	"strconv"
@@ -79,6 +80,25 @@ func shouldIncludePCIeDevice(class, vendor, device string) bool {
 		}
 	}
 
+	// Exclude BMC/management virtual VGA adapters — these are firmware video chips,
+	// not real GPUs, and pollute the GPU inventory (e.g. iBMC, iDRAC, iLO VGA).
+	if strings.Contains(c, "vga") || strings.Contains(c, "display") || strings.Contains(c, "3d") {
+		bmcPatterns := []string{
+			"management system chip",
+			"management controller",
+			"ibmc",
+			"idrac",
+			"ilo vga",
+			"aspeed",
+			"matrox",
+		}
+		for _, bad := range bmcPatterns {
+			if strings.Contains(d, bad) {
+				return false
+			}
+		}
+	}
+
 	if strings.Contains(v, "advanced micro devices") || strings.Contains(v, "[amd]") {
 		internalAMDPatterns := []string{
 			"dummy function",
@@ -153,6 +173,9 @@ func parseLspciDevice(fields map[string]string) schema.HardwarePCIeDevice {
 
 	// SVendor/SDevice available but not in schema — skip
 
+	// Warn if PCIe link is running below its maximum negotiated speed.
+	applyPCIeLinkSpeedWarning(&dev)
+
 	return dev
 }
 
@@ -222,6 +245,41 @@ func readPCIStringAttribute(bdf, attribute string) (string, bool) {
 	return value, true
 }
 
+// applyPCIeLinkSpeedWarning sets the device status to Warning if the current PCIe link
+// speed is below the maximum negotiated speed supported by both ends.
+func applyPCIeLinkSpeedWarning(dev *schema.HardwarePCIeDevice) {
+	if dev.LinkSpeed == nil || dev.MaxLinkSpeed == nil {
+		return
+	}
+	if pcieLinkSpeedRank(*dev.LinkSpeed) < pcieLinkSpeedRank(*dev.MaxLinkSpeed) {
+		warn := statusWarning
+		dev.Status = &warn
+		desc := fmt.Sprintf("PCIe link speed degraded: running at %s, capable of %s", *dev.LinkSpeed, *dev.MaxLinkSpeed)
+		dev.ErrorDescription = &desc
+	}
+}
+
+// pcieLinkSpeedRank returns a numeric rank for a normalized Gen string (e.g. "Gen4" → 4).
+// Returns 0 for unrecognised values so comparisons fail safe.
+func pcieLinkSpeedRank(gen string) int {
+	switch gen {
+	case "Gen1":
+		return 1
+	case "Gen2":
+		return 2
+	case "Gen3":
+		return 3
+	case "Gen4":
+		return 4
+	case "Gen5":
+		return 5
+	case "Gen6":
+		return 6
+	default:
+		return 0
+	}
+}
+
 func normalizePCILinkSpeed(raw string) string {
 	raw = strings.TrimSpace(strings.ToLower(raw))
 	switch {

audit/internal/collector/pcie_filter_test.go

@@ -1,6 +1,7 @@
 package collector
 
 import (
+	"bee/audit/internal/schema"
 	"encoding/json"
 	"strings"
 	"testing"
@@ -29,6 +30,8 @@ func TestShouldIncludePCIeDevice(t *testing.T) {
 		{name: "raid", class: "RAID bus controller", want: true},
 		{name: "nvme", class: "Non-Volatile memory controller", want: true},
 		{name: "vga", class: "VGA compatible controller", want: true},
+		{name: "ibmc vga", class: "VGA compatible controller", vendor: "Huawei Technologies Co., Ltd.", device: "Hi171x Series [iBMC Intelligent Management system chip w/VGA support]", want: false},
+		{name: "aspeed vga", class: "VGA compatible controller", vendor: "ASPEED Technology, Inc.", device: "ASPEED Graphics Family", want: false},
 		{name: "other encryption controller", class: "Encryption controller", vendor: "Intel Corporation", device: "QuickAssist", want: true},
 	}
 
@@ -139,3 +142,77 @@ func TestNormalizePCILinkSpeed(t *testing.T) {
 		}
 	}
 }
+
+func TestApplyPCIeLinkSpeedWarning(t *testing.T) {
+	ptr := func(s string) *string { return &s }
+
+	tests := []struct {
+		name        string
+		linkSpeed   *string
+		maxSpeed    *string
+		wantWarning bool
+		wantGenIn   string // substring expected in ErrorDescription when warning
+	}{
+		{
+			name:        "degraded Gen1 vs Gen5",
+			linkSpeed:   ptr("Gen1"),
+			maxSpeed:    ptr("Gen5"),
+			wantWarning: true,
+			wantGenIn:   "Gen1",
+		},
+		{
+			name:        "at max Gen5",
+			linkSpeed:   ptr("Gen5"),
+			maxSpeed:    ptr("Gen5"),
+			wantWarning: false,
+		},
+		{
+			name:        "degraded Gen4 vs Gen5",
+			linkSpeed:   ptr("Gen4"),
+			maxSpeed:    ptr("Gen5"),
+			wantWarning: true,
+			wantGenIn:   "Gen4",
+		},
+		{
+			name:        "missing current speed — no warning",
+			linkSpeed:   nil,
+			maxSpeed:    ptr("Gen5"),
+			wantWarning: false,
+		},
+		{
+			name:        "missing max speed — no warning",
+			linkSpeed:   ptr("Gen1"),
+			maxSpeed:    nil,
+			wantWarning: false,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			dev := schema.HardwarePCIeDevice{}
+			ok := statusOK
+			dev.Status = &ok
+			dev.LinkSpeed = tt.linkSpeed
+			dev.MaxLinkSpeed = tt.maxSpeed
+
+			applyPCIeLinkSpeedWarning(&dev)
+
+			gotWarn := dev.Status != nil && *dev.Status == statusWarning
+			if gotWarn != tt.wantWarning {
+				t.Fatalf("wantWarning=%v gotWarning=%v (status=%v)", tt.wantWarning, gotWarn, dev.Status)
+			}
+			if tt.wantWarning {
+				if dev.ErrorDescription == nil {
+					t.Fatal("expected ErrorDescription to be set")
+				}
+				if !strings.Contains(*dev.ErrorDescription, tt.wantGenIn) {
+					t.Fatalf("ErrorDescription %q does not contain %q", *dev.ErrorDescription, tt.wantGenIn)
+				}
+			} else {
+				if dev.ErrorDescription != nil {
+					t.Fatalf("unexpected ErrorDescription: %s", *dev.ErrorDescription)
+				}
+			}
+		})
+	}
+}

audit/internal/platform/benchmark.go

@@ -121,15 +121,22 @@ func (s *System) RunNvidiaBenchmark(ctx context.Context, baseDir string, opts Nv
 	var serverIdleOK, serverLoadedOK bool
 	var serverLoadedSamples int
 
+	// Run nvidia-smi -q first: used both for the log file and as a fallback
+	// source of max clock values when CSV clock fields are unsupported.
+	var nvsmiQOut []byte
+	if out, err := runSATCommandCtx(ctx, verboseLog, "00-nvidia-smi-q.log", []string{"nvidia-smi", "-q"}, nil, nil); err == nil {
+		nvsmiQOut = out
+		_ = os.WriteFile(filepath.Join(runDir, "00-nvidia-smi-q.log"), out, 0644)
+	}
+
 	infoByIndex, infoErr := queryBenchmarkGPUInfo(selected)
 	if infoErr != nil {
 		result.Warnings = append(result.Warnings, "gpu inventory query failed: "+infoErr.Error())
 		result.Normalization.Status = "partial"
 	}
-
-	if out, err := runSATCommandCtx(ctx, verboseLog, "00-nvidia-smi-q.log", []string{"nvidia-smi", "-q"}, nil, nil); err == nil {
-		_ = os.WriteFile(filepath.Join(runDir, "00-nvidia-smi-q.log"), out, 0644)
-	}
+	// Enrich with max clocks from verbose output — covers GPUs where
+	// clocks.max.* CSV fields are unsupported (e.g. Blackwell / driver 98.x).
+	enrichGPUInfoWithMaxClocks(infoByIndex, nvsmiQOut)
 
 	activeApps, err := queryActiveComputeApps(selected)
 	if err == nil && len(activeApps) > 0 {
@@ -326,8 +333,8 @@ func (s *System) RunNvidiaBenchmark(ctx context.Context, baseDir string, opts Nv
 	}
 
 	report := renderBenchmarkReportWithCharts(result, loadBenchmarkReportCharts(runDir, selected))
-	if err := os.WriteFile(filepath.Join(runDir, "report.txt"), []byte(report), 0644); err != nil {
-		return "", fmt.Errorf("write report.txt: %w", err)
+	if err := os.WriteFile(filepath.Join(runDir, "report.md"), []byte(report), 0644); err != nil {
+		return "", fmt.Errorf("write report.md: %w", err)
 	}
 
 	summary := renderBenchmarkSummary(result)
@@ -335,11 +342,7 @@ func (s *System) RunNvidiaBenchmark(ctx context.Context, baseDir string, opts Nv
 		return "", fmt.Errorf("write summary.txt: %w", err)
 	}
 
-	archive := filepath.Join(baseDir, "gpu-benchmark-"+ts+".tar.gz")
-	if err := createTarGz(archive, runDir); err != nil {
-		return "", fmt.Errorf("pack benchmark archive: %w", err)
-	}
-	return archive, nil
+	return runDir, nil
 }
 
 func normalizeNvidiaBenchmarkOptionsForBenchmark(opts NvidiaBenchmarkOptions) NvidiaBenchmarkOptions {
@@ -374,9 +377,13 @@ func resolveBenchmarkProfile(profile string) benchmarkProfileSpec {
 // Fields are tried in order; the first successful query wins. Extended fields
 // (attribute.multiprocessor_count, power.default_limit) are not supported on
 // all driver versions, so we fall back to the base set if the full query fails.
+// The minimal fallback omits clock fields entirely — clocks.max.* returns
+// exit status 2 on some GPU generations (e.g. Blackwell); max clocks are
+// then recovered from nvidia-smi -q via enrichGPUInfoWithMaxClocks.
 var benchmarkGPUInfoQueries = []struct {
 	fields   string
 	extended bool // whether this query includes optional extended fields
+	minimal  bool // clock fields omitted; max clocks must be filled separately
 }{
 	{
 		fields:   "index,uuid,name,pci.bus_id,vbios_version,power.limit,clocks.max.graphics,clocks.max.memory,clocks.base.graphics,attribute.multiprocessor_count,power.default_limit",
@@ -386,6 +393,83 @@ var benchmarkGPUInfoQueries = []struct {
 		fields:   "index,uuid,name,pci.bus_id,vbios_version,power.limit,clocks.max.graphics,clocks.max.memory,clocks.base.graphics",
 		extended: false,
 	},
+	{
+		fields:  "index,uuid,name,pci.bus_id,vbios_version,power.limit",
+		minimal: true,
+	},
+}
+
+// enrichGPUInfoWithMaxClocks fills MaxGraphicsClockMHz / MaxMemoryClockMHz for
+// any GPU in infoByIndex where those values are still zero.  It parses the
+// "Max Clocks" section of nvidia-smi -q output (already available as nvsmiQ).
+// This is the fallback for GPUs (e.g. Blackwell) where clocks.max.* CSV fields
+// return exit status 2 but the verbose query works fine.
+func enrichGPUInfoWithMaxClocks(infoByIndex map[int]benchmarkGPUInfo, nvsmiQ []byte) {
+	if len(infoByIndex) == 0 || len(nvsmiQ) == 0 {
+		return
+	}
+
+	// Build bus_id → index map for matching verbose sections to GPU indices.
+	busToBenchIdx := make(map[string]int, len(infoByIndex))
+	for idx, info := range infoByIndex {
+		if info.BusID != "" {
+			// nvidia-smi -q uses "GPU 00000000:4E:00.0" (8-digit domain),
+			// while --query-gpu returns the same format; normalise to lower.
+			busToBenchIdx[strings.ToLower(strings.TrimSpace(info.BusID))] = idx
+		}
+	}
+
+	// Split the verbose output into per-GPU sections on "^GPU " lines.
+	gpuSectionRe := regexp.MustCompile(`(?m)^GPU\s+([\dA-Fa-f:\.]+)`)
+	maxGfxRe := regexp.MustCompile(`(?i)Max Clocks[\s\S]*?Graphics\s*:\s*(\d+)\s*MHz`)
+	maxMemRe := regexp.MustCompile(`(?i)Max Clocks[\s\S]*?Memory\s*:\s*(\d+)\s*MHz`)
+
+	sectionStarts := gpuSectionRe.FindAllSubmatchIndex(nvsmiQ, -1)
+	for i, loc := range sectionStarts {
+		busID := strings.ToLower(string(nvsmiQ[loc[2]:loc[3]]))
+		benchIdx, ok := busToBenchIdx[busID]
+		if !ok {
+			// Bus IDs from verbose output may have a different domain prefix;
+			// try suffix match on the slot portion (XX:XX.X).
+			for k, v := range busToBenchIdx {
+				if strings.HasSuffix(k, busID) || strings.HasSuffix(busID, k) {
+					benchIdx = v
+					ok = true
+					break
+				}
+			}
+		}
+		if !ok {
+			continue
+		}
+
+		info := infoByIndex[benchIdx]
+		if info.MaxGraphicsClockMHz > 0 && info.MaxMemoryClockMHz > 0 {
+			continue // already populated
+		}
+
+		end := len(nvsmiQ)
+		if i+1 < len(sectionStarts) {
+			end = sectionStarts[i+1][0]
+		}
+		section := nvsmiQ[loc[0]:end]
+
+		if info.MaxGraphicsClockMHz == 0 {
+			if m := maxGfxRe.FindSubmatch(section); m != nil {
+				if v, err := strconv.ParseFloat(string(m[1]), 64); err == nil {
+					info.MaxGraphicsClockMHz = v
+				}
+			}
+		}
+		if info.MaxMemoryClockMHz == 0 {
+			if m := maxMemRe.FindSubmatch(section); m != nil {
+				if v, err := strconv.ParseFloat(string(m[1]), 64); err == nil {
+					info.MaxMemoryClockMHz = v
+				}
+			}
+		}
+		infoByIndex[benchIdx] = info
+	}
 }
 
 func queryBenchmarkGPUInfo(gpuIndices []int) (map[int]benchmarkGPUInfo, error) {
@@ -413,9 +497,13 @@ func queryBenchmarkGPUInfo(gpuIndices []int) (map[int]benchmarkGPUInfo, error) {
 			continue
 		}
 
+		minFields := 6
+		if !q.minimal {
+			minFields = 9
+		}
 		infoByIndex := make(map[int]benchmarkGPUInfo, len(rows))
 		for _, row := range rows {
-			if len(row) < 9 {
+			if len(row) < minFields {
 				continue
 			}
 			idx, err := strconv.Atoi(strings.TrimSpace(row[0]))
@@ -423,24 +511,26 @@ func queryBenchmarkGPUInfo(gpuIndices []int) (map[int]benchmarkGPUInfo, error) {
 				continue
 			}
 			info := benchmarkGPUInfo{
-				Index:               idx,
-				UUID:                strings.TrimSpace(row[1]),
-				Name:                strings.TrimSpace(row[2]),
-				BusID:               strings.TrimSpace(row[3]),
-				VBIOS:               strings.TrimSpace(row[4]),
-				PowerLimitW:         parseBenchmarkFloat(row[5]),
-				MaxGraphicsClockMHz: parseBenchmarkFloat(row[6]),
-				MaxMemoryClockMHz:   parseBenchmarkFloat(row[7]),
+				Index:       idx,
+				UUID:        strings.TrimSpace(row[1]),
+				Name:        strings.TrimSpace(row[2]),
+				BusID:       strings.TrimSpace(row[3]),
+				VBIOS:       strings.TrimSpace(row[4]),
+				PowerLimitW: parseBenchmarkFloat(row[5]),
 			}
-			if len(row) >= 9 {
-				info.BaseGraphicsClockMHz = parseBenchmarkFloat(row[8])
-			}
-			if q.extended {
-				if len(row) >= 10 {
-					info.MultiprocessorCount = int(parseBenchmarkFloat(row[9]))
+			if !q.minimal {
+				info.MaxGraphicsClockMHz = parseBenchmarkFloat(row[6])
+				info.MaxMemoryClockMHz = parseBenchmarkFloat(row[7])
+				if len(row) >= 9 {
+					info.BaseGraphicsClockMHz = parseBenchmarkFloat(row[8])
 				}
-				if len(row) >= 11 {
-					info.DefaultPowerLimitW = parseBenchmarkFloat(row[10])
+				if q.extended {
+					if len(row) >= 10 {
+						info.MultiprocessorCount = int(parseBenchmarkFloat(row[9]))
+					}
+					if len(row) >= 11 {
+						info.DefaultPowerLimitW = parseBenchmarkFloat(row[10])
+					}
 				}
 			}
 			infoByIndex[idx] = info
@@ -1183,18 +1273,8 @@ func queryIPMIServerPowerW() (float64, error) {
 	if err != nil {
 		return 0, fmt.Errorf("ipmitool dcmi power reading: %w", err)
 	}
-	for _, line := range strings.Split(string(out), "\n") {
-		if strings.Contains(line, "Current Power") {
-			parts := strings.SplitN(line, ":", 2)
-			if len(parts) == 2 {
-				val := strings.TrimSpace(strings.TrimSuffix(strings.TrimSpace(parts[1]), "Watts"))
-				val = strings.TrimSpace(val)
-				w, err := strconv.ParseFloat(val, 64)
-				if err == nil && w > 0 {
-					return w, nil
-				}
-			}
-		}
+	if w := parseDCMIPowerReading(string(out)); w > 0 {
+		return w, nil
 	}
 	return 0, fmt.Errorf("could not parse ipmitool dcmi power reading output")
 }

audit/internal/platform/benchmark_report.go

@@ -22,18 +22,53 @@ var ansiEscapePattern = regexp.MustCompile(`\x1b\[[0-9;]*m`)
 
 func renderBenchmarkReportWithCharts(result NvidiaBenchmarkResult, charts []benchmarkReportChart) 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)
 
+	// ── Header ────────────────────────────────────────────────────────────────
+	b.WriteString("# Bee NVIDIA Benchmark Report\n\n")
+
+	// System identity block
+	if result.ServerModel != "" {
+		fmt.Fprintf(&b, "**Server:** %s  \n", result.ServerModel)
+	}
+	if result.Hostname != "" {
+		fmt.Fprintf(&b, "**Host:** %s  \n", result.Hostname)
+	}
+	// GPU models summary
+	if len(result.GPUs) > 0 {
+		modelCount := make(map[string]int)
+		var modelOrder []string
+		for _, g := range result.GPUs {
+			m := strings.TrimSpace(g.Name)
+			if m == "" {
+				m = "Unknown GPU"
+			}
+			if modelCount[m] == 0 {
+				modelOrder = append(modelOrder, m)
+			}
+			modelCount[m]++
+		}
+		var parts []string
+		for _, m := range modelOrder {
+			if modelCount[m] == 1 {
+				parts = append(parts, m)
+			} else {
+				parts = append(parts, fmt.Sprintf("%d× %s", modelCount[m], m))
+			}
+		}
+		fmt.Fprintf(&b, "**GPU(s):** %s  \n", strings.Join(parts, ", "))
+	}
+	fmt.Fprintf(&b, "**Profile:** %s  \n", result.BenchmarkProfile)
+	fmt.Fprintf(&b, "**App version:** %s  \n", result.BenchmarkVersion)
+	fmt.Fprintf(&b, "**Generated:** %s  \n", result.GeneratedAt.Format("2006-01-02 15:04:05 UTC"))
+	if result.ParallelGPUs {
+		fmt.Fprintf(&b, "**Mode:** parallel (all GPUs simultaneously)  \n")
+	}
+	fmt.Fprintf(&b, "**Overall status:** %s  \n", result.OverallStatus)
+	b.WriteString("\n")
+
+	// ── Executive Summary ─────────────────────────────────────────────────────
 	if len(result.Findings) > 0 {
-		fmt.Fprintf(&b, "Executive Summary\n")
-		fmt.Fprintf(&b, "-----------------\n")
+		b.WriteString("## Executive Summary\n\n")
 		for _, finding := range result.Findings {
 			fmt.Fprintf(&b, "- %s\n", finding)
 		}
@@ -41,149 +76,206 @@ func renderBenchmarkReportWithCharts(result NvidiaBenchmarkResult, charts []benc
 	}
 
 	if len(result.Warnings) > 0 {
-		fmt.Fprintf(&b, "Warnings\n")
-		fmt.Fprintf(&b, "--------\n")
+		b.WriteString("## Warnings\n\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")
+	// ── Scorecard table ───────────────────────────────────────────────────────
+	b.WriteString("## Scorecard\n\n")
+	b.WriteString("| GPU | Status | Composite | Compute | TOPS/SM/GHz | Power Sustain | Thermal Sustain | Stability | Interconnect |\n")
+	b.WriteString("|-----|--------|-----------|---------|-------------|---------------|-----------------|-----------|-------------|\n")
 	for _, gpu := range result.GPUs {
-		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)
-		if gpu.Scores.TOPSPerSMPerGHz > 0 {
-			fmt.Fprintf(&b, "  Compute efficiency: %.3f TOPS/SM/GHz\n", gpu.Scores.TOPSPerSMPerGHz)
+		name := strings.TrimSpace(gpu.Name)
+		if name == "" {
+			name = "Unknown GPU"
 		}
-		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)
+		interconnect := "-"
 		if gpu.Scores.InterconnectScore > 0 {
-			fmt.Fprintf(&b, "  Interconnect: %.1f\n", gpu.Scores.InterconnectScore)
+			interconnect = fmt.Sprintf("%.1f", gpu.Scores.InterconnectScore)
 		}
-		if len(gpu.DegradationReasons) > 0 {
-			fmt.Fprintf(&b, "  Degradation reasons: %s\n", strings.Join(gpu.DegradationReasons, ", "))
+		topsPerSM := "-"
+		if gpu.Scores.TOPSPerSMPerGHz > 0 {
+			topsPerSM = fmt.Sprintf("%.3f", gpu.Scores.TOPSPerSMPerGHz)
 		}
-		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, "| GPU %d %s | %s | **%.2f** | %.2f | %s | %.1f | %.1f | %.1f | %s |\n",
+			gpu.Index, name,
+			gpu.Status,
+			gpu.Scores.CompositeScore,
+			gpu.Scores.ComputeScore,
+			topsPerSM,
+			gpu.Scores.PowerSustainScore,
+			gpu.Scores.ThermalSustainScore,
+			gpu.Scores.StabilityScore,
+			interconnect,
+		)
+	}
+	b.WriteString("\n")
+
+	// ── Per GPU detail ────────────────────────────────────────────────────────
+	b.WriteString("## Per-GPU Details\n\n")
+	for _, gpu := range result.GPUs {
+		name := strings.TrimSpace(gpu.Name)
+		if name == "" {
+			name = "Unknown GPU"
 		}
-		fmt.Fprintf(&b, "  Throttle: %s\n", formatThrottleLine(gpu.Throttle, gpu.Steady.DurationSec))
-		if len(gpu.Notes) > 0 {
-			fmt.Fprintf(&b, "  Notes:\n")
-			for _, note := range gpu.Notes {
-				fmt.Fprintf(&b, "    - %s\n", note)
-			}
+		fmt.Fprintf(&b, "### GPU %d — %s\n\n", gpu.Index, name)
+
+		// Identity
+		if gpu.BusID != "" {
+			fmt.Fprintf(&b, "- **Bus ID:** %s\n", gpu.BusID)
+		}
+		if gpu.VBIOS != "" {
+			fmt.Fprintf(&b, "- **vBIOS:** %s\n", gpu.VBIOS)
+		}
+		if gpu.ComputeCapability != "" {
+			fmt.Fprintf(&b, "- **Compute capability:** %s\n", gpu.ComputeCapability)
+		}
+		if gpu.MultiprocessorCount > 0 {
+			fmt.Fprintf(&b, "- **SMs:** %d\n", gpu.MultiprocessorCount)
+		}
+		if gpu.PowerLimitW > 0 {
+			fmt.Fprintf(&b, "- **Power limit:** %.0f W (default %.0f W)\n", gpu.PowerLimitW, gpu.DefaultPowerLimitW)
+		}
+		if gpu.LockedGraphicsClockMHz > 0 {
+			fmt.Fprintf(&b, "- **Locked clocks:** GPU %.0f MHz / Mem %.0f MHz\n", gpu.LockedGraphicsClockMHz, gpu.LockedMemoryClockMHz)
 		}
 		b.WriteString("\n")
+
+		// Steady-state telemetry
+		fmt.Fprintf(&b, "**Steady-state telemetry** (%ds):\n\n", int(gpu.Steady.DurationSec))
+		b.WriteString("| | Avg | P95 |\n|---|---|---|\n")
+		fmt.Fprintf(&b, "| Power | %.1f W | %.1f W |\n", gpu.Steady.AvgPowerW, gpu.Steady.P95PowerW)
+		fmt.Fprintf(&b, "| Temperature | %.1f °C | %.1f °C |\n", gpu.Steady.AvgTempC, gpu.Steady.P95TempC)
+		fmt.Fprintf(&b, "| GPU clock | %.0f MHz | %.0f MHz |\n", gpu.Steady.AvgGraphicsClockMHz, gpu.Steady.P95GraphicsClockMHz)
+		fmt.Fprintf(&b, "| Memory clock | %.0f MHz | %.0f MHz |\n", gpu.Steady.AvgMemoryClockMHz, gpu.Steady.P95MemoryClockMHz)
+		fmt.Fprintf(&b, "| GPU utilisation | %.1f %% | — |\n", gpu.Steady.AvgUsagePct)
+		b.WriteString("\n")
+
+		// Throttle
+		throttle := formatThrottleLine(gpu.Throttle, gpu.Steady.DurationSec)
+		if throttle != "none" {
+			fmt.Fprintf(&b, "**Throttle:** %s\n\n", throttle)
+		}
+
+		// Precision results
+		if len(gpu.PrecisionResults) > 0 {
+			b.WriteString("**Precision results:**\n\n")
+			b.WriteString("| Precision | TOPS | Lanes | Iterations |\n|-----------|------|-------|------------|\n")
+			for _, p := range gpu.PrecisionResults {
+				if p.Supported {
+					fmt.Fprintf(&b, "| %s | %.2f | %d | %d |\n", p.Name, p.TeraOpsPerSec, p.Lanes, p.Iterations)
+				} else {
+					fmt.Fprintf(&b, "| %s | — (unsupported) | — | — |\n", p.Name)
+				}
+			}
+			b.WriteString("\n")
+		}
+
+		// Degradation / Notes
+		if len(gpu.DegradationReasons) > 0 {
+			fmt.Fprintf(&b, "**Degradation reasons:** %s\n\n", strings.Join(gpu.DegradationReasons, ", "))
+		}
+		if len(gpu.Notes) > 0 {
+			b.WriteString("**Notes:**\n\n")
+			for _, note := range gpu.Notes {
+				fmt.Fprintf(&b, "- %s\n", note)
+			}
+			b.WriteString("\n")
+		}
 	}
 
+	// ── Interconnect ──────────────────────────────────────────────────────────
 	if result.Interconnect != nil {
-		fmt.Fprintf(&b, "Interconnect\n")
-		fmt.Fprintf(&b, "------------\n")
-		fmt.Fprintf(&b, "Status: %s\n", result.Interconnect.Status)
+		b.WriteString("## Interconnect (NCCL)\n\n")
+		fmt.Fprintf(&b, "**Status:** %s\n\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)
+			b.WriteString("| Metric | Avg | Max |\n|--------|-----|-----|\n")
+			fmt.Fprintf(&b, "| Alg BW | %.1f GB/s | %.1f GB/s |\n", result.Interconnect.AvgAlgBWGBps, result.Interconnect.MaxAlgBWGBps)
+			fmt.Fprintf(&b, "| Bus BW | %.1f GB/s | %.1f GB/s |\n", result.Interconnect.AvgBusBWGBps, result.Interconnect.MaxBusBWGBps)
+			b.WriteString("\n")
 		}
 		for _, note := range result.Interconnect.Notes {
 			fmt.Fprintf(&b, "- %s\n", note)
 		}
-		b.WriteString("\n")
+		if len(result.Interconnect.Notes) > 0 {
+			b.WriteString("\n")
+		}
 	}
 
+	// ── Server Power (IPMI) ───────────────────────────────────────────────────
+	if sp := result.ServerPower; sp != nil {
+		b.WriteString("## Server Power (IPMI)\n\n")
+		if !sp.Available {
+			b.WriteString("IPMI power measurement unavailable.\n\n")
+		} else {
+			b.WriteString("| | Value |\n|---|---|\n")
+			fmt.Fprintf(&b, "| Server idle | %.0f W |\n", sp.IdleW)
+			fmt.Fprintf(&b, "| Server under load | %.0f W |\n", sp.LoadedW)
+			fmt.Fprintf(&b, "| Server delta (load − idle) | %.0f W |\n", sp.DeltaW)
+			fmt.Fprintf(&b, "| GPU-reported sum | %.0f W |\n", sp.GPUReportedSumW)
+			if sp.ReportingRatio > 0 {
+				fmt.Fprintf(&b, "| Reporting ratio | %.2f (1.0 = accurate, <0.75 = GPU over-reports) |\n", sp.ReportingRatio)
+			}
+			b.WriteString("\n")
+		}
+		for _, note := range sp.Notes {
+			fmt.Fprintf(&b, "- %s\n", note)
+		}
+		if len(sp.Notes) > 0 {
+			b.WriteString("\n")
+		}
+	}
+
+	// ── Terminal charts (steady-state only) ───────────────────────────────────
 	if len(charts) > 0 {
-		fmt.Fprintf(&b, "Terminal Charts\n")
-		fmt.Fprintf(&b, "---------------\n")
+		b.WriteString("## Steady-State Charts\n\n")
 		for _, chart := range charts {
 			content := strings.TrimSpace(stripANSIEscapeSequences(chart.Content))
 			if content == "" {
 				continue
 			}
-			fmt.Fprintf(&b, "%s\n", chart.Title)
-			fmt.Fprintf(&b, "%s\n", strings.Repeat("~", len(chart.Title)))
-			fmt.Fprintf(&b, "%s\n\n", content)
+			fmt.Fprintf(&b, "### %s\n\n```\n%s\n```\n\n", chart.Title, content)
 		}
 	}
 
-	if sp := result.ServerPower; sp != nil {
-		fmt.Fprintf(&b, "Server Power (IPMI)\n")
-		fmt.Fprintf(&b, "-------------------\n")
-		if !sp.Available {
-			fmt.Fprintf(&b, "Unavailable\n")
-		} else {
-			fmt.Fprintf(&b, "  Server idle:         %.0f W\n", sp.IdleW)
-			fmt.Fprintf(&b, "  Server under load:   %.0f W\n", sp.LoadedW)
-			fmt.Fprintf(&b, "  Server delta:        %.0f W\n", sp.DeltaW)
-			fmt.Fprintf(&b, "  GPU reported (sum):  %.0f W\n", sp.GPUReportedSumW)
-			if sp.ReportingRatio > 0 {
-				fmt.Fprintf(&b, "  Reporting ratio:     %.2f  (1.0 = accurate, <0.75 = GPU over-reports)\n", sp.ReportingRatio)
-			}
-		}
-		for _, note := range sp.Notes {
-			fmt.Fprintf(&b, "  Note: %s\n", note)
-		}
-		b.WriteString("\n")
-	}
+	// ── Methodology ───────────────────────────────────────────────────────────
+	b.WriteString("## Methodology\n\n")
+	fmt.Fprintf(&b, "- Profile `%s` uses standardized baseline → warmup → steady-state → interconnect → cooldown phases.\n", result.BenchmarkProfile)
+	b.WriteString("- Single-GPU compute score from bee-gpu-burn cuBLASLt when available.\n")
+	b.WriteString("- Thermal and power limitations inferred from NVIDIA clock event reason counters and sustained telemetry.\n")
+	b.WriteString("- `result.json` is the canonical machine-readable source for this benchmark run.\n\n")
 
-	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")
+	// ── Raw files ─────────────────────────────────────────────────────────────
+	b.WriteString("## Raw Files\n\n")
+	b.WriteString("- `result.json`\n- `report.md`\n- `summary.txt`\n- `verbose.log`\n")
+	b.WriteString("- `gpu-*-baseline-metrics.csv/html/term.txt`\n")
+	b.WriteString("- `gpu-*-warmup.log`\n")
+	b.WriteString("- `gpu-*-steady.log`\n")
+	b.WriteString("- `gpu-*-steady-metrics.csv/html/term.txt`\n")
+	b.WriteString("- `gpu-*-cooldown-metrics.csv/html/term.txt`\n")
 	if result.Interconnect != nil {
-		fmt.Fprintf(&b, "- nccl-all-reduce.log\n")
+		b.WriteString("- `nccl-all-reduce.log`\n")
 	}
 	return b.String()
 }
 
+// loadBenchmarkReportCharts loads only steady-state terminal charts (baseline and
+// cooldown charts are not useful for human review).
 func loadBenchmarkReportCharts(runDir string, gpuIndices []int) []benchmarkReportChart {
-	phases := []struct {
-		name  string
-		label string
-	}{
-		{name: "baseline", label: "Baseline"},
-		{name: "steady", label: "Steady State"},
-		{name: "cooldown", label: "Cooldown"},
-	}
 	var charts []benchmarkReportChart
 	for _, idx := range gpuIndices {
-		for _, phase := range phases {
-			path := filepath.Join(runDir, fmt.Sprintf("gpu-%d-%s-metrics-term.txt", idx, phase.name))
-			raw, err := os.ReadFile(path)
-			if err != nil || len(raw) == 0 {
-				continue
-			}
-			charts = append(charts, benchmarkReportChart{
-				Title:   fmt.Sprintf("GPU %d %s", idx, phase.label),
-				Content: string(raw),
-			})
+		path := filepath.Join(runDir, fmt.Sprintf("gpu-%d-steady-metrics-term.txt", idx))
+		raw, err := os.ReadFile(path)
+		if err != nil || len(raw) == 0 {
+			continue
 		}
+		charts = append(charts, benchmarkReportChart{
+			Title:   fmt.Sprintf("GPU %d — Steady State", idx),
+			Content: string(raw),
+		})
 	}
 	return charts
 }

audit/internal/platform/benchmark_test.go

@@ -137,8 +137,9 @@ func TestRenderBenchmarkReportIncludesFindingsAndScores(t *testing.T) {
 	for _, needle := range []string{
 		"Executive Summary",
 		"GPU 0 spent measurable time under SW power cap.",
-		"Composite score: 1176.00",
-		"fp16_tensor: 700.00 TOPS",
+		"1176.00",
+		"fp16_tensor",
+		"700.00",
 	} {
 		if !strings.Contains(report, needle) {
 			t.Fatalf("report missing %q\n%s", needle, report)
@@ -164,7 +165,7 @@ func TestRenderBenchmarkReportIncludesTerminalChartsWithoutANSI(t *testing.T) {
 	})
 
 	for _, needle := range []string{
-		"Terminal Charts",
+		"Steady-State Charts",
 		"GPU 0 Steady State",
 		"GPU 0 chart",
 		"42┤───",
@@ -177,3 +178,67 @@ func TestRenderBenchmarkReportIncludesTerminalChartsWithoutANSI(t *testing.T) {
 		t.Fatalf("report should not contain ANSI escapes\n%s", report)
 	}
 }
+
+func TestEnrichGPUInfoWithMaxClocks(t *testing.T) {
+	t.Parallel()
+
+	nvsmiQ := []byte(`
+GPU 00000000:4E:00.0
+    Product Name                          : NVIDIA RTX PRO 6000 Blackwell Server Edition
+    Clocks
+        Graphics                          : 2422 MHz
+        Memory                            : 12481 MHz
+    Max Clocks
+        Graphics                          : 2430 MHz
+        SM                                : 2430 MHz
+        Memory                            : 12481 MHz
+        Video                             : 2107 MHz
+
+GPU 00000000:4F:00.0
+    Product Name                          : NVIDIA RTX PRO 6000 Blackwell Server Edition
+    Max Clocks
+        Graphics                          : 2430 MHz
+        Memory                            : 12481 MHz
+`)
+
+	infoByIndex := map[int]benchmarkGPUInfo{
+		0: {Index: 0, BusID: "00000000:4E:00.0"},
+		1: {Index: 1, BusID: "00000000:4F:00.0"},
+	}
+
+	enrichGPUInfoWithMaxClocks(infoByIndex, nvsmiQ)
+
+	if infoByIndex[0].MaxGraphicsClockMHz != 2430 {
+		t.Errorf("GPU 0 MaxGraphicsClockMHz = %v, want 2430", infoByIndex[0].MaxGraphicsClockMHz)
+	}
+	if infoByIndex[0].MaxMemoryClockMHz != 12481 {
+		t.Errorf("GPU 0 MaxMemoryClockMHz = %v, want 12481", infoByIndex[0].MaxMemoryClockMHz)
+	}
+	if infoByIndex[1].MaxGraphicsClockMHz != 2430 {
+		t.Errorf("GPU 1 MaxGraphicsClockMHz = %v, want 2430", infoByIndex[1].MaxGraphicsClockMHz)
+	}
+	if infoByIndex[1].MaxMemoryClockMHz != 12481 {
+		t.Errorf("GPU 1 MaxMemoryClockMHz = %v, want 12481", infoByIndex[1].MaxMemoryClockMHz)
+	}
+}
+
+func TestEnrichGPUInfoWithMaxClocksSkipsPopulated(t *testing.T) {
+	t.Parallel()
+
+	nvsmiQ := []byte(`
+GPU 00000000:4E:00.0
+    Max Clocks
+        Graphics                          : 9999 MHz
+        Memory                            : 9999 MHz
+`)
+	// Already populated — must not be overwritten.
+	infoByIndex := map[int]benchmarkGPUInfo{
+		0: {Index: 0, BusID: "00000000:4E:00.0", MaxGraphicsClockMHz: 2430, MaxMemoryClockMHz: 12481},
+	}
+
+	enrichGPUInfoWithMaxClocks(infoByIndex, nvsmiQ)
+
+	if infoByIndex[0].MaxGraphicsClockMHz != 2430 {
+		t.Errorf("expected existing value to be preserved, got %v", infoByIndex[0].MaxGraphicsClockMHz)
+	}
+}

audit/internal/platform/gpu_metrics.go

@@ -383,10 +383,7 @@ func drawGPUChartSVG(rows []GPUMetricRow, gpuIdx int) string {
 }
 
 const (
-	ansiRed    = "\033[31m"
-	ansiBlue   = "\033[34m"
-	ansiGreen  = "\033[32m"
-	ansiYellow = "\033[33m"
+	ansiAmber  = "\033[38;5;214m"
 	ansiReset  = "\033[0m"
 )
 
@@ -415,10 +412,10 @@ func RenderGPUTerminalChart(rows []GPUMetricRow) string {
 		fn      func(GPUMetricRow) float64
 	}
 	defs := []seriesDef{
-		{"Temperature (°C)", ansiRed, func(r GPUMetricRow) float64 { return r.TempC }},
-		{"GPU Usage (%)", ansiBlue, func(r GPUMetricRow) float64 { return r.UsagePct }},
-		{"Power (W)", ansiGreen, func(r GPUMetricRow) float64 { return r.PowerW }},
-		{"Clock (MHz)", ansiYellow, func(r GPUMetricRow) float64 { return r.ClockMHz }},
+		{"Temperature (°C)", ansiAmber, func(r GPUMetricRow) float64 { return r.TempC }},
+		{"GPU Usage (%)", ansiAmber, func(r GPUMetricRow) float64 { return r.UsagePct }},
+		{"Power (W)", ansiAmber, func(r GPUMetricRow) float64 { return r.PowerW }},
+		{"Clock (MHz)", ansiAmber, func(r GPUMetricRow) float64 { return r.ClockMHz }},
 	}
 
 	var b strings.Builder

audit/internal/platform/install_to_ram.go

@@ -116,25 +116,47 @@ func (s *System) RunInstallToRAM(ctx context.Context, logFunc func(string)) erro
 	if err := ctx.Err(); err != nil {
 		return err
 	}
-	if err := exec.Command("mount", "--bind", dstDir, "/run/live/medium").Run(); err != nil {
-		log(fmt.Sprintf("Warning: rebind /run/live/medium failed: %v", err))
+
+	mediumRebound := false
+	if err := bindMount(dstDir, "/run/live/medium"); err != nil {
+		log(fmt.Sprintf("Warning: rebind /run/live/medium → %s failed: %v", dstDir, err))
+	} else {
+		mediumRebound = true
 	}
 
 	log("Verifying live medium now served from RAM...")
 	status := s.LiveBootSource()
-	if err := verifyInstallToRAMStatus(status); err != nil {
+	if err := verifyInstallToRAMStatus(status, dstDir, mediumRebound, log); err != nil {
 		return err
 	}
-	log(fmt.Sprintf("Verification passed: live medium now served from %s.", describeLiveBootSource(status)))
-	log("Done. Installation media can be safely disconnected.")
+	if status.InRAM {
+		log(fmt.Sprintf("Verification passed: live medium now served from %s.", describeLiveBootSource(status)))
+	}
+	log("Done. Squashfs files are in RAM. Installation media can be safely disconnected.")
 	return nil
 }
 
-func verifyInstallToRAMStatus(status LiveBootSource) error {
+func verifyInstallToRAMStatus(status LiveBootSource, dstDir string, mediumRebound bool, log func(string)) error {
 	if status.InRAM {
 		return nil
 	}
-	return fmt.Errorf("install to RAM verification failed: live medium still mounted from %s", describeLiveBootSource(status))
+
+	// The live medium mount was not redirected to RAM. This is expected when
+	// booting from an ISO/CD-ROM: the squashfs loop device has a non-zero
+	// offset and LOOP_CHANGE_FD cannot be used; the bind mount also fails
+	// because the CD-ROM mount is in use. Check whether files were at least
+	// copied to the tmpfs directory — that is sufficient for safe disconnection
+	// once the kernel has paged in all actively-used data.
+	files, _ := filepath.Glob(filepath.Join(dstDir, "*.squashfs"))
+	if len(files) > 0 {
+		if !mediumRebound {
+			log(fmt.Sprintf("Note: squashfs copied to RAM (%s) but /run/live/medium still shows the original source.", dstDir))
+			log("This is normal for CD-ROM boots. For a fully transparent RAM boot, add 'toram' to the kernel parameters.")
+		}
+		return nil
+	}
+
+	return fmt.Errorf("install to RAM verification failed: live medium still mounted from %s and no squashfs found in %s", describeLiveBootSource(status), dstDir)
 }
 
 func describeLiveBootSource(status LiveBootSource) string {
@@ -247,7 +269,31 @@ func findLoopForFile(backingFile string) (string, error) {
 	return "", fmt.Errorf("no loop device found for %s", backingFile)
 }
 
+// loopDeviceOffset returns the byte offset configured for the loop device,
+// or -1 if it cannot be determined.
+func loopDeviceOffset(loopDev string) int64 {
+	out, err := exec.Command("losetup", "--json", loopDev).Output()
+	if err != nil {
+		return -1
+	}
+	var result struct {
+		Loopdevices []struct {
+			Offset int64 `json:"offset"`
+		} `json:"loopdevices"`
+	}
+	if err := json.Unmarshal(out, &result); err != nil || len(result.Loopdevices) == 0 {
+		return -1
+	}
+	return result.Loopdevices[0].Offset
+}
+
 func reassociateLoopDevice(loopDev, newFile string) error {
+	// LOOP_CHANGE_FD requires lo_offset == 0. ISO/CD-ROM loop devices are
+	// typically set up with a non-zero offset (squashfs lives inside the ISO),
+	// so the ioctl returns EINVAL. Detect this early for a clear error message.
+	if off := loopDeviceOffset(loopDev); off > 0 {
+		return fmt.Errorf("loop device has non-zero offset (%d bytes, typical for ISO/CD-ROM) — LOOP_CHANGE_FD not supported; use 'toram' kernel parameter for RAM boot", off)
+	}
 	if err := exec.Command("losetup", "--replace", loopDev, newFile).Run(); err == nil {
 		return nil
 	}

audit/internal/platform/install_to_ram_linux.go

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

audit/internal/platform/install_to_ram_other.go

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

audit/internal/platform/install_to_ram_test.go

@@ -33,14 +33,17 @@ func TestInferLiveBootKind(t *testing.T) {
 func TestVerifyInstallToRAMStatus(t *testing.T) {
 	t.Parallel()
 
-	if err := verifyInstallToRAMStatus(LiveBootSource{InRAM: true, Kind: "ram", Source: "tmpfs"}); err != nil {
+	dstDir := t.TempDir()
+
+	if err := verifyInstallToRAMStatus(LiveBootSource{InRAM: true, Kind: "ram", Source: "tmpfs"}, dstDir, false, nil); err != nil {
 		t.Fatalf("expected success for RAM-backed status, got %v", err)
 	}
-	err := verifyInstallToRAMStatus(LiveBootSource{InRAM: false, Kind: "usb", Device: "/dev/sdb1"})
+
+	err := verifyInstallToRAMStatus(LiveBootSource{InRAM: false, Kind: "usb", Device: "/dev/sdb1"}, dstDir, false, nil)
 	if err == nil {
 		t.Fatal("expected verification failure when media is still on USB")
 	}
-	if got := err.Error(); got != "install to RAM verification failed: live medium still mounted from USB (/dev/sdb1)" {
+	if got := err.Error(); got != "install to RAM verification failed: live medium still mounted from USB (/dev/sdb1) and no squashfs found in "+dstDir {
 		t.Fatalf("error=%q", got)
 	}
 }

audit/internal/platform/nvidia_stress.go

@@ -49,6 +49,9 @@ func buildNvidiaStressJob(opts NvidiaStressOptions) (satJob, error) {
 			"--seconds", strconv.Itoa(opts.DurationSec),
 			"--size-mb", strconv.Itoa(opts.SizeMB),
 		}
+		if opts.StaggerSeconds > 0 && len(selected) > 1 {
+			cmd = append(cmd, "--stagger-seconds", strconv.Itoa(opts.StaggerSeconds))
+		}
 		if len(selected) > 0 {
 			cmd = append(cmd, "--devices", joinIndexList(selected))
 		}
@@ -63,6 +66,9 @@ func buildNvidiaStressJob(opts NvidiaStressOptions) (satJob, error) {
 			"bee-john-gpu-stress",
 			"--seconds", strconv.Itoa(opts.DurationSec),
 		}
+		if opts.StaggerSeconds > 0 && len(selected) > 1 {
+			cmd = append(cmd, "--stagger-seconds", strconv.Itoa(opts.StaggerSeconds))
+		}
 		if len(selected) > 0 {
 			cmd = append(cmd, "--devices", joinIndexList(selected))
 		}

audit/internal/platform/platform_stress.go

@@ -161,13 +161,7 @@ func (s *System) RunPlatformStress(
 	}
 	_ = os.WriteFile(filepath.Join(runDir, "summary.txt"), []byte(summary), 0644)
 
-	// Pack tar.gz
-	archivePath := filepath.Join(baseDir, "platform-stress-"+stamp+".tar.gz")
-	if err := packPlatformDir(runDir, archivePath); err != nil {
-		return "", fmt.Errorf("pack archive: %w", err)
-	}
-	_ = os.RemoveAll(runDir)
-	return archivePath, nil
+	return runDir, nil
 }
 
 // collectPhase samples live metrics every second until ctx is done.

audit/internal/platform/runtime.go

@@ -1,6 +1,7 @@
 package platform
 
 import (
+	"bufio"
 	"os"
 	"os/exec"
 	"strings"
@@ -114,6 +115,8 @@ func (s *System) CollectRuntimeHealth(exportDir string) (schema.RuntimeHealth, e
 	}
 
 	s.collectGPURuntimeHealth(vendor, &health)
+	s.collectToRAMHealth(&health)
+	s.collectUSBExportHealth(&health)
 
 	if health.Status != "FAILED" && len(health.Issues) > 0 {
 		health.Status = "PARTIAL"
@@ -168,6 +171,90 @@ func resolvedToolStatus(display string, candidates ...string) ToolStatus {
 	return ToolStatus{Name: display}
 }
 
+// collectToRAMHealth checks whether the LiveCD ISO has been copied to RAM.
+// Status values: "ok" = in RAM, "warning" = toram not active (no copy attempted),
+// "failed" = toram was requested but medium is not in RAM (copy failed or in progress).
+func (s *System) collectToRAMHealth(health *schema.RuntimeHealth) {
+	inRAM := s.IsLiveMediaInRAM()
+	active := toramActive()
+	switch {
+	case inRAM:
+		health.ToRAMStatus = "ok"
+	case active:
+		// toram was requested but medium is not yet/no longer in RAM
+		health.ToRAMStatus = "failed"
+		health.Issues = append(health.Issues, schema.RuntimeIssue{
+			Code:        "toram_copy_failed",
+			Severity:    "warning",
+			Description: "toram boot parameter is set but the live medium is not mounted from RAM.",
+		})
+	default:
+		health.ToRAMStatus = "warning"
+	}
+}
+
+// collectUSBExportHealth scans /proc/mounts for a writable USB-backed filesystem
+// suitable for log export. Sets USBExportPath to the first match found.
+func (s *System) collectUSBExportHealth(health *schema.RuntimeHealth) {
+	health.USBExportPath = findUSBExportMount()
+}
+
+// findUSBExportMount returns the mount point of the first writable USB filesystem
+// found in /proc/mounts (vfat, exfat, ext2/3/4, ntfs) whose backing block device
+// has USB transport. Returns "" if none found.
+func findUSBExportMount() string {
+	f, err := os.Open("/proc/mounts")
+	if err != nil {
+		return ""
+	}
+	defer f.Close()
+
+	// fs types that are expected on USB export drives
+	exportFSTypes := map[string]bool{
+		"vfat":  true,
+		"exfat": true,
+		"ext2":  true,
+		"ext3":  true,
+		"ext4":  true,
+		"ntfs":  true,
+		"ntfs3": true,
+		"fuseblk": true,
+	}
+
+	scanner := bufio.NewScanner(f)
+	for scanner.Scan() {
+		// fields: device mountpoint fstype options dump pass
+		fields := strings.Fields(scanner.Text())
+		if len(fields) < 4 {
+			continue
+		}
+		device, mountPoint, fsType, options := fields[0], fields[1], fields[2], fields[3]
+		if !exportFSTypes[strings.ToLower(fsType)] {
+			continue
+		}
+		// Skip read-only mounts
+		opts := strings.Split(options, ",")
+		readOnly := false
+		for _, o := range opts {
+			if strings.TrimSpace(o) == "ro" {
+				readOnly = true
+				break
+			}
+		}
+		if readOnly {
+			continue
+		}
+		// Check USB transport via lsblk on the device
+		if !strings.HasPrefix(device, "/dev/") {
+			continue
+		}
+		if blockDeviceTransport(device) == "usb" {
+			return mountPoint
+		}
+	}
+	return ""
+}
+
 func (s *System) collectGPURuntimeHealth(vendor string, health *schema.RuntimeHealth) {
 	lsmodText := commandText("lsmod")
 

audit/internal/platform/sat.go

@@ -88,6 +88,37 @@ type NvidiaGPU struct {
 	MemoryMB int    `json:"memory_mb"`
 }
 
+type NvidiaGPUStatus struct {
+	Index        int    `json:"index"`
+	Name         string `json:"name"`
+	BDF          string `json:"bdf,omitempty"`
+	Serial       string `json:"serial,omitempty"`
+	Status       string `json:"status"`
+	RawLine      string `json:"raw_line,omitempty"`
+	NeedsReset   bool   `json:"needs_reset"`
+	ParseFailure bool   `json:"parse_failure,omitempty"`
+}
+
+type nvidiaGPUHealth struct {
+	Index        int
+	Name         string
+	NeedsReset   bool
+	RawLine      string
+	ParseFailure bool
+}
+
+type nvidiaGPUStatusFile struct {
+	Index       int
+	Name        string
+	RunStatus   string
+	Reason      string
+	Health      string
+	HealthRaw   string
+	Observed    bool
+	Selected    bool
+	FailingJob  string
+}
+
 // AMDGPUInfo holds basic info about an AMD GPU from rocm-smi.
 type AMDGPUInfo struct {
 	Index int    `json:"index"`
@@ -269,6 +300,72 @@ func (s *System) ListNvidiaGPUs() ([]NvidiaGPU, error) {
 	return gpus, nil
 }
 
+func (s *System) ListNvidiaGPUStatuses() ([]NvidiaGPUStatus, error) {
+	out, err := satExecCommand(
+		"nvidia-smi",
+		"--query-gpu=index,name,pci.bus_id,serial,temperature.gpu,power.draw,utilization.gpu,memory.used,memory.total",
+		"--format=csv,noheader,nounits",
+	).Output()
+	if err != nil {
+		return nil, fmt.Errorf("nvidia-smi: %w", err)
+	}
+	var gpus []NvidiaGPUStatus
+	for _, line := range strings.Split(strings.TrimSpace(string(out)), "\n") {
+		line = strings.TrimSpace(line)
+		if line == "" {
+			continue
+		}
+		parts := strings.Split(line, ",")
+		if len(parts) < 4 {
+			gpus = append(gpus, NvidiaGPUStatus{RawLine: line, Status: "UNKNOWN", ParseFailure: true})
+			continue
+		}
+		idx, err := strconv.Atoi(strings.TrimSpace(parts[0]))
+		if err != nil {
+			gpus = append(gpus, NvidiaGPUStatus{RawLine: line, Status: "UNKNOWN", ParseFailure: true})
+			continue
+		}
+		upper := strings.ToUpper(line)
+		needsReset := strings.Contains(upper, "GPU REQUIRES RESET")
+		status := "OK"
+		if needsReset {
+			status = "RESET_REQUIRED"
+		}
+		gpus = append(gpus, NvidiaGPUStatus{
+			Index:      idx,
+			Name:       strings.TrimSpace(parts[1]),
+			BDF:        normalizeNvidiaBusID(strings.TrimSpace(parts[2])),
+			Serial:     strings.TrimSpace(parts[3]),
+			Status:     status,
+			RawLine:    line,
+			NeedsReset: needsReset,
+		})
+	}
+	sort.Slice(gpus, func(i, j int) bool { return gpus[i].Index < gpus[j].Index })
+	return gpus, nil
+}
+
+func normalizeNvidiaBusID(v string) string {
+	v = strings.TrimSpace(strings.ToLower(v))
+	parts := strings.Split(v, ":")
+	if len(parts) == 3 && len(parts[0]) > 4 {
+		parts[0] = parts[0][len(parts[0])-4:]
+		return strings.Join(parts, ":")
+	}
+	return v
+}
+
+func (s *System) ResetNvidiaGPU(index int) (string, error) {
+	if index < 0 {
+		return "", fmt.Errorf("gpu index must be >= 0")
+	}
+	raw, err := satExecCommand("nvidia-smi", "-r", "-i", strconv.Itoa(index)).CombinedOutput()
+	if len(raw) == 0 && err == nil {
+		raw = []byte("GPU reset completed.\n")
+	}
+	return string(raw), err
+}
+
 // RunNCCLTests runs nccl-tests all_reduce_perf across all NVIDIA GPUs.
 // Measures collective communication bandwidth over NVLink/PCIe.
 func (s *System) RunNCCLTests(ctx context.Context, baseDir string, logFunc func(string)) (string, error) {
@@ -287,25 +384,39 @@ func (s *System) RunNCCLTests(ctx context.Context, baseDir string, logFunc func(
 	), logFunc)
 }
 
-func (s *System) RunNvidiaOfficialComputePack(ctx context.Context, baseDir string, durationSec int, gpuIndices []int, logFunc func(string)) (string, error) {
+func (s *System) RunNvidiaOfficialComputePack(ctx context.Context, baseDir string, durationSec int, gpuIndices []int, staggerSec int, logFunc func(string)) (string, error) {
 	selected, err := resolveDCGMGPUIndices(gpuIndices)
 	if err != nil {
 		return "", err
 	}
-	profCmd, err := resolveDCGMProfTesterCommand("--no-dcgm-validation", "-t", "1004", "-d", strconv.Itoa(normalizeNvidiaBurnDuration(durationSec)))
-	if err != nil {
-		return "", err
+	var (
+		profCmd []string
+		profEnv []string
+	)
+	if staggerSec > 0 && len(selected) > 1 {
+		profCmd = []string{
+			"bee-dcgmproftester-staggered",
+			"--seconds", strconv.Itoa(normalizeNvidiaBurnDuration(durationSec)),
+			"--stagger-seconds", strconv.Itoa(staggerSec),
+			"--devices", joinIndexList(selected),
+		}
+	} else {
+		profCmd, err = resolveDCGMProfTesterCommand("--no-dcgm-validation", "-t", "1004", "-d", strconv.Itoa(normalizeNvidiaBurnDuration(durationSec)))
+		if err != nil {
+			return "", err
+		}
+		profEnv = nvidiaVisibleDevicesEnv(selected)
 	}
 	return runAcceptancePackCtx(ctx, baseDir, "gpu-nvidia-compute", withNvidiaPersistenceMode(
 		satJob{name: "01-nvidia-smi-q.log", cmd: []string{"nvidia-smi", "-q"}},
 		satJob{name: "02-dcgmi-version.log", cmd: []string{"dcgmi", "-v"}},
-		satJob{
-			name:       "03-dcgmproftester.log",
-			cmd:        profCmd,
-			env:        nvidiaVisibleDevicesEnv(selected),
-			collectGPU: true,
-			gpuIndices: selected,
-		},
+			satJob{
+				name:       "03-dcgmproftester.log",
+				cmd:        profCmd,
+				env:        profEnv,
+				collectGPU: true,
+				gpuIndices: selected,
+			},
 		satJob{name: "04-nvidia-smi-after.log", cmd: []string{"nvidia-smi", "--query-gpu=index,name,temperature.gpu,power.draw,utilization.gpu,memory.used,memory.total", "--format=csv,noheader,nounits"}},
 	), logFunc)
 }
@@ -434,9 +545,13 @@ func memoryStressSizeArg() string {
 	return fmt.Sprintf("%dM", targetMB)
 }
 
-func (s *System) RunMemoryAcceptancePack(ctx context.Context, baseDir string, logFunc func(string)) (string, error) {
-	sizeMB := envInt("BEE_MEMTESTER_SIZE_MB", 128)
-	passes := envInt("BEE_MEMTESTER_PASSES", 1)
+func (s *System) RunMemoryAcceptancePack(ctx context.Context, baseDir string, sizeMB, passes int, logFunc func(string)) (string, error) {
+	if sizeMB <= 0 {
+		sizeMB = 256
+	}
+	if passes <= 0 {
+		passes = 1
+	}
 	return runAcceptancePackCtx(ctx, baseDir, "memory", []satJob{
 		{name: "01-free-before.log", cmd: []string{"free", "-h"}},
 		{name: "02-memtester.log", cmd: []string{"memtester", fmt.Sprintf("%dM", sizeMB), fmt.Sprintf("%d", passes)}},
@@ -493,7 +608,7 @@ func (s *System) RunCPUAcceptancePack(ctx context.Context, baseDir string, durat
 	}, logFunc)
 }
 
-func (s *System) RunStorageAcceptancePack(ctx context.Context, baseDir string, logFunc func(string)) (string, error) {
+func (s *System) RunStorageAcceptancePack(ctx context.Context, baseDir string, extended bool, logFunc func(string)) (string, error) {
 	if baseDir == "" {
 		baseDir = "/var/log/bee-sat"
 	}
@@ -525,7 +640,7 @@ func (s *System) RunStorageAcceptancePack(ctx context.Context, baseDir string, l
 			break
 		}
 		prefix := fmt.Sprintf("%02d-%s", index+1, filepath.Base(devPath))
-		commands := storageSATCommands(devPath)
+		commands := storageSATCommands(devPath, extended)
 		for cmdIndex, job := range commands {
 			if ctx.Err() != nil {
 				break
@@ -547,11 +662,7 @@ func (s *System) RunStorageAcceptancePack(ctx context.Context, baseDir string, l
 	if err := os.WriteFile(filepath.Join(runDir, "summary.txt"), []byte(summary.String()), 0644); err != nil {
 		return "", err
 	}
-	archive := filepath.Join(baseDir, "storage-"+ts+".tar.gz")
-	if err := createTarGz(archive, runDir); err != nil {
-		return "", err
-	}
-	return archive, nil
+	return runDir, nil
 }
 
 type satJob struct {
@@ -604,7 +715,7 @@ func nvidiaDCGMJobs(diagLevel int, gpuIndices []int) []satJob {
 		satJob{name: "01-nvidia-smi-q.log", cmd: []string{"nvidia-smi", "-q"}},
 		satJob{name: "02-dmidecode-baseboard.log", cmd: []string{"dmidecode", "-t", "baseboard"}},
 		satJob{name: "03-dmidecode-system.log", cmd: []string{"dmidecode", "-t", "system"}},
-		satJob{name: "04-dcgmi-diag.log", cmd: diagArgs},
+		satJob{name: "04-dcgmi-diag.log", cmd: diagArgs, gpuIndices: gpuIndices},
 	)
 }
 
@@ -652,11 +763,23 @@ func runAcceptancePackCtx(ctx context.Context, baseDir, prefix string, jobs []sa
 
 	var summary strings.Builder
 	stats := satStats{}
+	nvidiaPack := strings.HasPrefix(prefix, "gpu-nvidia")
+	perGPU := map[int]*nvidiaGPUStatusFile{}
+	selectedGPUIndices := map[int]struct{}{}
 	fmt.Fprintf(&summary, "run_at_utc=%s\n", time.Now().UTC().Format(time.RFC3339))
 	for _, job := range jobs {
 		if ctx.Err() != nil {
 			break
 		}
+		for _, idx := range job.gpuIndices {
+			selectedGPUIndices[idx] = struct{}{}
+			status := perGPU[idx]
+			if status == nil {
+				status = &nvidiaGPUStatusFile{Index: idx}
+				perGPU[idx] = status
+			}
+			status.Selected = true
+		}
 		cmd := make([]string, 0, len(job.cmd))
 		for _, arg := range job.cmd {
 			cmd = append(cmd, strings.ReplaceAll(arg, "{{run_dir}}", runDir))
@@ -665,10 +788,37 @@ func runAcceptancePackCtx(ctx context.Context, baseDir, prefix string, jobs []sa
 		var out []byte
 		var err error
 
-		if job.collectGPU {
-			out, err = runSATCommandWithMetrics(ctx, verboseLog, job.name, cmd, job.env, job.gpuIndices, runDir, logFunc)
-		} else {
-			out, err = runSATCommandCtx(ctx, verboseLog, job.name, cmd, job.env, logFunc)
+		if nvidiaPack && nvidiaJobNeedsHealthCheck(job) {
+			if msg, healthErr := checkNvidiaJobHealth(job.gpuIndices); healthErr != nil {
+				if logFunc != nil {
+					logFunc(msg)
+				}
+				out = []byte(msg + "\n")
+				err = healthErr
+			}
+		}
+
+		if err == nil {
+			if job.collectGPU {
+				out, err = runSATCommandWithMetrics(ctx, verboseLog, job.name, cmd, job.env, job.gpuIndices, runDir, logFunc)
+			} else {
+				out, err = runSATCommandCtx(ctx, verboseLog, job.name, cmd, job.env, logFunc)
+			}
+		}
+
+		if nvidiaPack && nvidiaJobNeedsHealthCheck(job) {
+			if msg, healthErr := checkNvidiaJobHealth(job.gpuIndices); healthErr != nil {
+				if logFunc != nil {
+					logFunc(msg)
+				}
+				if len(out) > 0 && !bytes.HasSuffix(out, []byte("\n")) {
+					out = append(out, '\n')
+				}
+				out = append(out, []byte(msg+"\n")...)
+				if err == nil {
+					err = healthErr
+				}
+			}
 		}
 
 		if writeErr := os.WriteFile(filepath.Join(runDir, job.name), out, 0644); writeErr != nil {
@@ -679,6 +829,11 @@ func runAcceptancePackCtx(ctx context.Context, baseDir, prefix string, jobs []sa
 		}
 		status, rc := classifySATResult(job.name, out, err)
 		stats.Add(status)
+		if nvidiaPack && len(job.gpuIndices) > 0 && nvidiaJobNeedsHealthCheck(job) {
+			for _, idx := range job.gpuIndices {
+				updateNvidiaGPUStatus(perGPU, idx, status, job.name, string(out))
+			}
+		}
 		key := strings.TrimSuffix(strings.TrimPrefix(job.name, "0"), ".log")
 		fmt.Fprintf(&summary, "%s_rc=%d\n", key, rc)
 		fmt.Fprintf(&summary, "%s_status=%s\n", key, status)
@@ -687,12 +842,204 @@ func runAcceptancePackCtx(ctx context.Context, baseDir, prefix string, jobs []sa
 	if err := os.WriteFile(filepath.Join(runDir, "summary.txt"), []byte(summary.String()), 0644); err != nil {
 		return "", err
 	}
-
-	archive := filepath.Join(baseDir, prefix+"-"+ts+".tar.gz")
-	if err := createTarGz(archive, runDir); err != nil {
-		return "", err
+	if nvidiaPack {
+		if err := writeNvidiaGPUStatusFiles(runDir, stats.Overall(), perGPU, selectedGPUIndices); err != nil {
+			return "", err
+		}
 	}
-	return archive, nil
+
+	return runDir, nil
+}
+
+func updateNvidiaGPUStatus(perGPU map[int]*nvidiaGPUStatusFile, idx int, status, jobName, detail string) {
+	entry := perGPU[idx]
+	if entry == nil {
+		entry = &nvidiaGPUStatusFile{Index: idx}
+		perGPU[idx] = entry
+	}
+	if nvidiaSATStatusSeverity(status) >= nvidiaSATStatusSeverity(entry.RunStatus) {
+		entry.RunStatus = status
+		entry.FailingJob = jobName
+		entry.Reason = firstLine(detail)
+	}
+}
+
+func writeNvidiaGPUStatusFiles(runDir, overall string, perGPU map[int]*nvidiaGPUStatusFile, selected map[int]struct{}) error {
+	health, err := readNvidiaGPUHealth()
+	if err == nil {
+		for _, gpu := range health {
+			entry := perGPU[gpu.Index]
+			if entry == nil {
+				entry = &nvidiaGPUStatusFile{Index: gpu.Index}
+				perGPU[gpu.Index] = entry
+			}
+			entry.Name = gpu.Name
+			entry.Observed = true
+			entry.HealthRaw = gpu.RawLine
+			if gpu.NeedsReset {
+				entry.Health = "RESET_REQUIRED"
+				if entry.RunStatus == "" || nvidiaSATStatusSeverity("FAILED") >= nvidiaSATStatusSeverity(entry.RunStatus) {
+					entry.RunStatus = "FAILED"
+					if strings.TrimSpace(entry.Reason) == "" {
+						entry.Reason = "GPU requires reset"
+					}
+				}
+			} else {
+				entry.Health = "OK"
+			}
+		}
+	}
+	for idx := range selected {
+		entry := perGPU[idx]
+		if entry == nil {
+			entry = &nvidiaGPUStatusFile{Index: idx}
+			perGPU[idx] = entry
+		}
+		entry.Selected = true
+	}
+	var indices []int
+	for idx := range perGPU {
+		indices = append(indices, idx)
+	}
+	sort.Ints(indices)
+	for _, idx := range indices {
+		entry := perGPU[idx]
+		if entry.RunStatus == "" {
+			entry.RunStatus = overall
+		}
+		if entry.Health == "" {
+			entry.Health = "UNKNOWN"
+		}
+		if entry.Name == "" {
+			entry.Name = "Unknown GPU"
+		}
+		var body strings.Builder
+		fmt.Fprintf(&body, "gpu_index=%d\n", entry.Index)
+		fmt.Fprintf(&body, "gpu_name=%s\n", entry.Name)
+		fmt.Fprintf(&body, "selected=%t\n", entry.Selected)
+		fmt.Fprintf(&body, "observed=%t\n", entry.Observed)
+		fmt.Fprintf(&body, "run_status=%s\n", entry.RunStatus)
+		fmt.Fprintf(&body, "health_status=%s\n", entry.Health)
+		if strings.TrimSpace(entry.FailingJob) != "" {
+			fmt.Fprintf(&body, "failing_job=%s\n", entry.FailingJob)
+		}
+		if strings.TrimSpace(entry.Reason) != "" {
+			fmt.Fprintf(&body, "reason=%s\n", entry.Reason)
+		}
+		if strings.TrimSpace(entry.HealthRaw) != "" {
+			fmt.Fprintf(&body, "health_raw=%s\n", entry.HealthRaw)
+		}
+		if err := os.WriteFile(filepath.Join(runDir, fmt.Sprintf("gpu-%d-status.txt", idx)), []byte(body.String()), 0644); err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
+func nvidiaSATStatusSeverity(status string) int {
+	switch strings.ToUpper(strings.TrimSpace(status)) {
+	case "FAILED":
+		return 3
+	case "PARTIAL", "UNSUPPORTED":
+		return 2
+	case "OK":
+		return 1
+	default:
+		return 0
+	}
+}
+
+func firstLine(s string) string {
+	s = strings.TrimSpace(s)
+	if s == "" {
+		return ""
+	}
+	if idx := strings.IndexByte(s, '\n'); idx >= 0 {
+		return strings.TrimSpace(s[:idx])
+	}
+	return s
+}
+
+func nvidiaJobNeedsHealthCheck(job satJob) bool {
+	if job.collectGPU {
+		return true
+	}
+	name := strings.ToLower(strings.TrimSpace(job.name))
+	return strings.Contains(name, "dcgmi") ||
+		strings.Contains(name, "gpu-burn") ||
+		strings.Contains(name, "gpu-stress") ||
+		strings.Contains(name, "dcgmproftester")
+}
+
+func checkNvidiaJobHealth(selected []int) (string, error) {
+	health, err := readNvidiaGPUHealth()
+	if err != nil {
+		return "", nil
+	}
+	var bad []nvidiaGPUHealth
+	selectedSet := make(map[int]struct{}, len(selected))
+	for _, idx := range selected {
+		selectedSet[idx] = struct{}{}
+	}
+	for _, gpu := range health {
+		if len(selectedSet) > 0 {
+			if _, ok := selectedSet[gpu.Index]; !ok {
+				continue
+			}
+		}
+		if gpu.NeedsReset {
+			bad = append(bad, gpu)
+		}
+	}
+	if len(bad) == 0 {
+		return "", nil
+	}
+	lines := make([]string, 0, len(bad)+1)
+	lines = append(lines, "NVIDIA GPU health check failed:")
+	for _, gpu := range bad {
+		lines = append(lines, fmt.Sprintf("gpu %d (%s) requires reset: %s", gpu.Index, gpu.Name, gpu.RawLine))
+	}
+	return strings.Join(lines, "\n"), errors.New("nvidia gpu requires reset")
+}
+
+func readNvidiaGPUHealth() ([]nvidiaGPUHealth, error) {
+	out, err := satExecCommand(
+		"nvidia-smi",
+		"--query-gpu=index,name,temperature.gpu,power.draw,utilization.gpu,memory.used,memory.total",
+		"--format=csv,noheader,nounits",
+	).Output()
+	if err != nil {
+		return nil, fmt.Errorf("nvidia-smi: %w", err)
+	}
+	return parseNvidiaGPUHealth(string(out)), nil
+}
+
+func parseNvidiaGPUHealth(raw string) []nvidiaGPUHealth {
+	var gpus []nvidiaGPUHealth
+	for _, line := range strings.Split(strings.TrimSpace(raw), "\n") {
+		line = strings.TrimSpace(line)
+		if line == "" {
+			continue
+		}
+		parts := strings.Split(line, ",")
+		if len(parts) < 2 {
+			gpus = append(gpus, nvidiaGPUHealth{RawLine: line, ParseFailure: true})
+			continue
+		}
+		idx, err := strconv.Atoi(strings.TrimSpace(parts[0]))
+		if err != nil {
+			gpus = append(gpus, nvidiaGPUHealth{RawLine: line, ParseFailure: true})
+			continue
+		}
+		upper := strings.ToUpper(line)
+		gpus = append(gpus, nvidiaGPUHealth{
+			Index:      idx,
+			Name:       strings.TrimSpace(parts[1]),
+			NeedsReset: strings.Contains(upper, "GPU REQUIRES RESET"),
+			RawLine:    line,
+		})
+	}
+	return gpus
 }
 
 func runSATCommandCtx(ctx context.Context, verboseLog, name string, cmd []string, env []string, logFunc func(string)) ([]byte, error) {
@@ -749,17 +1096,25 @@ func listStorageDevices() ([]string, error) {
 	return parseStorageDevices(string(out)), nil
 }
 
-func storageSATCommands(devPath string) []satJob {
+func storageSATCommands(devPath string, extended bool) []satJob {
 	if strings.Contains(filepath.Base(devPath), "nvme") {
+		selfTestLevel := "1"
+		if extended {
+			selfTestLevel = "2"
+		}
 		return []satJob{
 			{name: "nvme-id-ctrl", cmd: []string{"nvme", "id-ctrl", devPath, "-o", "json"}},
 			{name: "nvme-smart-log", cmd: []string{"nvme", "smart-log", devPath, "-o", "json"}},
-			{name: "nvme-device-self-test", cmd: []string{"nvme", "device-self-test", devPath, "-s", "1", "--wait"}},
+			{name: "nvme-device-self-test", cmd: []string{"nvme", "device-self-test", devPath, "-s", selfTestLevel, "--wait"}},
 		}
 	}
+	smartTestType := "short"
+	if extended {
+		smartTestType = "long"
+	}
 	return []satJob{
 		{name: "smartctl-health", cmd: []string{"smartctl", "-H", "-A", devPath}},
-		{name: "smartctl-self-test-short", cmd: []string{"smartctl", "-t", "short", devPath}},
+		{name: "smartctl-self-test-short", cmd: []string{"smartctl", "-t", smartTestType, devPath}},
 	}
 }
 
@@ -818,6 +1173,11 @@ func classifySATResult(name string, out []byte, err error) (string, int) {
 		// nvidia-smi on a machine with no NVIDIA GPU
 		strings.Contains(text, "couldn't communicate with the nvidia driver") ||
 		strings.Contains(text, "no nvidia gpu") ||
+		// Some NVMe firmwares start self-test but never expose progress to nvme-cli
+		// while waiting, so the CLI stops polling without proving device failure.
+		(strings.Contains(name, "self-test") &&
+			strings.Contains(text, "no progress for") &&
+			strings.Contains(text, "stop waiting")) ||
 		(strings.Contains(name, "self-test") && strings.Contains(text, "aborted")) {
 		return "UNSUPPORTED", rc
 	}

audit/internal/platform/sat_fan_stress.go

@@ -20,7 +20,7 @@ type FanStressOptions struct {
 	Phase1DurSec int   // first load phase duration in seconds (default 300)
 	PauseSec     int   // pause between the two load phases (default 60)
 	Phase2DurSec int   // second load phase duration in seconds (default 300)
-	SizeMB       int   // GPU memory to allocate per GPU during stress (default 64)
+	SizeMB       int   // GPU memory to allocate per GPU during stress (0 = auto: 95% of VRAM)
 	GPUIndices   []int // which GPU indices to stress (empty = all detected)
 }
 
@@ -223,11 +223,7 @@ func (s *System) RunFanStressTest(ctx context.Context, baseDir string, opts FanS
 		return "", err
 	}
 
-	archive := filepath.Join(baseDir, "fan-stress-"+ts+".tar.gz")
-	if err := createTarGz(archive, runDir); err != nil {
-		return "", err
-	}
-	return archive, nil
+	return runDir, nil
 }
 
 func applyFanStressDefaults(opts *FanStressOptions) {
@@ -243,9 +239,8 @@ func applyFanStressDefaults(opts *FanStressOptions) {
 	if opts.Phase2DurSec <= 0 {
 		opts.Phase2DurSec = 300
 	}
-	if opts.SizeMB <= 0 {
-		opts.SizeMB = 64
-	}
+	// SizeMB == 0 means "auto" (worker picks 95% of GPU VRAM for maximum power draw).
+	// Leave at 0 to avoid passing a too-small size that starves the tensor-core path.
 }
 
 // sampleFanStressRow collects all metrics for one telemetry sample.

audit/internal/platform/sat_test.go

@@ -14,12 +14,12 @@ import (
 func TestStorageSATCommands(t *testing.T) {
 	t.Parallel()
 
-	nvme := storageSATCommands("/dev/nvme0n1")
+	nvme := storageSATCommands("/dev/nvme0n1", false)
 	if len(nvme) != 3 || nvme[2].cmd[0] != "nvme" {
 		t.Fatalf("unexpected nvme commands: %#v", nvme)
 	}
 
-	sata := storageSATCommands("/dev/sda")
+	sata := storageSATCommands("/dev/sda", false)
 	if len(sata) != 2 || sata[0].cmd[0] != "smartctl" {
 		t.Fatalf("unexpected sata commands: %#v", sata)
 	}
@@ -216,6 +216,74 @@ func TestResolveDCGMGPUIndicesKeepsExplicitSelection(t *testing.T) {
 	}
 }
 
+func TestParseNvidiaGPUHealthDetectsResetRequired(t *testing.T) {
+	t.Parallel()
+
+	got := parseNvidiaGPUHealth("0, NVIDIA H100 PCIe, 38, 46.89, 0, 0, 81559\n1, NVIDIA H100 PCIe, [GPU requires reset], [N/A], [N/A], 0, 81559\n")
+	if len(got) != 2 {
+		t.Fatalf("len=%d want 2", len(got))
+	}
+	if got[0].NeedsReset {
+		t.Fatalf("gpu0 unexpectedly marked reset-required")
+	}
+	if !got[1].NeedsReset {
+		t.Fatalf("gpu1 should be marked reset-required: %#v", got[1])
+	}
+}
+
+func TestCheckNvidiaJobHealthReturnsErrorForSelectedResetRequiredGPU(t *testing.T) {
+	oldExecCommand := satExecCommand
+	satExecCommand = func(name string, args ...string) *exec.Cmd {
+		if name == "nvidia-smi" {
+			return exec.Command("sh", "-c", "printf '0, NVIDIA H100 PCIe, 38, 46.89, 0, 0, 81559\n1, NVIDIA H100 PCIe, [GPU requires reset], [N/A], [N/A], 0, 81559\n'")
+		}
+		return exec.Command(name, args...)
+	}
+	t.Cleanup(func() { satExecCommand = oldExecCommand })
+
+	msg, err := checkNvidiaJobHealth([]int{1})
+	if err == nil {
+		t.Fatal("expected health check error")
+	}
+	if !strings.Contains(msg, "gpu 1") || !strings.Contains(strings.ToLower(msg), "requires reset") {
+		t.Fatalf("unexpected message: %q", msg)
+	}
+}
+
+func TestWriteNvidiaGPUStatusFilesCreatesPerGPUFiles(t *testing.T) {
+	dir := t.TempDir()
+	oldExecCommand := satExecCommand
+	satExecCommand = func(name string, args ...string) *exec.Cmd {
+		if name == "nvidia-smi" {
+			return exec.Command("sh", "-c", "printf '0, NVIDIA H100 PCIe, 38, 46.89, 0, 0, 81559\n1, NVIDIA H100 PCIe, [GPU requires reset], [N/A], [N/A], 0, 81559\n'")
+		}
+		return exec.Command(name, args...)
+	}
+	t.Cleanup(func() { satExecCommand = oldExecCommand })
+
+	perGPU := map[int]*nvidiaGPUStatusFile{
+		0: {Index: 0, RunStatus: "OK"},
+		1: {Index: 1, RunStatus: "FAILED", FailingJob: "02-dcgmi-targeted-stress.log", Reason: "NVIDIA GPU health check failed:"},
+	}
+	if err := writeNvidiaGPUStatusFiles(dir, "FAILED", perGPU, map[int]struct{}{0: {}, 1: {}}); err != nil {
+		t.Fatalf("writeNvidiaGPUStatusFiles error: %v", err)
+	}
+	raw, err := os.ReadFile(filepath.Join(dir, "gpu-1-status.txt"))
+	if err != nil {
+		t.Fatalf("ReadFile gpu-1-status.txt: %v", err)
+	}
+	text := string(raw)
+	if !strings.Contains(text, "run_status=FAILED") {
+		t.Fatalf("missing run status:\n%s", text)
+	}
+	if !strings.Contains(text, "health_status=RESET_REQUIRED") {
+		t.Fatalf("missing health status:\n%s", text)
+	}
+	if !strings.Contains(text, "failing_job=02-dcgmi-targeted-stress.log") {
+		t.Fatalf("missing failing job:\n%s", text)
+	}
+}
+
 func TestResolveDCGMProfTesterCommandUsesVersionedBinary(t *testing.T) {
 	oldLookPath := satLookPath
 	satLookPath = func(file string) (string, error) {
@@ -341,6 +409,7 @@ func TestClassifySATResult(t *testing.T) {
 	}{
 		{name: "ok", job: "memtester", out: "done", err: nil, status: "OK"},
 		{name: "unsupported", job: "smartctl-self-test-short", out: "Self-test not supported", err: errors.New("rc 1"), status: "UNSUPPORTED"},
+		{name: "nvme wait timeout without progress", job: "nvme-device-self-test", out: "Short Device self-test started\nWaiting for self test completion...\nno progress for 78 seconds, stop waiting", err: errors.New("rc 1"), status: "UNSUPPORTED"},
 		{name: "failed", job: "bee-gpu-burn", out: "cuda error", err: errors.New("rc 1"), status: "FAILED"},
 		{name: "cuda not ready", job: "bee-gpu-burn", out: "cuInit failed: CUDA_ERROR_SYSTEM_NOT_READY", err: errors.New("rc 1"), status: "UNSUPPORTED"},
 	}

audit/internal/platform/techdump.go

@@ -20,6 +20,7 @@ var techDumpFixedCommands = []struct {
 	{Name: "dmidecode", Args: []string{"-t", "4"}, File: "dmidecode-type4.txt"},
 	{Name: "dmidecode", Args: []string{"-t", "17"}, File: "dmidecode-type17.txt"},
 	{Name: "lspci", Args: []string{"-vmm", "-D"}, File: "lspci-vmm.txt"},
+	{Name: "lspci", Args: []string{"-vvv"}, File: "lspci-vvv.txt"},
 	{Name: "lsblk", Args: []string{"-J", "-d", "-o", "NAME,TYPE,SIZE,SERIAL,MODEL,TRAN,HCTL"}, File: "lsblk.json"},
 	{Name: "sensors", Args: []string{"-j"}, File: "sensors.json"},
 	{Name: "ipmitool", Args: []string{"fru", "print"}, File: "ipmitool-fru.txt"},

audit/internal/platform/types.go

@@ -70,6 +70,7 @@ type NvidiaStressOptions struct {
 	Loader            string
 	GPUIndices        []int
 	ExcludeGPUIndices []int
+	StaggerSeconds    int
 }
 
 func New() *System {

audit/internal/schema/hardware.go

@@ -22,6 +22,10 @@ type RuntimeHealth struct {
 	CUDAReady     bool                   `json:"cuda_ready,omitempty"`
 	NvidiaGSPMode string                 `json:"nvidia_gsp_mode,omitempty"` // "gsp-on", "gsp-off", "gsp-stuck"
 	NetworkStatus string                 `json:"network_status,omitempty"`
+	// ToRAMStatus: "ok" (ISO in RAM), "warning" (toram not active), "failed" (toram active but copy failed)
+	ToRAMStatus   string `json:"toram_status,omitempty"`
+	// USBExportPath: mount point of the first writable USB drive found, empty if none.
+	USBExportPath string `json:"usb_export_path,omitempty"`
 	Issues        []RuntimeIssue         `json:"issues,omitempty"`
 	Tools         []RuntimeToolStatus    `json:"tools,omitempty"`
 	Services      []RuntimeServiceStatus `json:"services,omitempty"`
@@ -183,6 +187,13 @@ type HardwarePCIeDevice struct {
 	BatteryTemperatureC    *float64       `json:"battery_temperature_c,omitempty"`
 	BatteryVoltageV        *float64       `json:"battery_voltage_v,omitempty"`
 	BatteryReplaceRequired *bool          `json:"battery_replace_required,omitempty"`
+	SFPPresent             *bool          `json:"sfp_present,omitempty"`
+	SFPIdentifier          *string        `json:"sfp_identifier,omitempty"`
+	SFPConnector           *string        `json:"sfp_connector,omitempty"`
+	SFPVendor              *string        `json:"sfp_vendor,omitempty"`
+	SFPPartNumber          *string        `json:"sfp_part_number,omitempty"`
+	SFPSerialNumber        *string        `json:"sfp_serial_number,omitempty"`
+	SFPWavelengthNM        *float64       `json:"sfp_wavelength_nm,omitempty"`
 	SFPTemperatureC        *float64       `json:"sfp_temperature_c,omitempty"`
 	SFPTXPowerDBM          *float64       `json:"sfp_tx_power_dbm,omitempty"`
 	SFPRXPowerDBM          *float64       `json:"sfp_rx_power_dbm,omitempty"`

audit/internal/webui/api.go

@@ -28,6 +28,12 @@ var apiListNvidiaGPUs = func(a *app.App) ([]platform.NvidiaGPU, error) {
 	}
 	return a.ListNvidiaGPUs()
 }
+var apiListNvidiaGPUStatuses = func(a *app.App) ([]platform.NvidiaGPUStatus, error) {
+	if a == nil {
+		return nil, fmt.Errorf("app not configured")
+	}
+	return a.ListNvidiaGPUStatuses()
+}
 
 // ── Job ID counter ────────────────────────────────────────────────────────────
 
@@ -216,7 +222,21 @@ func formatSplitTaskName(baseName, selectionLabel string) string {
 }
 
 func buildNvidiaTaskSet(target string, priority int, createdAt time.Time, params taskParams, baseName string, appRef *app.App, idPrefix string) ([]*Task, error) {
-	if !shouldSplitHomogeneousNvidiaTarget(target) {
+	if !shouldSplitHomogeneousNvidiaTarget(target) || params.ParallelGPUs {
+		// Parallel mode (or non-splittable target): one task for all selected GPUs.
+		if params.ParallelGPUs && shouldSplitHomogeneousNvidiaTarget(target) {
+			// Resolve the selected GPU indices so ExcludeGPUIndices is applied.
+			gpus, err := apiListNvidiaGPUs(appRef)
+			if err != nil {
+				return nil, err
+			}
+			resolved, err := expandSelectedGPUIndices(gpus, params.GPUIndices, params.ExcludeGPUIndices)
+			if err != nil {
+				return nil, err
+			}
+			params.GPUIndices = resolved
+			params.ExcludeGPUIndices = nil
+		}
 		t := &Task{
 			ID:        newJobID(idPrefix),
 			Name:      baseName,
@@ -256,6 +276,53 @@ func buildNvidiaTaskSet(target string, priority int, createdAt time.Time, params
 	return tasks, nil
 }
 
+// expandSelectedGPUIndices returns the sorted list of selected GPU indices after
+// applying include/exclude filters, without splitting by model.
+func expandSelectedGPUIndices(gpus []platform.NvidiaGPU, include, exclude []int) ([]int, error) {
+	indexed := make(map[int]struct{}, len(gpus))
+	allIndices := make([]int, 0, len(gpus))
+	for _, gpu := range gpus {
+		indexed[gpu.Index] = struct{}{}
+		allIndices = append(allIndices, gpu.Index)
+	}
+	sort.Ints(allIndices)
+
+	selected := allIndices
+	if len(include) > 0 {
+		selected = make([]int, 0, len(include))
+		seen := make(map[int]struct{}, len(include))
+		for _, idx := range include {
+			if _, ok := indexed[idx]; !ok {
+				continue
+			}
+			if _, dup := seen[idx]; dup {
+				continue
+			}
+			seen[idx] = struct{}{}
+			selected = append(selected, idx)
+		}
+		sort.Ints(selected)
+	}
+	if len(exclude) > 0 {
+		skip := make(map[int]struct{}, len(exclude))
+		for _, idx := range exclude {
+			skip[idx] = struct{}{}
+		}
+		filtered := selected[:0]
+		for _, idx := range selected {
+			if _, ok := skip[idx]; ok {
+				continue
+			}
+			filtered = append(filtered, idx)
+		}
+		selected = filtered
+	}
+	if len(selected) == 0 {
+		return nil, fmt.Errorf("no NVIDIA GPUs selected")
+	}
+	return selected, nil
+}
+
 // ── SSE helpers ───────────────────────────────────────────────────────────────
 
 func sseWrite(w http.ResponseWriter, event, data string) bool {
@@ -415,12 +482,13 @@ func (h *handler) handleAPISATRun(target string) http.HandlerFunc {
 			return
 		}
 
-		var body struct {
-			Duration           int      `json:"duration"`
-			DiagLevel          int      `json:"diag_level"`
-			GPUIndices         []int    `json:"gpu_indices"`
-			ExcludeGPUIndices  []int    `json:"exclude_gpu_indices"`
-			Loader             string   `json:"loader"`
+			var body struct {
+				Duration           int      `json:"duration"`
+				StressMode         bool     `json:"stress_mode"`
+				GPUIndices         []int    `json:"gpu_indices"`
+				ExcludeGPUIndices  []int    `json:"exclude_gpu_indices"`
+				StaggerGPUStart    bool     `json:"stagger_gpu_start"`
+				Loader             string   `json:"loader"`
 			Profile            string   `json:"profile"`
 			DisplayName        string   `json:"display_name"`
 			PlatformComponents []string `json:"platform_components"`
@@ -436,12 +504,13 @@ func (h *handler) handleAPISATRun(target string) http.HandlerFunc {
 		if strings.TrimSpace(body.DisplayName) != "" {
 			name = body.DisplayName
 		}
-		params := taskParams{
-			Duration:           body.Duration,
-			DiagLevel:          body.DiagLevel,
-			GPUIndices:         body.GPUIndices,
-			ExcludeGPUIndices:  body.ExcludeGPUIndices,
-			Loader:             body.Loader,
+			params := taskParams{
+				Duration:           body.Duration,
+				StressMode:         body.StressMode,
+				GPUIndices:         body.GPUIndices,
+				ExcludeGPUIndices:  body.ExcludeGPUIndices,
+				StaggerGPUStart:    body.StaggerGPUStart,
+				Loader:             body.Loader,
 			BurnProfile:        body.Profile,
 			DisplayName:        body.DisplayName,
 			PlatformComponents: body.PlatformComponents,
@@ -788,6 +857,42 @@ func (h *handler) handleAPIGNVIDIAGPUs(w http.ResponseWriter, _ *http.Request) {
 	writeJSON(w, gpus)
 }
 
+func (h *handler) handleAPIGNVIDIAGPUStatuses(w http.ResponseWriter, _ *http.Request) {
+	if h.opts.App == nil {
+		writeError(w, http.StatusServiceUnavailable, "app not configured")
+		return
+	}
+	gpus, err := apiListNvidiaGPUStatuses(h.opts.App)
+	if err != nil {
+		writeError(w, http.StatusInternalServerError, err.Error())
+		return
+	}
+	if gpus == nil {
+		gpus = []platform.NvidiaGPUStatus{}
+	}
+	writeJSON(w, gpus)
+}
+
+func (h *handler) handleAPIGNVIDIAReset(w http.ResponseWriter, r *http.Request) {
+	if h.opts.App == nil {
+		writeError(w, http.StatusServiceUnavailable, "app not configured")
+		return
+	}
+	var req struct {
+		Index int `json:"index"`
+	}
+	if err := json.NewDecoder(r.Body).Decode(&req); err != nil {
+		writeError(w, http.StatusBadRequest, "invalid request body")
+		return
+	}
+	result, err := h.opts.App.ResetNvidiaGPU(req.Index)
+	status := "ok"
+	if err != nil {
+		status = "error"
+	}
+	writeJSON(w, map[string]string{"status": status, "output": result.Body})
+}
+
 func (h *handler) handleAPIGPUPresence(w http.ResponseWriter, r *http.Request) {
 	if h.opts.App == nil {
 		writeError(w, http.StatusServiceUnavailable, "app not configured")
@@ -1273,107 +1378,3 @@ func (h *handler) rollbackPendingNetworkChange() error {
 	return nil
 }
 
-// ── Display / Screen Resolution ───────────────────────────────────────────────
-
-type displayMode struct {
-	Output  string `json:"output"`
-	Mode    string `json:"mode"`
-	Current bool   `json:"current"`
-}
-
-type displayInfo struct {
-	Output  string        `json:"output"`
-	Modes   []displayMode `json:"modes"`
-	Current string        `json:"current"`
-}
-
-var xrandrOutputRE = regexp.MustCompile(`^(\S+)\s+connected`)
-var xrandrModeRE = regexp.MustCompile(`^\s{3}(\d+x\d+)\s`)
-var xrandrCurrentRE = regexp.MustCompile(`\*`)
-
-func parseXrandrOutput(out string) []displayInfo {
-	var infos []displayInfo
-	var cur *displayInfo
-	for _, line := range strings.Split(out, "\n") {
-		if m := xrandrOutputRE.FindStringSubmatch(line); m != nil {
-			if cur != nil {
-				infos = append(infos, *cur)
-			}
-			cur = &displayInfo{Output: m[1]}
-			continue
-		}
-		if cur == nil {
-			continue
-		}
-		if m := xrandrModeRE.FindStringSubmatch(line); m != nil {
-			isCurrent := xrandrCurrentRE.MatchString(line)
-			mode := displayMode{Output: cur.Output, Mode: m[1], Current: isCurrent}
-			cur.Modes = append(cur.Modes, mode)
-			if isCurrent {
-				cur.Current = m[1]
-			}
-		}
-	}
-	if cur != nil {
-		infos = append(infos, *cur)
-	}
-	return infos
-}
-
-func xrandrCommand(args ...string) *exec.Cmd {
-	cmd := exec.Command("xrandr", args...)
-	env := append([]string{}, os.Environ()...)
-	hasDisplay := false
-	hasXAuthority := false
-	for _, kv := range env {
-		if strings.HasPrefix(kv, "DISPLAY=") && strings.TrimPrefix(kv, "DISPLAY=") != "" {
-			hasDisplay = true
-		}
-		if strings.HasPrefix(kv, "XAUTHORITY=") && strings.TrimPrefix(kv, "XAUTHORITY=") != "" {
-			hasXAuthority = true
-		}
-	}
-	if !hasDisplay {
-		env = append(env, "DISPLAY=:0")
-	}
-	if !hasXAuthority {
-		env = append(env, "XAUTHORITY=/home/bee/.Xauthority")
-	}
-	cmd.Env = env
-	return cmd
-}
-
-func (h *handler) handleAPIDisplayResolutions(w http.ResponseWriter, _ *http.Request) {
-	out, err := xrandrCommand().Output()
-	if err != nil {
-		writeError(w, http.StatusInternalServerError, "xrandr: "+err.Error())
-		return
-	}
-	writeJSON(w, parseXrandrOutput(string(out)))
-}
-
-func (h *handler) handleAPIDisplaySet(w http.ResponseWriter, r *http.Request) {
-	var req struct {
-		Output string `json:"output"`
-		Mode   string `json:"mode"`
-	}
-	if err := json.NewDecoder(r.Body).Decode(&req); err != nil || req.Output == "" || req.Mode == "" {
-		writeError(w, http.StatusBadRequest, "output and mode are required")
-		return
-	}
-	// Validate mode looks like WxH to prevent injection
-	if !regexp.MustCompile(`^\d+x\d+$`).MatchString(req.Mode) {
-		writeError(w, http.StatusBadRequest, "invalid mode format")
-		return
-	}
-	// Validate output name (no special chars)
-	if !regexp.MustCompile(`^[A-Za-z0-9_\-]+$`).MatchString(req.Output) {
-		writeError(w, http.StatusBadRequest, "invalid output name")
-		return
-	}
-	if out, err := xrandrCommand("--output", req.Output, "--mode", req.Mode).CombinedOutput(); err != nil {
-		writeError(w, http.StatusInternalServerError, "xrandr: "+strings.TrimSpace(string(out)))
-		return
-	}
-	writeJSON(w, map[string]string{"status": "ok", "output": req.Output, "mode": req.Mode})
-}

audit/internal/webui/api_test.go

@@ -10,30 +10,6 @@ import (
 	"bee/audit/internal/platform"
 )
 
-func TestXrandrCommandAddsDefaultX11Env(t *testing.T) {
-	t.Setenv("DISPLAY", "")
-	t.Setenv("XAUTHORITY", "")
-
-	cmd := xrandrCommand("--query")
-
-	var hasDisplay bool
-	var hasXAuthority bool
-	for _, kv := range cmd.Env {
-		if kv == "DISPLAY=:0" {
-			hasDisplay = true
-		}
-		if kv == "XAUTHORITY=/home/bee/.Xauthority" {
-			hasXAuthority = true
-		}
-	}
-	if !hasDisplay {
-		t.Fatalf("DISPLAY not injected: %v", cmd.Env)
-	}
-	if !hasXAuthority {
-		t.Fatalf("XAUTHORITY not injected: %v", cmd.Env)
-	}
-}
-
 func TestHandleAPISATRunDecodesBodyWithoutContentLength(t *testing.T) {
 	globalQueue.mu.Lock()
 	originalTasks := globalQueue.tasks

audit/internal/webui/charts_svg.go

@@ -83,6 +83,10 @@ func renderMetricChartSVG(title string, labels []string, times []time.Time, data
 		}
 	}
 
+	// Downsample to at most ~1400 points (one per pixel) before building SVG.
+	times, datasets = downsampleTimeSeries(times, datasets, 1400)
+	pointCount = len(times)
+
 	statsLabel := chartStatsLabel(datasets)
 
 	legendItems := []metricChartSeries{}
@@ -196,6 +200,19 @@ func drawGPUOverviewChartSVG(title string, labels []string, times []time.Time, s
 		}
 	}
 
+	// Downsample to at most ~1400 points before building SVG.
+	{
+		datasets := make([][]float64, len(series))
+		for i := range series {
+			datasets[i] = series[i].Values
+		}
+		times, datasets = downsampleTimeSeries(times, datasets, 1400)
+		pointCount = len(times)
+		for i := range series {
+			series[i].Values = datasets[i]
+		}
+	}
+
 	scales := make([]chartScale, len(series))
 	for i := range series {
 		min, max := chartSeriesBounds(series[i].Values)
@@ -626,6 +643,87 @@ func writeTimelineBoundaries(b *strings.Builder, layout chartLayout, start, end
 	b.WriteString(`</g>` + "\n")
 }
 
+// downsampleTimeSeries reduces the time series to at most maxPts points using
+// min-max bucketing. Each bucket contributes the index of its min and max value
+// (using the first full-length dataset as the reference series). All parallel
+// datasets are sampled at those same indices so all series stay aligned.
+// If len(times) <= maxPts the inputs are returned unchanged.
+func downsampleTimeSeries(times []time.Time, datasets [][]float64, maxPts int) ([]time.Time, [][]float64) {
+	n := len(times)
+	if n <= maxPts || maxPts <= 0 {
+		return times, datasets
+	}
+	buckets := maxPts / 2
+	if buckets < 1 {
+		buckets = 1
+	}
+	// Use the first dataset that has the same length as times as the reference
+	// for deciding which two indices to keep per bucket.
+	var ref []float64
+	for _, ds := range datasets {
+		if len(ds) == n {
+			ref = ds
+			break
+		}
+	}
+	selected := make([]int, 0, maxPts)
+	bucketSize := float64(n) / float64(buckets)
+	for b := 0; b < buckets; b++ {
+		lo := int(math.Round(float64(b) * bucketSize))
+		hi := int(math.Round(float64(b+1) * bucketSize))
+		if hi > n {
+			hi = n
+		}
+		if lo >= hi {
+			continue
+		}
+		if ref == nil {
+			selected = append(selected, lo)
+			if hi-1 != lo {
+				selected = append(selected, hi-1)
+			}
+			continue
+		}
+		minIdx, maxIdx := lo, lo
+		for i := lo + 1; i < hi; i++ {
+			if ref[i] < ref[minIdx] {
+				minIdx = i
+			}
+			if ref[i] > ref[maxIdx] {
+				maxIdx = i
+			}
+		}
+		if minIdx <= maxIdx {
+			selected = append(selected, minIdx)
+			if maxIdx != minIdx {
+				selected = append(selected, maxIdx)
+			}
+		} else {
+			selected = append(selected, maxIdx)
+			if minIdx != maxIdx {
+				selected = append(selected, minIdx)
+			}
+		}
+	}
+	outTimes := make([]time.Time, len(selected))
+	for i, idx := range selected {
+		outTimes[i] = times[idx]
+	}
+	outDatasets := make([][]float64, len(datasets))
+	for d, ds := range datasets {
+		if len(ds) != n {
+			outDatasets[d] = ds
+			continue
+		}
+		out := make([]float64, len(selected))
+		for i, idx := range selected {
+			out[i] = ds[idx]
+		}
+		outDatasets[d] = out
+	}
+	return outTimes, outDatasets
+}
+
 func chartXForTime(ts, start, end time.Time, left, right int) float64 {
 	if !end.After(start) {
 		return float64(left+right) / 2

audit/internal/webui/server.go

@@ -295,13 +295,11 @@ func NewHandler(opts HandlerOptions) http.Handler {
 	// Tools
 	mux.HandleFunc("GET /api/tools/check", h.handleAPIToolsCheck)
 
-	// Display
-	mux.HandleFunc("GET /api/display/resolutions", h.handleAPIDisplayResolutions)
-	mux.HandleFunc("POST /api/display/set", h.handleAPIDisplaySet)
-
 	// GPU presence / tools
 	mux.HandleFunc("GET /api/gpu/presence", h.handleAPIGPUPresence)
 	mux.HandleFunc("GET /api/gpu/nvidia", h.handleAPIGNVIDIAGPUs)
+	mux.HandleFunc("GET /api/gpu/nvidia-status", h.handleAPIGNVIDIAGPUStatuses)
+	mux.HandleFunc("POST /api/gpu/nvidia-reset", h.handleAPIGNVIDIAReset)
 	mux.HandleFunc("GET /api/gpu/tools", h.handleAPIGPUTools)
 
 	// System

audit/internal/webui/server_test.go

@@ -591,7 +591,7 @@ func TestTasksPageRendersOpenLinksAndPaginationControls(t *testing.T) {
 	}
 }
 
-func TestToolsPageRendersRestartGPUDriversButton(t *testing.T) {
+func TestToolsPageRendersNvidiaSelfHealSection(t *testing.T) {
 	handler := NewHandler(HandlerOptions{})
 	rec := httptest.NewRecorder()
 	handler.ServeHTTP(rec, httptest.NewRequest(http.MethodGet, "/tools", nil))
@@ -599,11 +599,20 @@ func TestToolsPageRendersRestartGPUDriversButton(t *testing.T) {
 		t.Fatalf("status=%d", rec.Code)
 	}
 	body := rec.Body.String()
+	if !strings.Contains(body, `NVIDIA Self Heal`) {
+		t.Fatalf("tools page missing nvidia self heal section: %s", body)
+	}
 	if !strings.Contains(body, `Restart GPU Drivers`) {
 		t.Fatalf("tools page missing restart gpu drivers button: %s", body)
 	}
-	if !strings.Contains(body, `restartGPUDrivers()`) {
-		t.Fatalf("tools page missing restartGPUDrivers action: %s", body)
+	if !strings.Contains(body, `nvidiaRestartDrivers()`) {
+		t.Fatalf("tools page missing nvidiaRestartDrivers action: %s", body)
+	}
+	if !strings.Contains(body, `/api/gpu/nvidia-status`) {
+		t.Fatalf("tools page missing nvidia status api usage: %s", body)
+	}
+	if !strings.Contains(body, `nvidiaResetGPU(`) {
+		t.Fatalf("tools page missing nvidiaResetGPU action: %s", body)
 	}
 	if !strings.Contains(body, `id="boot-source-text"`) {
 		t.Fatalf("tools page missing boot source field: %s", body)
@@ -684,8 +693,8 @@ func TestBenchmarkPageRendersSavedResultsTable(t *testing.T) {
 	for _, needle := range []string{
 		`Benchmark Results`,
 		`Composite score by saved benchmark run and GPU.`,
-		`NVIDIA H100 PCIe / GPU 0`,
-		`NVIDIA H100 PCIe / GPU 1`,
+		`GPU #0 — NVIDIA H100 PCIe`,
+		`GPU #1 — NVIDIA H100 PCIe`,
 		`#1`,
 		wantTime,
 		`1176.25`,
@@ -732,8 +741,8 @@ func TestBurnPageRendersGoalBasedNVIDIACards(t *testing.T) {
 	for _, needle := range []string{
 		`NVIDIA Max Compute Load`,
 		`dcgmproftester`,
-		`targeted_stress remain in <a href="/validate">Validate</a>`,
-		`NVIDIA Interconnect Test (NCCL all_reduce_perf)`,
+		`NCCL`,
+		`Validate → Stress mode`,
 		`id="burn-gpu-list"`,
 	} {
 		if !strings.Contains(body, needle) {
@@ -1085,6 +1094,7 @@ func TestDashboardRendersRuntimeHealthTable(t *testing.T) {
 	}
 	body := rec.Body.String()
 	for _, needle := range []string{
+		// Runtime Health card — LiveCD checks only
 		`Runtime Health`,
 		`<th>Check</th><th>Status</th><th>Source</th><th>Issue</th>`,
 		`Export Directory`,
@@ -1093,16 +1103,18 @@ func TestDashboardRendersRuntimeHealthTable(t *testing.T) {
 		`CUDA / ROCm`,
 		`Required Utilities`,
 		`Bee Services`,
-		`<td>CPU</td>`,
-		`<td>Memory</td>`,
-		`<td>Storage</td>`,
-		`<td>GPU</td>`,
 		`CUDA runtime is not ready for GPU SAT.`,
 		`Missing: nvidia-smi`,
 		`bee-nvidia=inactive`,
-		`cpu SAT: FAILED`,
-		`storage SAT: FAILED`,
-		`sat:nvidia`,
+		// Hardware Summary card — component health badges
+		`Hardware Summary`,
+		`>CPU<`,
+		`>Memory<`,
+		`>Storage<`,
+		`>GPU<`,
+		`>PSU<`,
+		`badge-warn`,   // cpu Warning badge
+		`badge-err`,    // storage Critical badge
 	} {
 		if !strings.Contains(body, needle) {
 			t.Fatalf("dashboard missing %q: %s", needle, body)

audit/internal/webui/tasks.go

@@ -115,10 +115,12 @@ type Task struct {
 // taskParams holds optional parameters parsed from the run request.
 type taskParams struct {
 	Duration           int      `json:"duration,omitempty"`
-	DiagLevel          int      `json:"diag_level,omitempty"`
+	StressMode         bool     `json:"stress_mode,omitempty"`
 	GPUIndices         []int    `json:"gpu_indices,omitempty"`
 	ExcludeGPUIndices  []int    `json:"exclude_gpu_indices,omitempty"`
+	StaggerGPUStart    bool     `json:"stagger_gpu_start,omitempty"`
 	SizeMB             int      `json:"size_mb,omitempty"`
+	Passes             int      `json:"passes,omitempty"`
 	Loader             string   `json:"loader,omitempty"`
 	BurnProfile        string   `json:"burn_profile,omitempty"`
 	BenchmarkProfile   string   `json:"benchmark_profile,omitempty"`
@@ -161,6 +163,13 @@ func resolveBurnPreset(profile string) burnPreset {
 	}
 }
 
+func boolToNvidiaStaggerSeconds(enabled bool, selected []int) int {
+	if enabled && len(selected) > 1 {
+		return 180
+	}
+	return 0
+}
+
 func resolvePlatformStressPreset(profile string) platform.PlatformStressOptions {
 	acceptanceCycles := []platform.PlatformStressCycle{
 		{LoadSec: 85, IdleSec: 5},
@@ -215,11 +224,11 @@ var globalQueue = &taskQueue{trigger: make(chan struct{}, 1)}
 const maxTaskHistory = 50
 
 var (
-	runMemoryAcceptancePackCtx = func(a *app.App, ctx context.Context, baseDir string, logFunc func(string)) (string, error) {
-		return a.RunMemoryAcceptancePackCtx(ctx, baseDir, logFunc)
+	runMemoryAcceptancePackCtx = func(a *app.App, ctx context.Context, baseDir string, sizeMB, passes int, logFunc func(string)) (string, error) {
+		return a.RunMemoryAcceptancePackCtx(ctx, baseDir, sizeMB, passes, logFunc)
 	}
-	runStorageAcceptancePackCtx = func(a *app.App, ctx context.Context, baseDir string, logFunc func(string)) (string, error) {
-		return a.RunStorageAcceptancePackCtx(ctx, baseDir, logFunc)
+	runStorageAcceptancePackCtx = func(a *app.App, ctx context.Context, baseDir string, extended bool, logFunc func(string)) (string, error) {
+		return a.RunStorageAcceptancePackCtx(ctx, baseDir, extended, logFunc)
 	}
 	runCPUAcceptancePackCtx = func(a *app.App, ctx context.Context, baseDir string, durationSec int, logFunc func(string)) (string, error) {
 		return a.RunCPUAcceptancePackCtx(ctx, baseDir, durationSec, logFunc)
@@ -552,7 +561,10 @@ func (q *taskQueue) runTask(t *Task, j *jobState, ctx context.Context) {
 			err = fmt.Errorf("app not configured")
 			break
 		}
-		diagLevel := t.params.DiagLevel
+		diagLevel := 2
+		if t.params.StressMode {
+			diagLevel = 3
+		}
 		if len(t.params.GPUIndices) > 0 || diagLevel > 0 {
 			result, e := a.RunNvidiaAcceptancePackWithOptions(
 				ctx, "", diagLevel, t.params.GPUIndices, j.append,
@@ -588,7 +600,7 @@ func (q *taskQueue) runTask(t *Task, j *jobState, ctx context.Context) {
 			RunNCCL:           t.params.RunNCCL,
 			ParallelGPUs:      t.params.ParallelGPUs,
 		}, j.append)
-	case "nvidia-compute":
+		case "nvidia-compute":
 		if a == nil {
 			err = fmt.Errorf("app not configured")
 			break
@@ -597,7 +609,11 @@ func (q *taskQueue) runTask(t *Task, j *jobState, ctx context.Context) {
 		if t.params.BurnProfile != "" && dur <= 0 {
 			dur = resolveBurnPreset(t.params.BurnProfile).DurationSec
 		}
-		archive, err = a.RunNvidiaOfficialComputePack(ctx, "", dur, t.params.GPUIndices, j.append)
+			staggerSec := boolToNvidiaStaggerSeconds(t.params.StaggerGPUStart, t.params.GPUIndices)
+			if staggerSec > 0 {
+				j.append(fmt.Sprintf("NVIDIA staggered ramp-up enabled: %ds per GPU", staggerSec))
+			}
+			archive, err = a.RunNvidiaOfficialComputePack(ctx, "", dur, t.params.GPUIndices, staggerSec, j.append)
 	case "nvidia-targeted-power":
 		if a == nil {
 			err = fmt.Errorf("app not configured")
@@ -647,24 +663,29 @@ func (q *taskQueue) runTask(t *Task, j *jobState, ctx context.Context) {
 		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)
+			archive, err = runNvidiaStressPackCtx(a, ctx, "", platform.NvidiaStressOptions{
+				DurationSec:       dur,
+				Loader:            t.params.Loader,
+				GPUIndices:        t.params.GPUIndices,
+				ExcludeGPUIndices: t.params.ExcludeGPUIndices,
+				StaggerSeconds:    boolToNvidiaStaggerSeconds(t.params.StaggerGPUStart, t.params.GPUIndices),
+			}, j.append)
 	case "memory":
 		if a == nil {
 			err = fmt.Errorf("app not configured")
 			break
 		}
-		archive, err = runMemoryAcceptancePackCtx(a, ctx, "", j.append)
+		sizeMB, passes := 256, 1
+		if t.params.StressMode {
+			sizeMB, passes = 1024, 3
+		}
+		archive, err = runMemoryAcceptancePackCtx(a, ctx, "", sizeMB, passes, j.append)
 	case "storage":
 		if a == nil {
 			err = fmt.Errorf("app not configured")
 			break
 		}
-		archive, err = runStorageAcceptancePackCtx(a, ctx, "", j.append)
+		archive, err = runStorageAcceptancePackCtx(a, ctx, "", t.params.StressMode, j.append)
 	case "cpu":
 		if a == nil {
 			err = fmt.Errorf("app not configured")
@@ -675,7 +696,11 @@ func (q *taskQueue) runTask(t *Task, j *jobState, ctx context.Context) {
 			dur = resolveBurnPreset(t.params.BurnProfile).DurationSec
 		}
 		if dur <= 0 {
-			dur = 60
+			if t.params.StressMode {
+				dur = 1800
+			} else {
+				dur = 60
+			}
 		}
 		j.append(fmt.Sprintf("CPU stress duration: %ds", dur))
 		archive, err = runCPUAcceptancePackCtx(a, ctx, "", dur, j.append)

audit/internal/webui/tasks_test.go

@@ -422,7 +422,7 @@ func TestWriteTaskReportArtifactsIncludesBenchmarkResultsForTask(t *testing.T) {
 	for _, needle := range []string{
 		`Benchmark Results`,
 		`Composite score for this benchmark task.`,
-		`NVIDIA H100 PCIe / GPU 0`,
+		`GPU #0 — NVIDIA H100 PCIe`,
 		`1176.25`,
 	} {
 		if !strings.Contains(html, needle) {

bible-local/docs/gpu-model-propagation.md

@@ -0,0 +1,117 @@
+# GPU Model Name Propagation
+
+How GPU model names are detected, stored, and displayed throughout the project.
+
+---
+
+## Detection Sources
+
+There are **two separate pipelines** for GPU model names — they use different structs and don't share state.
+
+### Pipeline A — Live / SAT (nvidia-smi query at runtime)
+
+**File:** `audit/internal/platform/sat.go`
+
+- `ListNvidiaGPUs()` → `NvidiaGPU.Name` (field: `name`, from `nvidia-smi --query-gpu=index,name,...`)
+- `ListNvidiaGPUStatuses()` → `NvidiaGPUStatus.Name`
+- Used by: GPU selection UI, live metrics labels, burn/stress test logic
+
+### Pipeline B — Benchmark results
+
+**File:** `audit/internal/platform/benchmark.go`, line 124
+
+- `queryBenchmarkGPUInfo(selected)` → `benchmarkGPUInfo.Name`
+- Stored in `BenchmarkGPUResult.Name` (`json:"name,omitempty"`)
+- Used by: benchmark history table, benchmark report
+
+### Pipeline C — Hardware audit JSON (PCIe schema)
+
+**File:** `audit/internal/schema/hardware.go`
+
+- `HardwarePCIeDevice.Model *string` (field name is **Model**, not Name)
+- For AMD GPUs: populated by `audit/internal/collector/amdgpu.go` from `info.Product`
+- For NVIDIA GPUs: **NOT populated** by `audit/internal/collector/nvidia.go` — the NVIDIA enricher sets telemetry/status but skips the Model field
+- Used by: hardware summary page (`hwDescribeGPU` in `pages.go:487`)
+
+---
+
+## Key Inconsistency: NVIDIA PCIe Model is Never Set
+
+`audit/internal/collector/nvidia.go` — `enrichPCIeWithNVIDIAData()` enriches NVIDIA PCIe devices with telemetry and status but does **not** populate `HardwarePCIeDevice.Model`.
+
+This means:
+- Hardware summary page shows "Unknown GPU" for all NVIDIA devices (falls back at `pages.go:486`)
+- AMD GPUs do have their model populated
+
+The fix would be: copy `gpu.Name` from the SAT pipeline into `dev.Model` inside `enrichPCIeWithNVIDIAData`.
+
+---
+
+## Benchmark History "Unknown GPU" Issue
+
+**Symptom:** Benchmark history table shows "GPU #N — Unknown GPU" columns instead of real GPU model names.
+
+**Root cause:** `BenchmarkGPUResult.Name` has tag `json:"name,omitempty"`. If `queryBenchmarkGPUInfo()` fails (warns at `benchmark.go:126`) or returns empty names, the Name field is never set and is omitted from JSON. Loaded results have empty Name → falls back to "Unknown GPU" at `pages.go:2226, 2237`.
+
+This happens for:
+- Older result files saved before the `Name` field was added
+- Runs where nvidia-smi query failed before the benchmark started
+
+---
+
+## Fallback Strings — Current State
+
+| Location | File | Fallback string |
+|---|---|---|
+| Hardware summary (PCIe) | `pages.go:486` | `"Unknown GPU"` |
+| Benchmark report summary | `benchmark_report.go:43` | `"Unknown GPU"` |
+| Benchmark report scorecard | `benchmark_report.go:93` | `"Unknown"` ← inconsistent |
+| Benchmark report detail | `benchmark_report.go:122` | `"Unknown GPU"` |
+| Benchmark history per-GPU col | `pages.go:2226` | `"Unknown GPU"` |
+| Benchmark history parallel col | `pages.go:2237` | `"Unknown GPU"` |
+| SAT status file write | `sat.go:922` | `"unknown"` ← lowercase, inconsistent |
+| GPU selection API | `api.go:163` | `"GPU N"` (no "Unknown") |
+
+**Rule:** all UI fallbacks should use `"Unknown GPU"`. The two outliers are `benchmark_report.go:93` (`"Unknown"`) and `sat.go:922` (`"unknown"`).
+
+---
+
+## GPU Selection UI
+
+**File:** `audit/internal/webui/pages.go`
+
+- Source: `GET /api/gpus` → `api.go` → `ListNvidiaGPUs()` → live nvidia-smi
+- Render: `'GPU ' + gpu.index + ' — ' + gpu.name + ' · ' + mem`
+- Fallback: `gpu.name || 'GPU ' + idx` (JS, line ~1432)
+
+This always shows the correct model because it queries nvidia-smi live. It is **not** connected to benchmark result data.
+
+---
+
+## Data Flow Summary
+
+```
+nvidia-smi (live)
+  └─ ListNvidiaGPUs() → NvidiaGPU.Name
+       ├─ GPU selection UI (always correct)
+       ├─ Live metrics labels (charts_svg.go)
+       └─ SAT/burn status file (sat.go)
+
+nvidia-smi (at benchmark start)
+  └─ queryBenchmarkGPUInfo() → benchmarkGPUInfo.Name
+       └─ BenchmarkGPUResult.Name (json:"name,omitempty")
+            ├─ Benchmark report
+            └─ Benchmark history table columns
+
+nvidia-smi / lspci (audit collection)
+  └─ HardwarePCIeDevice.Model (NVIDIA: NOT populated; AMD: populated)
+       └─ Hardware summary page hwDescribeGPU()
+```
+
+---
+
+## What Needs Fixing
+
+1. **NVIDIA PCIe Model** — `enrichPCIeWithNVIDIAData()` should set `dev.Model = &gpu.Name`
+2. **Fallback consistency** — `benchmark_report.go:93` should say `"Unknown GPU"` not `"Unknown"`; `sat.go:922` should say `"Unknown GPU"` not `"unknown"`
+3. **Old benchmark JSONs** — no fix possible for already-saved results with missing names (display-only issue)

iso/builder/bee-gpu-stress.c

@@ -36,7 +36,6 @@ typedef void *CUstream;
 #define MAX_CUBLAS_PROFILES 5
 #define MIN_PROFILE_BUDGET_BYTES ((size_t)4u * 1024u * 1024u)
 #define MIN_STREAM_BUDGET_BYTES ((size_t)64u * 1024u * 1024u)
-#define STRESS_LAUNCH_DEPTH 8
 
 static const char *ptx_source =
     ".version 6.0\n"
@@ -344,7 +343,6 @@ static int run_ptx_fallback(struct cuda_api *api,
     unsigned long iterations = 0;
     int mp_count = 0;
     int stream_count = 1;
-    int launches_per_wave = 0;
 
     memset(report, 0, sizeof(*report));
     snprintf(report->backend, sizeof(report->backend), "driver-ptx");
@@ -419,44 +417,42 @@ static int run_ptx_fallback(struct cuda_api *api,
 
     unsigned int threads = 256;
 
-    double start = now_seconds();
-    double deadline = start + (double)seconds;
+    double deadline = now_seconds() + (double)seconds;
+    double next_sync = now_seconds() + 1.0;
     while (now_seconds() < deadline) {
-        launches_per_wave = 0;
-        for (int depth = 0; depth < STRESS_LAUNCH_DEPTH && now_seconds() < deadline; depth++) {
-            int launched_this_batch = 0;
-            for (int lane = 0; lane < stream_count; lane++) {
-                unsigned int blocks = (unsigned int)((words[lane] + threads - 1) / threads);
-                if (!check_rc(api,
-                              "cuLaunchKernel",
-                              api->cuLaunchKernel(kernel,
-                                                  blocks,
-                                                  1,
-                                                  1,
-                                                  threads,
-                                                  1,
-                                                  1,
-                                                  0,
-                                                  streams[lane],
-                                                  params[lane],
-                                                  NULL))) {
-                    goto fail;
-                }
-                launches_per_wave++;
-                launched_this_batch++;
-            }
-            if (launched_this_batch <= 0) {
-                break;
+        int launched = 0;
+        for (int lane = 0; lane < stream_count; lane++) {
+            unsigned int blocks = (unsigned int)((words[lane] + threads - 1) / threads);
+            if (!check_rc(api,
+                          "cuLaunchKernel",
+                          api->cuLaunchKernel(kernel,
+                                              blocks,
+                                              1,
+                                              1,
+                                              threads,
+                                              1,
+                                              1,
+                                              0,
+                                              streams[lane],
+                                              params[lane],
+                                              NULL))) {
+                goto fail;
             }
+            launched++;
+            iterations++;
         }
-        if (launches_per_wave <= 0) {
+        if (launched <= 0) {
             goto fail;
         }
-        if (!check_rc(api, "cuCtxSynchronize", api->cuCtxSynchronize())) {
-            goto fail;
+        double now = now_seconds();
+        if (now >= next_sync || now >= deadline) {
+            if (!check_rc(api, "cuCtxSynchronize", api->cuCtxSynchronize())) {
+                goto fail;
+            }
+            next_sync = now + 1.0;
         }
-        iterations += (unsigned long)launches_per_wave;
     }
+    api->cuCtxSynchronize();
 
     if (!check_rc(api, "cuMemcpyDtoH", api->cuMemcpyDtoH(sample, device_mem[0], sizeof(sample)))) {
         goto fail;
@@ -468,11 +464,10 @@ static int run_ptx_fallback(struct cuda_api *api,
     report->iterations = iterations;
     snprintf(report->details,
              sizeof(report->details),
-             "fallback_int32=OK requested_mb=%d actual_mb=%d streams=%d queue_depth=%d per_stream_mb=%zu iterations=%lu\n",
+             "fallback_int32=OK requested_mb=%d actual_mb=%d streams=%d per_stream_mb=%zu iterations=%lu\n",
              size_mb,
              report->buffer_mb,
              report->stream_count,
-             STRESS_LAUNCH_DEPTH,
              bytes_per_stream[0] / (1024u * 1024u),
              iterations);
 
@@ -1140,7 +1135,6 @@ static int run_cublaslt_stress(struct cuda_api *cuda,
     int stream_count = 1;
     int profile_count = (int)(sizeof(k_profiles) / sizeof(k_profiles[0]));
     int prepared_count = 0;
-    int wave_launches = 0;
     size_t requested_budget = 0;
     size_t total_budget = 0;
     size_t per_profile_budget = 0;
@@ -1207,11 +1201,10 @@ static int run_cublaslt_stress(struct cuda_api *cuda,
     report->buffer_mb = (int)(total_budget / (1024u * 1024u));
     append_detail(report->details,
                   sizeof(report->details),
-                  "requested_mb=%d actual_mb=%d streams=%d queue_depth=%d mp_count=%d per_worker_mb=%zu\n",
+                  "requested_mb=%d actual_mb=%d streams=%d mp_count=%d per_worker_mb=%zu\n",
                   size_mb,
                   report->buffer_mb,
                   report->stream_count,
-                  STRESS_LAUNCH_DEPTH,
                   mp_count,
                   per_profile_budget / (1024u * 1024u));
 
@@ -1260,50 +1253,55 @@ static int run_cublaslt_stress(struct cuda_api *cuda,
         return 0;
     }
 
+    /* Keep the GPU queue continuously full by submitting kernels without
+     * synchronizing after every wave.  A sync barrier after each small batch
+     * creates CPU↔GPU ping-pong gaps that prevent full TDP utilisation,
+     * especially when individual kernels are short.  Instead we sync at most
+     * once per second (for error detection) and once at the very end. */
     double deadline = now_seconds() + (double)seconds;
+    double next_sync = now_seconds() + 1.0;
     while (now_seconds() < deadline) {
-        wave_launches = 0;
-        for (int depth = 0; depth < STRESS_LAUNCH_DEPTH && now_seconds() < deadline; depth++) {
-            int launched_this_batch = 0;
-            for (int i = 0; i < prepared_count; i++) {
-                if (!prepared[i].ready) {
-                    continue;
-                }
-                if (!run_cublas_profile(handle, &cublas, &prepared[i])) {
-                    append_detail(report->details,
-                                  sizeof(report->details),
-                                  "%s=FAILED runtime\n",
-                                  prepared[i].desc.name);
-                    for (int j = 0; j < prepared_count; j++) {
-                        destroy_profile(&cublas, cuda, &prepared[j]);
-                    }
-                    cublas.cublasLtDestroy(handle);
-                    destroy_streams(cuda, streams, stream_count);
-                    cuda->cuCtxDestroy(ctx);
-                    return 0;
-                }
-                prepared[i].iterations++;
-                report->iterations++;
-                wave_launches++;
-                launched_this_batch++;
+        int launched = 0;
+        for (int i = 0; i < prepared_count; i++) {
+            if (!prepared[i].ready) {
+                continue;
             }
-            if (launched_this_batch <= 0) {
-                break;
+            if (!run_cublas_profile(handle, &cublas, &prepared[i])) {
+                append_detail(report->details,
+                              sizeof(report->details),
+                              "%s=FAILED runtime\n",
+                              prepared[i].desc.name);
+                for (int j = 0; j < prepared_count; j++) {
+                    destroy_profile(&cublas, cuda, &prepared[j]);
+                }
+                cublas.cublasLtDestroy(handle);
+                destroy_streams(cuda, streams, stream_count);
+                cuda->cuCtxDestroy(ctx);
+                return 0;
             }
+            prepared[i].iterations++;
+            report->iterations++;
+            launched++;
         }
-        if (wave_launches <= 0) {
+        if (launched <= 0) {
             break;
         }
-        if (!check_rc(cuda, "cuCtxSynchronize", cuda->cuCtxSynchronize())) {
-            for (int i = 0; i < prepared_count; i++) {
-                destroy_profile(&cublas, cuda, &prepared[i]);
+        double now = now_seconds();
+        if (now >= next_sync || now >= deadline) {
+            if (!check_rc(cuda, "cuCtxSynchronize", cuda->cuCtxSynchronize())) {
+                for (int i = 0; i < prepared_count; i++) {
+                    destroy_profile(&cublas, cuda, &prepared[i]);
+                }
+                cublas.cublasLtDestroy(handle);
+                destroy_streams(cuda, streams, stream_count);
+                cuda->cuCtxDestroy(ctx);
+                return 0;
             }
-            cublas.cublasLtDestroy(handle);
-            destroy_streams(cuda, streams, stream_count);
-            cuda->cuCtxDestroy(ctx);
-            return 0;
+            next_sync = now + 1.0;
         }
     }
+    /* Final drain — ensure all queued work finishes before we read results. */
+    cuda->cuCtxSynchronize();
 
     for (int i = 0; i < prepared_count; i++) {
         if (!prepared[i].ready) {

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

@@ -11,18 +11,18 @@ echo "  Hardware Audit LiveCD"
 echo ""
 
 menuentry "EASY-BEE" {
-    linux   @KERNEL_LIVE@ @APPEND_LIVE@ nomodeset bee.nvidia.mode=normal net.ifnames=0 biosdevname=0 mitigations=off transparent_hugepage=always numa_balancing=disable nowatchdog nosoftlockup
+    linux   @KERNEL_LIVE@ @APPEND_LIVE@ nomodeset bee.nvidia.mode=normal net.ifnames=0 biosdevname=0 mitigations=off transparent_hugepage=always numa_balancing=disable pcie_aspm=off intel_idle.max_cstate=1 processor.max_cstate=1 nowatchdog nosoftlockup
     initrd  @INITRD_LIVE@
 }
 
 submenu "EASY-BEE (advanced options) -->" {
     menuentry "EASY-BEE — GSP=off" {
-        linux   @KERNEL_LIVE@ @APPEND_LIVE@ nomodeset bee.nvidia.mode=gsp-off net.ifnames=0 biosdevname=0 mitigations=off transparent_hugepage=always numa_balancing=disable nowatchdog nosoftlockup
+        linux   @KERNEL_LIVE@ @APPEND_LIVE@ nomodeset bee.nvidia.mode=gsp-off net.ifnames=0 biosdevname=0 mitigations=off transparent_hugepage=always numa_balancing=disable pcie_aspm=off intel_idle.max_cstate=1 processor.max_cstate=1 nowatchdog nosoftlockup
         initrd  @INITRD_LIVE@
     }
 
     menuentry "EASY-BEE — KMS (no nomodeset)" {
-        linux   @KERNEL_LIVE@ @APPEND_LIVE@ bee.nvidia.mode=normal net.ifnames=0 biosdevname=0 mitigations=off transparent_hugepage=always numa_balancing=disable nowatchdog nosoftlockup
+        linux   @KERNEL_LIVE@ @APPEND_LIVE@ bee.nvidia.mode=normal net.ifnames=0 biosdevname=0 mitigations=off transparent_hugepage=always numa_balancing=disable pcie_aspm=off intel_idle.max_cstate=1 processor.max_cstate=1 nowatchdog nosoftlockup
         initrd  @INITRD_LIVE@
     }
 

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

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

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

@@ -3,31 +3,31 @@ label live-@FLAVOUR@-normal
     menu default
     linux @LINUX@
     initrd @INITRD@
-    append @APPEND_LIVE@ bee.nvidia.mode=normal
+    append @APPEND_LIVE@ bee.nvidia.mode=normal pcie_aspm=off intel_idle.max_cstate=1 processor.max_cstate=1
 
 label live-@FLAVOUR@-kms
     menu label EASY-BEE (^graphics/KMS)
     linux @LINUX@
     initrd @INITRD@
-    append @APPEND_LIVE@ bee.display=kms bee.nvidia.mode=normal
+    append @APPEND_LIVE@ bee.display=kms bee.nvidia.mode=normal pcie_aspm=off intel_idle.max_cstate=1 processor.max_cstate=1
 
 label live-@FLAVOUR@-toram
     menu label EASY-BEE (^load to RAM)
     linux @LINUX@
     initrd @INITRD@
-    append @APPEND_LIVE@ toram bee.nvidia.mode=normal
+    append @APPEND_LIVE@ toram bee.nvidia.mode=normal pcie_aspm=off intel_idle.max_cstate=1 processor.max_cstate=1
 
 label live-@FLAVOUR@-gsp-off
     menu label EASY-BEE (^NVIDIA GSP=off)
     linux @LINUX@
     initrd @INITRD@
-    append @APPEND_LIVE@ nomodeset bee.nvidia.mode=gsp-off
+    append @APPEND_LIVE@ nomodeset bee.nvidia.mode=gsp-off pcie_aspm=off intel_idle.max_cstate=1 processor.max_cstate=1
 
 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
+    append @APPEND_LIVE@ bee.display=kms bee.nvidia.mode=gsp-off pcie_aspm=off intel_idle.max_cstate=1 processor.max_cstate=1
 
 label live-@FLAVOUR@-failsafe
     menu label EASY-BEE (^fail-safe)

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

@@ -25,6 +25,7 @@ ensure_bee_console_user() {
 ensure_bee_console_user
 
 # Enable common bee services
+systemctl enable bee-hpc-tuning.service
 systemctl enable bee-network.service
 systemctl enable bee-preflight.service
 systemctl enable bee-audit.service
@@ -55,6 +56,7 @@ fi
 # nogpu: no GPU services needed
 
 # Ensure scripts are executable
+chmod +x /usr/local/bin/bee-hpc-tuning  2>/dev/null || true
 chmod +x /usr/local/bin/bee-network.sh  2>/dev/null || true
 chmod +x /usr/local/bin/bee-sshsetup   2>/dev/null || true
 chmod +x /usr/local/bin/bee-smoketest  2>/dev/null || true

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

@@ -10,20 +10,15 @@ import os
 
 W, H = 1920, 1080
 
-GLYPHS = {
-    'E': ["11111", "10000", "11110", "10000", "10000", "10000", "11111"],
-    'A': ["01110", "10001", "10001", "11111", "10001", "10001", "10001"],
-    'S': ["01111", "10000", "10000", "01110", "00001", "00001", "11110"],
-    'Y': ["10001", "10001", "01010", "00100", "00100", "00100", "00100"],
-    'B': ["11110", "10001", "10001", "11110", "10001", "10001", "11110"],
-    '-': ["00000", "00000", "11111", "00000", "00000", "00000", "00000"],
-}
-
-TITLE = "EASY-BEE"
-SUBTITLE = "Hardware Audit LiveCD"
-CELL = 30
-GLYPH_GAP = 18
-ROW_GAP = 6
+ASCII_ART = [
+    "  ███████╗ █████╗ ███████╗██╗   ██╗      ██████╗ ███████╗███████╗",
+    "  ██╔════╝██╔══██╗██╔════╝╚██╗ ██╔╝      ██╔══██╗██╔════╝██╔════╝",
+    "  █████╗  ███████║███████╗ ╚████╔╝ █████╗██████╔╝█████╗  █████╗",
+    "  ██╔══╝  ██╔══██║╚════██║  ╚██╔╝  ╚════╝██╔══██╗██╔══╝  ██╔══╝",
+    "  ███████╗██║  ██║███████║   ██║         ██████╔╝███████╗███████╗",
+    "  ╚══════╝╚═╝  ╚═╝╚══════╝   ╚═╝         ╚═════╝ ╚══════╝╚══════╝",
+]
+SUBTITLE = "  Hardware Audit LiveCD"
 
 FG = (0xF6, 0xD0, 0x47)
 FG_DIM = (0xD4, 0xA9, 0x1C)
@@ -31,6 +26,12 @@ SHADOW = (0x5E, 0x47, 0x05)
 SUB = (0x96, 0x7A, 0x17)
 BG = (0x05, 0x05, 0x05)
 
+MONO_FONT_CANDIDATES = [
+    '/usr/share/fonts/truetype/dejavu/DejaVuSansMono-Bold.ttf',
+    '/usr/share/fonts/truetype/liberation2/LiberationMono-Bold.ttf',
+    '/usr/share/fonts/truetype/liberation/LiberationMono-Bold.ttf',
+    '/usr/share/fonts/truetype/freefont/FreeMonoBold.ttf',
+]
 SUB_FONT_CANDIDATES = [
     '/usr/share/fonts/truetype/dejavu/DejaVuSans-Bold.ttf',
     '/usr/share/fonts/truetype/liberation2/LiberationSans-Bold.ttf',
@@ -39,43 +40,34 @@ SUB_FONT_CANDIDATES = [
 ]
 
 
-def load_font(size):
-    for path in SUB_FONT_CANDIDATES:
+def load_font(candidates, size):
+    for path in candidates:
         if os.path.exists(path):
             return ImageFont.truetype(path, size)
     return ImageFont.load_default()
 
 
-def glyph_width(ch):
-    return len(GLYPHS[ch][0])
+def mono_metrics(font):
+    probe = Image.new('L', (W, H), 0)
+    draw = ImageDraw.Draw(probe)
+    char_w = int(round(draw.textlength("M", font=font)))
+    bb = draw.textbbox((0, 0), "Mg", font=font)
+    char_h = bb[3] - bb[1]
+    return char_w, char_h
 
 
-def render_logo_mask():
-    width_cells = 0
-    for idx, ch in enumerate(TITLE):
-        width_cells += glyph_width(ch)
-        if idx != len(TITLE) - 1:
-            width_cells += 1
-    mask_w = width_cells * CELL + (len(TITLE) - 1) * GLYPH_GAP
-    mask_h = 7 * CELL + 6 * ROW_GAP
-    mask = Image.new('L', (mask_w, mask_h), 0)
+def render_ascii_mask(font, lines, char_w, char_h, line_gap):
+    width = max(len(line) for line in lines) * char_w
+    height = len(lines) * char_h + line_gap * (len(lines) - 1)
+    mask = Image.new('L', (width, height), 0)
     draw = ImageDraw.Draw(mask)
-
-    cx = 0
-    for idx, ch in enumerate(TITLE):
-        glyph = GLYPHS[ch]
-        for row_idx, row in enumerate(glyph):
-            for col_idx, cell in enumerate(row):
-                if cell != '1':
-                    continue
-                x0 = cx + col_idx * CELL
-                y0 = row_idx * (CELL + ROW_GAP)
-                x1 = x0 + CELL - 4
-                y1 = y0 + CELL - 4
-                draw.rounded_rectangle((x0, y0, x1, y1), radius=4, fill=255)
-        cx += glyph_width(ch) * CELL
-        if idx != len(TITLE) - 1:
-            cx += CELL + GLYPH_GAP
+    for row, line in enumerate(lines):
+        y = row * (char_h + line_gap)
+        for col, ch in enumerate(line):
+            if ch == ' ':
+                continue
+            x = col * char_w
+            draw.text((x, y), ch, font=font, fill=255)
     return mask
 
 
@@ -90,20 +82,28 @@ glow_draw.ellipse((520, 340, 1400, 760), fill=(255, 190, 40, 36))
 glow = glow.filter(ImageFilter.GaussianBlur(60))
 img = Image.alpha_composite(img.convert('RGBA'), glow)
 
-logo_mask = render_logo_mask()
+TARGET_LOGO_W = 400
+max_chars = max(len(line) for line in ASCII_ART)
+_probe_font = load_font(MONO_FONT_CANDIDATES, 64)
+_probe_cw, _ = mono_metrics(_probe_font)
+font_size_logo = max(6, int(64 * TARGET_LOGO_W / (_probe_cw * max_chars)))
+font_logo = load_font(MONO_FONT_CANDIDATES, font_size_logo)
+char_w, char_h = mono_metrics(font_logo)
+logo_mask = render_ascii_mask(font_logo, ASCII_ART, char_w, char_h, 2)
 logo_w, logo_h = logo_mask.size
 logo_x = (W - logo_w) // 2
-logo_y = 290
+logo_y = 380
 
-shadow_mask = logo_mask.filter(ImageFilter.GaussianBlur(2))
-img.paste(SHADOW, (logo_x + 16, logo_y + 14), shadow_mask)
-img.paste(FG_DIM, (logo_x + 8, logo_y + 7), logo_mask)
+sh_off = max(1, font_size_logo // 6)
+shadow_mask = logo_mask.filter(ImageFilter.GaussianBlur(1))
+img.paste(SHADOW, (logo_x + sh_off * 2, logo_y + sh_off * 2), shadow_mask)
+img.paste(FG_DIM, (logo_x + sh_off, logo_y + sh_off), logo_mask)
 img.paste(FG, (logo_x, logo_y), logo_mask)
 
-font_sub = load_font(30)
+font_sub = load_font(SUB_FONT_CANDIDATES, 30)
 sub_bb = draw.textbbox((0, 0), SUBTITLE, font=font_sub)
 sub_x = (W - (sub_bb[2] - sub_bb[0])) // 2
-sub_y = logo_y + logo_h + 54
+sub_y = logo_y + logo_h + 48
 draw = ImageDraw.Draw(img)
 draw.text((sub_x + 2, sub_y + 2), SUBTITLE, font=font_sub, fill=(35, 28, 6))
 draw.text((sub_x, sub_y), SUBTITLE, font=font_sub, fill=SUB)

iso/overlay/etc/systemd/system/bee-hpc-tuning.service

@@ -0,0 +1,14 @@
+[Unit]
+Description=Bee: HPC tuning (CPU governor, C-states)
+After=local-fs.target
+Before=bee-nvidia.service bee-audit.service
+
+[Service]
+Type=oneshot
+ExecStart=/usr/local/bin/bee-log-run /appdata/bee/export/bee-hpc-tuning.log /usr/local/bin/bee-hpc-tuning
+StandardOutput=journal
+StandardError=journal
+RemainAfterExit=yes
+
+[Install]
+WantedBy=multi-user.target

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

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

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

@@ -2,13 +2,14 @@
 set -eu
 
 SECONDS=5
+STAGGER_SECONDS=0
 SIZE_MB=0
 DEVICES=""
 EXCLUDE=""
 WORKER="/usr/local/lib/bee/bee-gpu-burn-worker"
 
 usage() {
-    echo "usage: $0 [--seconds N] [--size-mb N] [--devices 0,1] [--exclude 2,3]" >&2
+    echo "usage: $0 [--seconds N] [--stagger-seconds N] [--size-mb N] [--devices 0,1] [--exclude 2,3]" >&2
     exit 2
 }
 
@@ -25,6 +26,7 @@ contains_csv() {
 while [ "$#" -gt 0 ]; do
     case "$1" in
         --seconds|-t) [ "$#" -ge 2 ] || usage; SECONDS="$2"; shift 2 ;;
+        --stagger-seconds) [ "$#" -ge 2 ] || usage; STAGGER_SECONDS="$2"; shift 2 ;;
         --size-mb|-m) [ "$#" -ge 2 ] || usage; SIZE_MB="$2"; shift 2 ;;
         --devices) [ "$#" -ge 2 ] || usage; DEVICES="$2"; shift 2 ;;
         --exclude) [ "$#" -ge 2 ] || usage; EXCLUDE="$2"; shift 2 ;;
@@ -61,14 +63,18 @@ done
 
 echo "loader=bee-gpu-burn"
 echo "selected_gpus=${FINAL}"
+echo "stagger_seconds=${STAGGER_SECONDS}"
 
 export CUDA_DEVICE_ORDER="PCI_BUS_ID"
 
 TMP_DIR=$(mktemp -d)
 trap 'rm -rf "${TMP_DIR}"' EXIT INT TERM
 
+GPU_COUNT=$(echo "${FINAL}" | tr ',' '\n' | awk 'NF' | wc -l | tr -d '[:space:]')
+gpu_pos=0
 WORKERS=""
 for id in $(echo "${FINAL}" | tr ',' ' '); do
+    gpu_pos=$((gpu_pos + 1))
     log="${TMP_DIR}/gpu-${id}.log"
     gpu_size_mb="${SIZE_MB}"
     if [ "${gpu_size_mb}" -le 0 ] 2>/dev/null; then
@@ -79,11 +85,16 @@ for id in $(echo "${FINAL}" | tr ',' ' '); do
             gpu_size_mb=512
         fi
     fi
-    echo "starting gpu ${id} size=${gpu_size_mb}MB"
+    extra_sec=$(( STAGGER_SECONDS * (GPU_COUNT - gpu_pos) ))
+    gpu_seconds=$(( SECONDS + extra_sec ))
+    echo "starting gpu ${id} size=${gpu_size_mb}MB seconds=${gpu_seconds}"
     CUDA_VISIBLE_DEVICES="${id}" \
-        "${WORKER}" --device 0 --seconds "${SECONDS}" --size-mb "${gpu_size_mb}" >"${log}" 2>&1 &
+        "${WORKER}" --device 0 --seconds "${gpu_seconds}" --size-mb "${gpu_size_mb}" >"${log}" 2>&1 &
     pid=$!
     WORKERS="${WORKERS} ${pid}:${id}:${log}"
+    if [ "${STAGGER_SECONDS}" -gt 0 ] && [ "${gpu_pos}" -lt "${GPU_COUNT}" ]; then
+        sleep "${STAGGER_SECONDS}"
+    fi
 done
 
 status=0

iso/overlay/usr/local/bin/bee-hpc-tuning

@@ -0,0 +1,41 @@
+#!/bin/sh
+# bee-hpc-tuning — apply HPC tuning for deterministic benchmarking
+# Called by bee-hpc-tuning.service at boot.
+
+log() { echo "[bee-hpc-tuning] $*"; }
+
+# ── CPU governor ────────────────────────────────────────────────────────────
+# Set all CPU cores to performance governor via sysfs.
+# cpupower is not available; write directly to scaling_governor.
+governor_ok=0
+governor_fail=0
+for gov_path in /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor; do
+    [ -f "$gov_path" ] || continue
+    if echo performance > "$gov_path" 2>/dev/null; then
+        governor_ok=$((governor_ok + 1))
+    else
+        governor_fail=$((governor_fail + 1))
+    fi
+done
+
+if [ "$governor_ok" -gt 0 ] && [ "$governor_fail" -eq 0 ]; then
+    log "CPU governor set to performance on ${governor_ok} core(s)"
+elif [ "$governor_ok" -gt 0 ]; then
+    log "WARN: CPU governor: ${governor_ok} OK, ${governor_fail} failed"
+elif [ "$governor_fail" -gt 0 ]; then
+    log "WARN: failed to set CPU governor on ${governor_fail} core(s)"
+else
+    log "WARN: no cpufreq scaling_governor paths found (C-state governor or HW-controlled)"
+fi
+
+# ── Transparent Huge Pages ───────────────────────────────────────────────────
+# Kernel cmdline sets transparent_hugepage=always at boot, but confirm and log.
+thp_path=/sys/kernel/mm/transparent_hugepage/enabled
+if [ -f "$thp_path" ]; then
+    current=$(cat "$thp_path" 2>/dev/null)
+    log "transparent_hugepage: ${current}"
+else
+    log "WARN: transparent_hugepage sysfs path not found"
+fi
+
+log "done"

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

@@ -2,6 +2,7 @@
 set -eu
 
 DURATION_SEC=300
+STAGGER_SECONDS=0
 DEVICES=""
 EXCLUDE=""
 FORMAT=""
@@ -12,7 +13,7 @@ export OCL_ICD_VENDORS="/etc/OpenCL/vendors"
 export LD_LIBRARY_PATH="/usr/lib:/usr/local/lib${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}"
 
 usage() {
-    echo "usage: $0 [--seconds N] [--devices 0,1] [--exclude 2,3] [--format name]" >&2
+    echo "usage: $0 [--seconds N] [--stagger-seconds N] [--devices 0,1] [--exclude 2,3] [--format name]" >&2
     exit 2
 }
 
@@ -118,6 +119,7 @@ ensure_opencl_ready() {
 while [ "$#" -gt 0 ]; do
     case "$1" in
         --seconds|-t) [ "$#" -ge 2 ] || usage; DURATION_SEC="$2"; shift 2 ;;
+        --stagger-seconds) [ "$#" -ge 2 ] || usage; STAGGER_SECONDS="$2"; shift 2 ;;
         --devices) [ "$#" -ge 2 ] || usage; DEVICES="$2"; shift 2 ;;
         --exclude) [ "$#" -ge 2 ] || usage; EXCLUDE="$2"; shift 2 ;;
         --format) [ "$#" -ge 2 ] || usage; FORMAT="$2"; shift 2 ;;
@@ -170,6 +172,7 @@ done
 echo "loader=john"
 echo "selected_gpus=${FINAL}"
 echo "john_devices=${JOHN_DEVICES}"
+echo "stagger_seconds=${STAGGER_SECONDS}"
 
 cd "${JOHN_DIR}"
 
@@ -232,14 +235,21 @@ trap cleanup EXIT INT TERM
 echo "format=${CHOSEN_FORMAT}"
 echo "target_seconds=${DURATION_SEC}"
 echo "slice_seconds=${TEST_SLICE_SECONDS}"
-DEADLINE=$(( $(date +%s) + DURATION_SEC ))
+TOTAL_DEVICES=$(echo "${JOHN_DEVICES}" | tr ',' '\n' | awk 'NF' | wc -l | tr -d '[:space:]')
 _first=1
+pos=0
 for opencl_id in $(echo "${JOHN_DEVICES}" | tr ',' ' '); do
+    pos=$((pos + 1))
     [ "${_first}" = "1" ] || sleep 3
     _first=0
-    run_john_loop "${opencl_id}" "${DEADLINE}" &
+    extra_sec=$(( STAGGER_SECONDS * (TOTAL_DEVICES - pos) ))
+    deadline=$(( $(date +%s) + DURATION_SEC + extra_sec ))
+    run_john_loop "${opencl_id}" "${deadline}" &
     pid=$!
     PIDS="${PIDS} ${pid}"
+    if [ "${STAGGER_SECONDS}" -gt 0 ] && [ "${pos}" -lt "${TOTAL_DEVICES}" ]; then
+        sleep "${STAGGER_SECONDS}"
+    fi
 done
 FAIL=0
 for pid in ${PIDS}; do

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

@@ -21,8 +21,13 @@ read_nvidia_modules_flavor() {
 
 log "kernel: $(uname -r)"
 
-# Skip if no NVIDIA GPU present (PCI vendor 10de)
-if ! lspci -nn 2>/dev/null | grep -qi '10de:'; then
+# Skip if no NVIDIA display/compute GPU is present.
+# Match only display-class PCI functions (0300 VGA, 0302 3D controller) from vendor 10de.
+have_nvidia_gpu() {
+    lspci -Dn 2>/dev/null | awk '$2 ~ /^03(00|02):$/ && $3 ~ /^10de:/ { found=1; exit } END { exit(found ? 0 : 1) }'
+}
+
+if ! have_nvidia_gpu; then
     log "no NVIDIA GPU detected — skipping module load"
     exit 0
 fi

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

@@ -14,7 +14,7 @@ log() {
 }
 
 have_nvidia_gpu() {
-    lspci -nn 2>/dev/null | grep -qi '10de:'
+    lspci -Dn 2>/dev/null | awk '$2 ~ /^03(00|02):$/ && $3 ~ /^10de:/ { found=1; exit } END { exit(found ? 0 : 1) }'
 }
 
 service_active() {