Rework Power Fit report: 90 min stability, aligned tables, PSU/fan sections

- Increase stability profile duration from 33 min to 90 min by wiring
  powerBenchDurationSec() into runBenchmarkPowerCalibration (was discarded)
- Collect per-step PSU slot readings, fan RPM/duty, and per-GPU telemetry
  in ramp loop; add matching fields to NvidiaPowerBenchStep/NvidiaPowerBenchGPU
- Rewrite renderPowerBenchReport: replace Per-Slot Results with Single GPU
  section, rework Ramp Sequence rows=runs/cols=GPUs, add PSU Performance
  section (conditional on IPMI data), add transposed Single vs All-GPU
  comparison table in per-GPU sections
- Add fmtMDTable helper (benchmark_table.go) and apply to all tables in
  both power and performance reports so columns align in plain-text view

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

audit/internal/app/app.go

@@ -304,7 +304,7 @@ func (a *App) ExportLatestAudit(target platform.RemovableTarget) (string, error)
 	}
 	filename := fmt.Sprintf("audit-%s-%s.json", sanitizeFilename(hostnameOr("unknown")), time.Now().UTC().Format("20060102-150405"))
 	tmpPath := filepath.Join(os.TempDir(), filename)
-	data, err := os.ReadFile(DefaultAuditJSONPath)
+	data, err := readFileLimited(DefaultAuditJSONPath, 100<<20)
 	if err != nil {
 		return "", err
 	}

audit/internal/app/atomic_write.go

@@ -2,10 +2,29 @@ package app
 
 import (
 	"fmt"
+	"io"
 	"os"
 	"path/filepath"
 )
 
+// readFileLimited reads path into memory, refusing files larger than maxBytes.
+// Prevents OOM on corrupted or unexpectedly large data files.
+func readFileLimited(path string, maxBytes int64) ([]byte, error) {
+	f, err := os.Open(path)
+	if err != nil {
+		return nil, err
+	}
+	defer f.Close()
+	data, err := io.ReadAll(io.LimitReader(f, maxBytes+1))
+	if err != nil {
+		return nil, err
+	}
+	if int64(len(data)) > maxBytes {
+		return nil, fmt.Errorf("file %s too large (exceeds %d bytes)", path, maxBytes)
+	}
+	return data, nil
+}
+
 func atomicWriteFile(path string, data []byte, perm os.FileMode) error {
 	if err := os.MkdirAll(filepath.Dir(path), 0755); err != nil {
 		return fmt.Errorf("mkdir %s: %w", filepath.Dir(path), err)

audit/internal/app/component_status_db.go

@@ -46,7 +46,7 @@ func OpenComponentStatusDB(path string) (*ComponentStatusDB, error) {
 	if err := os.MkdirAll(filepath.Dir(path), 0755); err != nil {
 		return nil, err
 	}
-	data, err := os.ReadFile(path)
+	data, err := readFileLimited(path, 10<<20)
 	if err != nil && !os.IsNotExist(err) {
 		return nil, err
 	}

audit/internal/collector/psu.go

@@ -160,11 +160,54 @@ type psuSDR struct {
 }
 
 var psuSlotPatterns = []*regexp.Regexp{
-	regexp.MustCompile(`(?i)\bpsu?\s*([0-9]+)\b`),
-	regexp.MustCompile(`(?i)\bps\s*([0-9]+)\b`),
-	regexp.MustCompile(`(?i)\bpws\s*([0-9]+)\b`),
-	regexp.MustCompile(`(?i)\bpower\s*supply(?:\s*bay)?\s*([0-9]+)\b`),
-	regexp.MustCompile(`(?i)\bbay\s*([0-9]+)\b`),
+	regexp.MustCompile(`(?i)\bpsu?\s*([0-9]+)\b`),                    // PSU1, PS1, ps 2
+	regexp.MustCompile(`(?i)\bps\s*([0-9]+)\b`),                      // PS 6, PS6
+	regexp.MustCompile(`(?i)\bpws\s*([0-9]+)\b`),                     // PWS1
+	regexp.MustCompile(`(?i)\bpower\s*supply(?:\s*bay)?\s*([0-9]+)\b`), // Power Supply 1, Power Supply Bay 3
+	regexp.MustCompile(`(?i)\bbay\s*([0-9]+)\b`),                     // Bay 1
+	// Fallback for xFusion-style generic numbered PSU sensors (Power1, Power2, …).
+	// Must be last: "power supply N" is already caught by the pattern above.
+	regexp.MustCompile(`(?i)\bpower([0-9]+)\b`),
+}
+
+// psuInputPowerKeywords matches AC-input power sensor names across vendors:
+//   MSI:     PSU1_POWER_IN, PSU1_PIN
+//   MLT:     PSU1_PIN
+//   xFusion: (matched via default fallback — no explicit keyword)
+//   HPE:     PS1 Input Power, PS1 Input Watts
+func isPSUInputPower(name string) bool {
+	return strings.Contains(name, "input power") ||
+		strings.Contains(name, "input watts") ||
+		strings.Contains(name, "_pin") ||
+		strings.Contains(name, " pin") ||
+		strings.Contains(name, "_power_in") ||
+		strings.Contains(name, "power_in")
+}
+
+// isPSUOutputPower matches DC-output power sensor names across vendors:
+//   MSI:     PSU1_POWER_OUT
+//   MLT:     PSU1_POUT
+//   xFusion: PS1 POut
+func isPSUOutputPower(name string) bool {
+	return strings.Contains(name, "output power") ||
+		strings.Contains(name, "output watts") ||
+		strings.Contains(name, "_pout") ||
+		strings.Contains(name, " pout") ||
+		strings.Contains(name, "_power_out") ||
+		strings.Contains(name, "power_out") ||
+		strings.Contains(name, "power supply bay") ||
+		strings.Contains(name, "psu bay")
+}
+
+// parseBoundedFloat parses a numeric value from an SDR value field and
+// validates it is within (0, max]. Returns nil for zero, negative, or
+// out-of-range values — these indicate missing/off/fault sensor readings.
+func parseBoundedFloat(raw string, max float64) *float64 {
+	v := parseFloatPtr(raw)
+	if v == nil || *v <= 0 || *v > max {
+		return nil
+	}
+	return v
 }
 
 func parsePSUSDR(raw string) map[int]psuSDR {
@@ -194,24 +237,59 @@ func parsePSUSDR(raw string) map[int]psuSDR {
 
 		lowerName := strings.ToLower(name)
 		switch {
-		case strings.Contains(lowerName, "input power"):
-			entry.inputPowerW = parseFloatPtr(value)
-		case strings.Contains(lowerName, "output power"):
-			entry.outputPowerW = parseFloatPtr(value)
-		case strings.Contains(lowerName, "power supply bay"), strings.Contains(lowerName, "psu bay"):
-			entry.outputPowerW = parseFloatPtr(value)
+		case isPSUInputPower(lowerName):
+			entry.inputPowerW = parseBoundedFloat(value, 6000)
+		case isPSUOutputPower(lowerName):
+			entry.outputPowerW = parseBoundedFloat(value, 6000)
 		case strings.Contains(lowerName, "input voltage"), strings.Contains(lowerName, "ac input"):
 			entry.inputVoltage = parseFloatPtr(value)
 		case strings.Contains(lowerName, "temp"):
 			entry.temperatureC = parseFloatPtr(value)
 		case strings.Contains(lowerName, "health"), strings.Contains(lowerName, "remaining life"), strings.Contains(lowerName, "life remaining"):
 			entry.healthPct = parsePercentPtr(value)
+		default:
+			// Generic PSU power reading: sensor matched a slot pattern but carries
+			// no input/output keyword (e.g. xFusion "Power1", "Power2"). Treat as
+			// AC input if the value looks like wattage and no better data is set yet.
+			if entry.inputPowerW == nil {
+				entry.inputPowerW = parseBoundedFloat(value, 6000)
+			}
 		}
 		out[slot] = entry
 	}
 	return out
 }
 
+// PSUSlotPower holds SDR power readings for one PSU slot.
+// Slot key used by PSUSlotsFromSDR is the 0-based index string,
+// matching HardwarePowerSupply.Slot in the audit schema.
+type PSUSlotPower struct {
+	InputW  *float64 `json:"input_w,omitempty"`
+	OutputW *float64 `json:"output_w,omitempty"`
+	Status  string   `json:"status,omitempty"`
+}
+
+// PSUSlotsFromSDR parses `ipmitool sdr` output and returns per-slot PSU data
+// using the same battle-tested slot patterns as the hardware audit collector.
+// Works across MSI (PSU1_POWER_IN), xFusion (Power1, PS1 POut), MLT (PSU1_PIN).
+// Slot keys are 0-based index strings matching HardwarePowerSupply.Slot.
+func PSUSlotsFromSDR(sdrOutput string) map[string]PSUSlotPower {
+	sdr := parsePSUSDR(sdrOutput)
+	if len(sdr) == 0 {
+		return nil
+	}
+	out := make(map[string]PSUSlotPower, len(sdr))
+	for slot, entry := range sdr {
+		key := strconv.Itoa(slot - 1) // audit uses 0-based slot
+		out[key] = PSUSlotPower{
+			InputW:  entry.inputPowerW,
+			OutputW: entry.outputPowerW,
+			Status:  entry.status,
+		}
+	}
+	return out
+}
+
 func synthesizePSUsFromSDR(sdr map[int]psuSDR) []schema.HardwarePowerSupply {
 	if len(sdr) == 0 {
 		return nil

audit/internal/platform/benchmark_report.go

@@ -89,136 +89,159 @@ func renderBenchmarkReportWithCharts(result NvidiaBenchmarkResult) string {
 
 	// Perspective 1: Compatibility — hard stops
 	b.WriteString("### 1. Compatibility\n\n")
-	b.WriteString("| GPU | Thermal throttle | Fan duty at throttle | ECC uncorr | Status |\n")
-	b.WriteString("|-----|------------------|----------------------|------------|--------|\n")
-	for _, gpu := range result.GPUs {
-		thermalThrottle := "-"
-		if gpu.Scores.ThermalThrottlePct > 0 {
-			thermalThrottle = fmt.Sprintf("%.1f%%", gpu.Scores.ThermalThrottlePct)
+	{
+		var rows [][]string
+		for _, gpu := range result.GPUs {
+			thermalThrottle := "-"
+			if gpu.Scores.ThermalThrottlePct > 0 {
+				thermalThrottle = fmt.Sprintf("%.1f%%", gpu.Scores.ThermalThrottlePct)
+			}
+			fanAtThrottle := "-"
+			if result.Cooling != nil && result.Cooling.FanDutyCycleAvailable && gpu.Scores.ThermalThrottlePct > 0 {
+				fanAtThrottle = fmt.Sprintf("%.0f%%", result.Cooling.P95FanDutyCyclePct)
+			}
+			ecc := "-"
+			if gpu.ECC.Uncorrected > 0 {
+				ecc = fmt.Sprintf("⛔ %d", gpu.ECC.Uncorrected)
+			}
+			compatStatus := "✓ OK"
+			if gpu.ECC.Uncorrected > 0 || (gpu.Scores.ThermalThrottlePct > 0 && result.Cooling != nil && result.Cooling.FanDutyCycleAvailable && result.Cooling.P95FanDutyCyclePct < 95) {
+				compatStatus = "⛔ HARD STOP"
+			}
+			rows = append(rows, []string{fmt.Sprintf("GPU %d", gpu.Index), thermalThrottle, fanAtThrottle, ecc, compatStatus})
 		}
-		fanAtThrottle := "-"
-		if result.Cooling != nil && result.Cooling.FanDutyCycleAvailable && gpu.Scores.ThermalThrottlePct > 0 {
-			fanAtThrottle = fmt.Sprintf("%.0f%%", result.Cooling.P95FanDutyCyclePct)
-		}
-		ecc := "-"
-		if gpu.ECC.Uncorrected > 0 {
-			ecc = fmt.Sprintf("⛔ %d", gpu.ECC.Uncorrected)
-		}
-		compatStatus := "✓ OK"
-		if gpu.ECC.Uncorrected > 0 || (gpu.Scores.ThermalThrottlePct > 0 && result.Cooling != nil && result.Cooling.FanDutyCycleAvailable && result.Cooling.P95FanDutyCyclePct < 95) {
-			compatStatus = "⛔ HARD STOP"
-		}
-		fmt.Fprintf(&b, "| GPU %d | %s | %s | %s | %s |\n",
-			gpu.Index, thermalThrottle, fanAtThrottle, ecc, compatStatus)
+		b.WriteString(fmtMDTable([]string{"GPU", "Thermal throttle", "Fan duty at throttle", "ECC uncorr", "Status"}, rows))
+		b.WriteString("\n")
 	}
-	b.WriteString("\n")
 
 	// Perspective 2: Thermal headroom
 	b.WriteString("### 2. Thermal Headroom\n\n")
-	b.WriteString("| GPU | p95 temp | Slowdown limit | Shutdown limit | Headroom | Thermal throttle | Status |\n")
-	b.WriteString("|-----|----------|----------------|----------------|----------|------------------|--------|\n")
-	for _, gpu := range result.GPUs {
-		shutdownTemp := gpu.ShutdownTempC
-		if shutdownTemp <= 0 {
-			shutdownTemp = 90
+	{
+		var rows [][]string
+		for _, gpu := range result.GPUs {
+			shutdownTemp := gpu.ShutdownTempC
+			if shutdownTemp <= 0 {
+				shutdownTemp = 90
+			}
+			slowdownTemp := gpu.SlowdownTempC
+			if slowdownTemp <= 0 {
+				slowdownTemp = 80
+			}
+			headroom := gpu.Scores.TempHeadroomC
+			thermalStatus := "✓ OK"
+			switch {
+			case headroom < 10:
+				thermalStatus = "⛔ CRITICAL"
+			case gpu.Steady.P95TempC >= slowdownTemp:
+				thermalStatus = "⚠ WARNING"
+			}
+			throttlePct := "-"
+			if gpu.Scores.ThermalThrottlePct > 0 {
+				throttlePct = fmt.Sprintf("%.1f%%", gpu.Scores.ThermalThrottlePct)
+			}
+			rows = append(rows, []string{
+				fmt.Sprintf("GPU %d", gpu.Index),
+				fmt.Sprintf("%.1f°C", gpu.Steady.P95TempC),
+				fmt.Sprintf("%.0f°C", slowdownTemp),
+				fmt.Sprintf("%.0f°C", shutdownTemp),
+				fmt.Sprintf("%.1f°C", headroom),
+				throttlePct,
+				thermalStatus,
+			})
 		}
-		slowdownTemp := gpu.SlowdownTempC
-		if slowdownTemp <= 0 {
-			slowdownTemp = 80
-		}
-		headroom := gpu.Scores.TempHeadroomC
-		thermalStatus := "✓ OK"
-		switch {
-		case headroom < 10:
-			thermalStatus = "⛔ CRITICAL"
-		case gpu.Steady.P95TempC >= slowdownTemp:
-			thermalStatus = "⚠ WARNING"
-		}
-		throttlePct := "-"
-		if gpu.Scores.ThermalThrottlePct > 0 {
-			throttlePct = fmt.Sprintf("%.1f%%", gpu.Scores.ThermalThrottlePct)
-		}
-		fmt.Fprintf(&b, "| GPU %d | %.1f°C | %.0f°C | %.0f°C | %.1f°C | %s | %s |\n",
-			gpu.Index, gpu.Steady.P95TempC, slowdownTemp, shutdownTemp, headroom, throttlePct, thermalStatus)
+		b.WriteString(fmtMDTable([]string{"GPU", "p95 temp", "Slowdown limit", "Shutdown limit", "Headroom", "Thermal throttle", "Status"}, rows))
+		b.WriteString("\n")
 	}
-	b.WriteString("\n")
 
 	// Perspective 3: Power delivery
 	b.WriteString("### 3. Power Delivery\n\n")
-	b.WriteString("| GPU | Power cap throttle | Power stability | Fan duty (p95) | Status |\n")
-	b.WriteString("|-----|-------------------|-----------------|----------------|--------|\n")
-	for _, gpu := range result.GPUs {
-		powerCap := "-"
-		if gpu.Scores.PowerCapThrottlePct > 0 {
-			powerCap = fmt.Sprintf("%.1f%%", gpu.Scores.PowerCapThrottlePct)
+	{
+		var rows [][]string
+		for _, gpu := range result.GPUs {
+			powerCap := "-"
+			if gpu.Scores.PowerCapThrottlePct > 0 {
+				powerCap = fmt.Sprintf("%.1f%%", gpu.Scores.PowerCapThrottlePct)
+			}
+			fanDuty := "-"
+			if result.Cooling != nil && result.Cooling.FanDutyCycleAvailable {
+				fanDuty = fmt.Sprintf("%.0f%%", result.Cooling.P95FanDutyCyclePct)
+			}
+			powerStatus := "✓ OK"
+			if gpu.Scores.PowerCapThrottlePct > 5 {
+				powerStatus = "⚠ POWER LIMITED"
+			}
+			rows = append(rows, []string{
+				fmt.Sprintf("GPU %d", gpu.Index),
+				powerCap,
+				fmt.Sprintf("%.1f", gpu.Scores.PowerSustainScore),
+				fanDuty,
+				powerStatus,
+			})
 		}
-		fanDuty := "-"
-		if result.Cooling != nil && result.Cooling.FanDutyCycleAvailable {
-			fanDuty = fmt.Sprintf("%.0f%%", result.Cooling.P95FanDutyCyclePct)
-		}
-		powerStatus := "✓ OK"
-		if gpu.Scores.PowerCapThrottlePct > 5 {
-			powerStatus = "⚠ POWER LIMITED"
-		}
-		fmt.Fprintf(&b, "| GPU %d | %s | %.1f | %s | %s |\n",
-			gpu.Index, powerCap, gpu.Scores.PowerSustainScore, fanDuty, powerStatus)
+		b.WriteString(fmtMDTable([]string{"GPU", "Power cap throttle", "Power stability", "Fan duty (p95)", "Status"}, rows))
+		b.WriteString("\n")
 	}
-	b.WriteString("\n")
 
 	// Perspective 4: Performance
 	b.WriteString("### 4. Performance\n\n")
-	b.WriteString("| GPU | Compute TOPS | Synthetic | Mixed | Mixed Eff. | TOPS/SM/GHz |\n")
-	b.WriteString("|-----|--------------|-----------|-------|------------|-------------|\n")
-	for _, gpu := range result.GPUs {
-		synthetic := "-"
-		if gpu.Scores.SyntheticScore > 0 {
-			synthetic = fmt.Sprintf("%.2f", gpu.Scores.SyntheticScore)
+	{
+		var rows [][]string
+		for _, gpu := range result.GPUs {
+			synthetic := "-"
+			if gpu.Scores.SyntheticScore > 0 {
+				synthetic = fmt.Sprintf("%.2f", gpu.Scores.SyntheticScore)
+			}
+			mixed := "-"
+			if gpu.Scores.MixedScore > 0 {
+				mixed = fmt.Sprintf("%.2f", gpu.Scores.MixedScore)
+			}
+			mixedEff := "-"
+			if gpu.Scores.MixedEfficiency > 0 {
+				mixedEff = fmt.Sprintf("%.1f%%", gpu.Scores.MixedEfficiency*100)
+			}
+			topsPerSM := "-"
+			if gpu.Scores.TOPSPerSMPerGHz > 0 {
+				topsPerSM = fmt.Sprintf("%.3f", gpu.Scores.TOPSPerSMPerGHz)
+			}
+			rows = append(rows, []string{
+				fmt.Sprintf("GPU %d", gpu.Index),
+				fmt.Sprintf("**%.2f**", gpu.Scores.CompositeScore),
+				synthetic, mixed, mixedEff, topsPerSM,
+			})
 		}
-		mixed := "-"
-		if gpu.Scores.MixedScore > 0 {
-			mixed = fmt.Sprintf("%.2f", gpu.Scores.MixedScore)
+		b.WriteString(fmtMDTable([]string{"GPU", "Compute TOPS", "Synthetic", "Mixed", "Mixed Eff.", "TOPS/SM/GHz"}, rows))
+		if len(result.PerformanceRampSteps) > 0 {
+			fmt.Fprintf(&b, "\n**Platform power score (scalability):** %.1f%%\n", result.PlatformPowerScore)
 		}
-		mixedEff := "-"
-		if gpu.Scores.MixedEfficiency > 0 {
-			mixedEff = fmt.Sprintf("%.1f%%", gpu.Scores.MixedEfficiency*100)
-		}
-		topsPerSM := "-"
-		if gpu.Scores.TOPSPerSMPerGHz > 0 {
-			topsPerSM = fmt.Sprintf("%.3f", gpu.Scores.TOPSPerSMPerGHz)
-		}
-		fmt.Fprintf(&b, "| GPU %d | **%.2f** | %s | %s | %s | %s |\n",
-			gpu.Index, gpu.Scores.CompositeScore, synthetic, mixed, mixedEff, topsPerSM)
+		b.WriteString("\n")
 	}
-	if len(result.PerformanceRampSteps) > 0 {
-		fmt.Fprintf(&b, "\n**Platform power score (scalability):** %.1f%%\n", result.PlatformPowerScore)
-	}
-	b.WriteString("\n")
 
 	// Perspective 5: Anomaly flags
 	b.WriteString("### 5. Anomalies\n\n")
-	b.WriteString("| GPU | ECC corrected | Sync boost throttle | Power instability | Thermal instability |\n")
-	b.WriteString("|-----|---------------|---------------------|-------------------|---------------------|\n")
-	for _, gpu := range result.GPUs {
-		eccCorr := "-"
-		if gpu.ECC.Corrected > 0 {
-			eccCorr = fmt.Sprintf("⚠ %d", gpu.ECC.Corrected)
+	{
+		var rows [][]string
+		for _, gpu := range result.GPUs {
+			eccCorr := "-"
+			if gpu.ECC.Corrected > 0 {
+				eccCorr = fmt.Sprintf("⚠ %d", gpu.ECC.Corrected)
+			}
+			syncBoost := "-"
+			if gpu.Scores.SyncBoostThrottlePct > 0 {
+				syncBoost = fmt.Sprintf("%.1f%%", gpu.Scores.SyncBoostThrottlePct)
+			}
+			powerVar := "OK"
+			if gpu.Scores.PowerSustainScore < 70 {
+				powerVar = "⚠ unstable"
+			}
+			thermalVar := "OK"
+			if gpu.Scores.ThermalSustainScore < 70 {
+				thermalVar = "⚠ unstable"
+			}
+			rows = append(rows, []string{fmt.Sprintf("GPU %d", gpu.Index), eccCorr, syncBoost, powerVar, thermalVar})
 		}
-		syncBoost := "-"
-		if gpu.Scores.SyncBoostThrottlePct > 0 {
-			syncBoost = fmt.Sprintf("%.1f%%", gpu.Scores.SyncBoostThrottlePct)
-		}
-		powerVar := "OK"
-		if gpu.Scores.PowerSustainScore < 70 {
-			powerVar = "⚠ unstable"
-		}
-		thermalVar := "OK"
-		if gpu.Scores.ThermalSustainScore < 70 {
-			thermalVar = "⚠ unstable"
-		}
-		fmt.Fprintf(&b, "| GPU %d | %s | %s | %s | %s |\n",
-			gpu.Index, eccCorr, syncBoost, powerVar, thermalVar)
+		b.WriteString(fmtMDTable([]string{"GPU", "ECC corrected", "Sync boost throttle", "Power instability", "Thermal instability"}, rows))
+		b.WriteString("\n")
 	}
-	b.WriteString("\n")
 
 	// ── Per GPU detail ────────────────────────────────────────────────────────
 	b.WriteString("## Per-GPU Details\n\n")
@@ -263,12 +286,16 @@ func renderBenchmarkReportWithCharts(result NvidiaBenchmarkResult) string {
 		// Steady-state telemetry
 		if benchmarkTelemetryAvailable(gpu.Steady) {
 			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(fmtMDTable(
+				[]string{"", "Avg", "P95"},
+				[][]string{
+					{"Power", fmt.Sprintf("%.1f W", gpu.Steady.AvgPowerW), fmt.Sprintf("%.1f W", gpu.Steady.P95PowerW)},
+					{"Temperature", fmt.Sprintf("%.1f °C", gpu.Steady.AvgTempC), fmt.Sprintf("%.1f °C", gpu.Steady.P95TempC)},
+					{"GPU clock", fmt.Sprintf("%.0f MHz", gpu.Steady.AvgGraphicsClockMHz), fmt.Sprintf("%.0f MHz", gpu.Steady.P95GraphicsClockMHz)},
+					{"Memory clock", fmt.Sprintf("%.0f MHz", gpu.Steady.AvgMemoryClockMHz), fmt.Sprintf("%.0f MHz", gpu.Steady.P95MemoryClockMHz)},
+					{"GPU utilisation", fmt.Sprintf("%.1f %%", gpu.Steady.AvgUsagePct), "—"},
+				},
+			))
 			b.WriteString("\n")
 		} else {
 			b.WriteString("**Steady-state telemetry:** unavailable\n\n")
@@ -277,7 +304,7 @@ func renderBenchmarkReportWithCharts(result NvidiaBenchmarkResult) string {
 		// Per-precision stability phases.
 		if len(gpu.PrecisionSteady) > 0 {
 			b.WriteString("**Per-precision stability:**\n\n")
-			b.WriteString("| Precision | Status | Clock CV | Power CV | Clock Drift | ECC corr | ECC uncorr |\n|-----------|--------|----------|----------|-------------|----------|------------|\n")
+			var precRows [][]string
 			for _, p := range gpu.PrecisionSteady {
 				eccCorr := "—"
 				eccUncorr := "—"
@@ -289,10 +316,15 @@ func renderBenchmarkReportWithCharts(result NvidiaBenchmarkResult) string {
 				if strings.TrimSpace(status) == "" {
 					status = "OK"
 				}
-				fmt.Fprintf(&b, "| %s | %s | %.1f%% | %.1f%% | %.1f%% | %s | %s |\n",
-					p.Precision, status, p.Steady.ClockCVPct, p.Steady.PowerCVPct, p.Steady.ClockDriftPct,
-					eccCorr, eccUncorr)
+				precRows = append(precRows, []string{
+					p.Precision, status,
+					fmt.Sprintf("%.1f%%", p.Steady.ClockCVPct),
+					fmt.Sprintf("%.1f%%", p.Steady.PowerCVPct),
+					fmt.Sprintf("%.1f%%", p.Steady.ClockDriftPct),
+					eccCorr, eccUncorr,
+				})
 			}
+			b.WriteString(fmtMDTable([]string{"Precision", "Status", "Clock CV", "Power CV", "Clock Drift", "ECC corr", "ECC uncorr"}, precRows))
 			b.WriteString("\n")
 		} else {
 			// Legacy: show combined-window variance.
@@ -315,16 +347,22 @@ func renderBenchmarkReportWithCharts(result NvidiaBenchmarkResult) string {
 		// Precision results
 		if len(gpu.PrecisionResults) > 0 {
 			b.WriteString("**Precision results:**\n\n")
-			b.WriteString("| Precision | TOPS (raw) | Weight | TOPS (fp32-eq) | Lanes | Iterations |\n|-----------|------------|--------|----------------|-------|------------|\n")
+			var presRows [][]string
 			for _, p := range gpu.PrecisionResults {
 				if p.Supported {
-					weightStr := fmt.Sprintf("×%.3g", p.Weight)
-					fmt.Fprintf(&b, "| %s | %.2f | %s | %.2f | %d | %d |\n",
-						p.Name, p.TeraOpsPerSec, weightStr, p.WeightedTeraOpsPerSec, p.Lanes, p.Iterations)
+					presRows = append(presRows, []string{
+						p.Name,
+						fmt.Sprintf("%.2f", p.TeraOpsPerSec),
+						fmt.Sprintf("×%.3g", p.Weight),
+						fmt.Sprintf("%.2f", p.WeightedTeraOpsPerSec),
+						fmt.Sprintf("%d", p.Lanes),
+						fmt.Sprintf("%d", p.Iterations),
+					})
 				} else {
-					fmt.Fprintf(&b, "| %s | — (unsupported) | — | — | — | — |\n", p.Name)
+					presRows = append(presRows, []string{p.Name, "— (unsupported)", "—", "—", "—", "—"})
 				}
 			}
+			b.WriteString(fmtMDTable([]string{"Precision", "TOPS (raw)", "Weight", "TOPS (fp32-eq)", "Lanes", "Iterations"}, presRows))
 			b.WriteString("\n")
 		}
 
@@ -346,9 +384,13 @@ func renderBenchmarkReportWithCharts(result NvidiaBenchmarkResult) string {
 		b.WriteString("## Interconnect (NCCL)\n\n")
 		fmt.Fprintf(&b, "**Status:** %s\n\n", result.Interconnect.Status)
 		if result.Interconnect.Supported {
-			b.WriteString("| Metric | Avg | Max |\n|--------|-----|-----|\n")
-			fmt.Fprintf(&b, "| Alg BW | %.1f GB/s | %.1f GB/s |\n", result.Interconnect.AvgAlgBWGBps, result.Interconnect.MaxAlgBWGBps)
-			fmt.Fprintf(&b, "| Bus BW | %.1f GB/s | %.1f GB/s |\n", result.Interconnect.AvgBusBWGBps, result.Interconnect.MaxBusBWGBps)
+			b.WriteString(fmtMDTable(
+				[]string{"Metric", "Avg", "Max"},
+				[][]string{
+					{"Alg BW", fmt.Sprintf("%.1f GB/s", result.Interconnect.AvgAlgBWGBps), fmt.Sprintf("%.1f GB/s", result.Interconnect.MaxAlgBWGBps)},
+					{"Bus BW", fmt.Sprintf("%.1f GB/s", result.Interconnect.AvgBusBWGBps), fmt.Sprintf("%.1f GB/s", result.Interconnect.MaxBusBWGBps)},
+				},
+			))
 			b.WriteString("\n")
 		}
 		for _, note := range result.Interconnect.Notes {
@@ -365,14 +407,16 @@ func renderBenchmarkReportWithCharts(result NvidiaBenchmarkResult) string {
 		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)
+			spRows := [][]string{
+				{"Server idle", fmt.Sprintf("%.0f W", sp.IdleW)},
+				{"Server under load", fmt.Sprintf("%.0f W", sp.LoadedW)},
+				{"Server delta (load − idle)", fmt.Sprintf("%.0f W", sp.DeltaW)},
+				{"GPU-reported sum", fmt.Sprintf("%.0f W", sp.GPUReportedSumW)},
 			}
+			if sp.ReportingRatio > 0 {
+				spRows = append(spRows, []string{"Reporting ratio", fmt.Sprintf("%.2f (1.0 = accurate, <0.75 = GPU over-reports)", sp.ReportingRatio)})
+			}
+			b.WriteString(fmtMDTable([]string{"", "Value"}, spRows))
 			b.WriteString("\n")
 		}
 		for _, note := range sp.Notes {
@@ -383,19 +427,33 @@ func renderBenchmarkReportWithCharts(result NvidiaBenchmarkResult) string {
 		}
 	}
 
+	// ── PSU Issues ────────────────────────────────────────────────────────────
+	if len(result.PSUIssues) > 0 {
+		b.WriteString("## PSU Issues\n\n")
+		b.WriteString("The following power supply anomalies were detected during the benchmark:\n\n")
+		for _, issue := range result.PSUIssues {
+			fmt.Fprintf(&b, "- ⛔ %s\n", issue)
+		}
+		b.WriteString("\n")
+	}
+
 	// ── Cooling ───────────────────────────────────────────────────────────────
 	if cooling := result.Cooling; cooling != nil {
 		b.WriteString("## Cooling\n\n")
 		if cooling.Available {
-			b.WriteString("| Metric | Value |\n|--------|-------|\n")
-			fmt.Fprintf(&b, "| Average fan speed | %.0f RPM |\n", cooling.AvgFanRPM)
+			dutyAvg, dutyP95 := "N/A", "N/A"
 			if cooling.FanDutyCycleAvailable {
-				fmt.Fprintf(&b, "| Average fan duty cycle | %.1f%% |\n", cooling.AvgFanDutyCyclePct)
-				fmt.Fprintf(&b, "| P95 fan duty cycle | %.1f%% |\n", cooling.P95FanDutyCyclePct)
-			} else {
-				b.WriteString("| Average fan duty cycle | N/A |\n")
-				b.WriteString("| P95 fan duty cycle | N/A |\n")
+				dutyAvg = fmt.Sprintf("%.1f%%", cooling.AvgFanDutyCyclePct)
+				dutyP95 = fmt.Sprintf("%.1f%%", cooling.P95FanDutyCyclePct)
 			}
+			b.WriteString(fmtMDTable(
+				[]string{"Metric", "Value"},
+				[][]string{
+					{"Average fan speed", fmt.Sprintf("%.0f RPM", cooling.AvgFanRPM)},
+					{"Average fan duty cycle", dutyAvg},
+					{"P95 fan duty cycle", dutyP95},
+				},
+			))
 			b.WriteString("\n")
 		} else {
 			b.WriteString("Cooling telemetry unavailable.\n\n")
@@ -412,12 +470,16 @@ func renderBenchmarkReportWithCharts(result NvidiaBenchmarkResult) string {
 	if len(result.PerformanceRampSteps) > 0 {
 		b.WriteString("## Platform Scalability (Performance Ramp)\n\n")
 		fmt.Fprintf(&b, "**Platform power score:** %.1f%%  \n\n", result.PlatformPowerScore)
-		b.WriteString("| k GPUs | GPU Indices | Total Synthetic TOPS | Scalability |\n")
-		b.WriteString("|--------|-------------|----------------------|-------------|\n")
+		var scalRows [][]string
 		for _, step := range result.PerformanceRampSteps {
-			fmt.Fprintf(&b, "| %d | %s | %.2f | %.1f%% |\n",
-				step.StepIndex, joinIndexList(step.GPUIndices), step.TotalSyntheticTOPS, step.ScalabilityPct)
+			scalRows = append(scalRows, []string{
+				fmt.Sprintf("%d", step.StepIndex),
+				joinIndexList(step.GPUIndices),
+				fmt.Sprintf("%.2f", step.TotalSyntheticTOPS),
+				fmt.Sprintf("%.1f%%", step.ScalabilityPct),
+			})
 		}
+		b.WriteString(fmtMDTable([]string{"k GPUs", "GPU Indices", "Total Synthetic TOPS", "Scalability"}, scalRows))
 		b.WriteString("\n")
 	}
 

audit/internal/platform/benchmark_table.go

@@ -0,0 +1,75 @@
+package platform
+
+import (
+	"strings"
+)
+
+// fmtMDTable renders a markdown table with column widths padded so the table
+// is readable as plain text without a markdown renderer.
+//
+// headers contains the column header strings.
+// rows contains data rows; each row must have the same number of cells as headers.
+// Cells with fewer entries than headers are treated as empty.
+func fmtMDTable(headers []string, rows [][]string) string {
+	ncols := len(headers)
+	if ncols == 0 {
+		return ""
+	}
+
+	// Compute max width per column.
+	widths := make([]int, ncols)
+	for i, h := range headers {
+		if len(h) > widths[i] {
+			widths[i] = len(h)
+		}
+	}
+	for _, row := range rows {
+		for i := 0; i < ncols; i++ {
+			cell := ""
+			if i < len(row) {
+				cell = row[i]
+			}
+			if len(cell) > widths[i] {
+				widths[i] = len(cell)
+			}
+		}
+	}
+
+	var b strings.Builder
+
+	// Header row.
+	b.WriteByte('|')
+	for i, h := range headers {
+		b.WriteByte(' ')
+		b.WriteString(h)
+		b.WriteString(strings.Repeat(" ", widths[i]-len(h)))
+		b.WriteString(" |")
+	}
+	b.WriteByte('\n')
+
+	// Separator row.
+	b.WriteByte('|')
+	for i := range headers {
+		b.WriteString(strings.Repeat("-", widths[i]+2))
+		b.WriteByte('|')
+	}
+	b.WriteByte('\n')
+
+	// Data rows.
+	for _, row := range rows {
+		b.WriteByte('|')
+		for i := 0; i < ncols; i++ {
+			cell := ""
+			if i < len(row) {
+				cell = row[i]
+			}
+			b.WriteByte(' ')
+			b.WriteString(cell)
+			b.WriteString(strings.Repeat(" ", widths[i]-len(cell)))
+			b.WriteString(" |")
+		}
+		b.WriteByte('\n')
+	}
+
+	return b.String()
+}

audit/internal/platform/benchmark_types.go

@@ -43,6 +43,31 @@ const (
 	NvidiaBenchmarkProfileOvernight = "overnight"
 )
 
+// Estimated wall-clock durations for benchmark runs, derived from real _v8 logs.
+// Rule: when changing profile phase durations in resolveBenchmarkProfile(),
+// re-measure from actual task logs and update the constants here.
+//
+// Sources:
+//   - BenchmarkEstimatedPerfStandardSec:   MLT v8.22 ramp 1-4: 927 s; xFusion v8.22 parallel 8GPU: 1080 s
+//   - BenchmarkEstimatedPerfStabilitySec:  xFusion v8.22 ramp 1-8: 5532 s
+//   - BenchmarkEstimatedPerfOvernightSec:  derived from profile phases (SteadySec=27000)
+//   - BenchmarkEstimatedPowerStandardSec:  MLT v8.22 ramp 1-4: 2663 s; MSI v8.22 ramp 1-8: 2375 s
+//   - BenchmarkEstimatedPowerStabilitySec: target ~90 min with calibDurationSec=300 (8 GPU × ~2-3 attempts)
+const (
+	// Performance Benchmark (bee-gpu-burn).
+	// Duration is per full ramp-up run (ramp 1→N) or per single parallel run.
+	// Sequential per-GPU mode scales approximately linearly.
+	BenchmarkEstimatedPerfStandardSec  = 960  // ~16 min; ramp-up 1-4: 927 s, parallel 8GPU: 1080 s
+	BenchmarkEstimatedPerfStabilitySec = 5532 // ~92 min; ramp-up 1-8 measured
+	BenchmarkEstimatedPerfOvernightSec = 8 * 3600
+
+	// Power / Thermal Fit (dcgmi targeted_power binary-search calibration).
+	// Duration is for the full ramp-up run; individual steps vary with convergence speed.
+	BenchmarkEstimatedPowerStandardSec  = 2600 // ~43 min; ramp 1-4: 2663 s, ramp 1-8: 2375 s
+	BenchmarkEstimatedPowerStabilitySec = 5400 // ~90 min; calibDurationSec=300 × 8 GPU × ~2-3 attempts
+	BenchmarkEstimatedPowerOvernightSec = 3 * 3600
+)
+
 type NvidiaBenchmarkOptions struct {
 	Profile           string
 	SizeMB            int
@@ -82,6 +107,10 @@ type NvidiaBenchmarkResult struct {
 	GPUs                 []BenchmarkGPUResult         `json:"gpus"`
 	Interconnect         *BenchmarkInterconnectResult `json:"interconnect,omitempty"`
 	ServerPower          *BenchmarkServerPower        `json:"server_power,omitempty"`
+	// PSUIssues holds power supply fault events detected by comparing IPMI PSU
+	// sensor states before and after the benchmark run. Empty when IPMI is
+	// unavailable or no PSU faults occurred during the test.
+	PSUIssues []string `json:"psu_issues,omitempty"`
 }
 
 type BenchmarkNormalization struct {
@@ -246,18 +275,55 @@ type BenchmarkScorecard struct {
 	TOPSPerSMPerGHz float64 `json:"tops_per_sm_per_ghz,omitempty"`
 }
 
-// BenchmarkServerPower captures server-side power via IPMI alongside GPU-reported
-// power. The reporting_ratio (delta / gpu_reported_sum) near 1.0 means GPU power
-// telemetry is accurate; a ratio well below 1.0 (e.g. 0.5) means the GPU is
-// over-reporting its power consumption.
+// BenchmarkPSUSlotPower holds SDR power readings for one PSU slot sampled
+// during the benchmark. Slot keys match audit HardwarePowerSupply.Slot (0-based)
+// so benchmark and audit data can be correlated by slot.
+type BenchmarkPSUSlotPower struct {
+	InputW  *float64 `json:"input_w,omitempty"`  // AC wall input (PSUx_POWER_IN)
+	OutputW *float64 `json:"output_w,omitempty"` // DC output (PSUx_POWER_OUT)
+	Status  string   `json:"status,omitempty"`
+}
+
+// BenchmarkServerPower captures server-side power from multiple independent
+// sources: IPMI DCMI (high-level), IPMI SDR per-PSU sensors (granular), and
+// GPU-reported power (nvidia-smi). Cross-comparing sources detects when DCMI
+// covers only a subset of installed PSUs (partial coverage).
+//
+// Source legend:
+//   - DCMI      — `ipmitool dcmi power reading`; fast but may miss PSUs
+//   - SDR       — `ipmitool sdr` PSUx_POWER_IN/OUT; per-PSU, reliable
+//   - nvidia-smi — GPU self-reported via internal shunt; accurate for GPU load
 type BenchmarkServerPower struct {
-	Available       bool     `json:"available"`
-	IdleW           float64  `json:"idle_w,omitempty"`
-	LoadedW         float64  `json:"loaded_w,omitempty"`
-	DeltaW          float64  `json:"delta_w,omitempty"`
-	GPUReportedSumW float64  `json:"gpu_reported_sum_w,omitempty"`
-	ReportingRatio  float64  `json:"reporting_ratio,omitempty"`
-	Notes           []string `json:"notes,omitempty"`
+	Available       bool    `json:"available"`
+	IdleW           float64 `json:"idle_w,omitempty"`   // DCMI at idle
+	LoadedW         float64 `json:"loaded_w,omitempty"` // DCMI at peak load
+	DeltaW          float64 `json:"delta_w,omitempty"`  // DCMI loaded − idle
+	GPUReportedSumW float64 `json:"gpu_reported_sum_w,omitempty"`
+	ReportingRatio  float64 `json:"reporting_ratio,omitempty"`
+
+	// PSU AC input sum — sampled at idle and at peak load using collector's
+	// slot patterns (PSU1_POWER_IN, PSU1_PIN, PS1 POut, Power1…).
+	PSUInputIdleW   float64 `json:"psu_input_idle_w,omitempty"`
+	PSUInputLoadedW float64 `json:"psu_input_loaded_w,omitempty"`
+
+	// PSU DC output sum — power delivered to server internals after conversion.
+	PSUOutputIdleW   float64 `json:"psu_output_idle_w,omitempty"`
+	PSUOutputLoadedW float64 `json:"psu_output_loaded_w,omitempty"`
+
+	// Per-slot PSU readings at idle and at peak load.
+	// Keys are 0-based slot strings matching audit HardwarePowerSupply.Slot.
+	PSUSlotReadingsIdle   map[string]BenchmarkPSUSlotPower `json:"psu_slot_readings_idle,omitempty"`
+	PSUSlotReadingsLoaded map[string]BenchmarkPSUSlotPower `json:"psu_slot_readings_loaded,omitempty"`
+
+	// GPUSlotTotalW is the sum of GPU_POWER_SLOTx SDR sensors at peak load.
+	// PCIe slot delivery only (excludes 16-pin connector power).
+	GPUSlotTotalW float64 `json:"gpu_slot_total_w,omitempty"`
+
+	// DCMICoverageRatio = DCMI_idle / SDR_PSU_IN_idle.
+	// Near 1.0 → DCMI tracks all PSUs. Near 0.5 → DCMI tracks half the PSUs.
+	DCMICoverageRatio float64 `json:"dcmi_coverage_ratio,omitempty"`
+
+	Notes []string `json:"notes,omitempty"`
 }
 
 // BenchmarkPrecisionSteadyPhase holds per-precision-category telemetry collected
@@ -308,6 +374,10 @@ type NvidiaPowerBenchResult struct {
 	ServerPower *BenchmarkServerPower `json:"server_power,omitempty"`
 	Findings    []string              `json:"findings,omitempty"`
 	GPUs        []NvidiaPowerBenchGPU `json:"gpus"`
+	// PSUIssues holds power supply fault events detected by comparing IPMI PSU
+	// sensor states before and after the power benchmark run. Empty when IPMI is
+	// unavailable or no PSU faults occurred during the test.
+	PSUIssues []string `json:"psu_issues,omitempty"`
 }
 
 type NvidiaPowerBenchGPU struct {
@@ -338,6 +408,9 @@ type NvidiaPowerBenchGPU struct {
 	// Telemetry holds the aggregated stats from the final converged calibration
 	// attempt for this GPU (temperature, power, fan, clock percentiles).
 	Telemetry *BenchmarkTelemetrySummary `json:"telemetry,omitempty"`
+	// Fan state sampled at the end of single-card calibration.
+	AvgFanRPM          float64 `json:"avg_fan_rpm,omitempty"`
+	AvgFanDutyCyclePct float64 `json:"avg_fan_duty_cycle_pct,omitempty"`
 }
 
 type NvidiaPowerBenchStep struct {
@@ -356,6 +429,13 @@ type NvidiaPowerBenchStep struct {
 	// ramp step's calibration run. ServerDeltaW = ServerLoadedW − idle.
 	ServerLoadedW float64 `json:"server_loaded_w,omitempty"`
 	ServerDeltaW  float64 `json:"server_delta_w,omitempty"`
+	// PSU slot readings sampled at end of this ramp step.
+	PSUSlotReadings map[string]BenchmarkPSUSlotPower `json:"psu_slot_readings,omitempty"`
+	// Fan state at end of this ramp step.
+	AvgFanRPM          float64 `json:"avg_fan_rpm,omitempty"`
+	AvgFanDutyCyclePct float64 `json:"avg_fan_duty_cycle_pct,omitempty"`
+	// Per-GPU telemetry from this step's calibration, keyed by GPU index.
+	PerGPUTelemetry map[int]*BenchmarkTelemetrySummary `json:"per_gpu_telemetry,omitempty"`
 }
 
 // NvidiaPerformanceRampStep holds per-step performance data for the

audit/internal/platform/kill_workers.go

@@ -1,11 +1,14 @@
 package platform
 
 import (
+	"context"
 	"fmt"
+	"log/slog"
 	"os"
 	"strconv"
 	"strings"
 	"syscall"
+	"time"
 )
 
 // workerPatterns are substrings matched against /proc/<pid>/cmdline to identify
@@ -30,7 +33,12 @@ type KilledProcess struct {
 // KillTestWorkers scans /proc for running test worker processes and sends
 // SIGKILL to each one found. It returns a list of killed processes.
 // Errors for individual processes (e.g. already exited) are silently ignored.
+// The scan runs under a 5-second deadline to avoid blocking if the process
+// table is very large (e.g. after a stress test with thousands of children).
 func KillTestWorkers() []KilledProcess {
+	ctx, cancel := context.WithTimeout(context.Background(), 5*time.Second)
+	defer cancel()
+
 	entries, err := os.ReadDir("/proc")
 	if err != nil {
 		return nil
@@ -38,6 +46,13 @@ func KillTestWorkers() []KilledProcess {
 
 	var killed []KilledProcess
 	for _, e := range entries {
+		select {
+		case <-ctx.Done():
+			slog.Warn("KillTestWorkers scan timed out", "killed_so_far", len(killed))
+			return killed
+		default:
+		}
+
 		if !e.IsDir() {
 			continue
 		}

audit/internal/platform/live_metrics.go

@@ -18,11 +18,19 @@ type LiveMetricSample struct {
 	Fans       []FanReading   `json:"fans"`
 	Temps      []TempReading  `json:"temps"`
 	PowerW     float64        `json:"power_w"`
+	PSUs       []PSUReading   `json:"psus,omitempty"`
 	CPULoadPct float64        `json:"cpu_load_pct"`
 	MemLoadPct float64        `json:"mem_load_pct"`
 	GPUs       []GPUMetricRow `json:"gpus"`
 }
 
+// PSUReading is a per-slot power supply input power reading.
+type PSUReading struct {
+	Slot   int     `json:"slot"`
+	Name   string  `json:"name"`
+	PowerW float64 `json:"power_w"`
+}
+
 // TempReading is a named temperature sensor value.
 type TempReading struct {
 	Name    string  `json:"name"`
@@ -57,6 +65,9 @@ func SampleLiveMetrics() LiveMetricSample {
 	// System power — returns 0 if unavailable
 	s.PowerW = sampleSystemPower()
 
+	// Per-PSU power — populated when IPMI SDR has Power Supply entities with Watt readings
+	s.PSUs = samplePSUPower()
+
 	// CPU load — from /proc/stat
 	s.CPULoadPct = sampleCPULoadPct()
 
@@ -326,3 +337,65 @@ func compactAmbientTempName(chip, name string) string {
 	}
 	return chip + " / " + name
 }
+
+// samplePSUPower reads per-PSU input power via IPMI SDR.
+// It parses `ipmitool sdr elist full` output looking for Power Supply entity
+// sensors (entity ID "10.N") that report a value in Watts.
+// Returns nil when IPMI is unavailable or no PSU Watt sensors exist.
+func samplePSUPower() []PSUReading {
+	out, err := exec.Command("ipmitool", "sdr", "elist", "full").Output()
+	if err != nil || len(out) == 0 {
+		return nil
+	}
+	// map slot → reading (keep highest-watt value per slot in case of duplicates)
+	type entry struct {
+		name   string
+		powerW float64
+	}
+	bySlot := map[int]entry{}
+	for _, line := range strings.Split(string(out), "\n") {
+		parts := strings.Split(line, "|")
+		if len(parts) < 5 {
+			continue
+		}
+		entityID := strings.TrimSpace(parts[3]) // e.g. "10.1"
+		if !strings.HasPrefix(entityID, "10.") {
+			continue // not a Power Supply entity
+		}
+		slotStr := strings.TrimPrefix(entityID, "10.")
+		slot, err := strconv.Atoi(slotStr)
+		if err != nil {
+			continue
+		}
+		valueField := strings.TrimSpace(parts[4]) // e.g. "740.00 Watts"
+		if !strings.Contains(strings.ToLower(valueField), "watts") {
+			continue
+		}
+		valueFields := strings.Fields(valueField)
+		if len(valueFields) < 2 {
+			continue
+		}
+		w, err := strconv.ParseFloat(valueFields[0], 64)
+		if err != nil || w <= 0 {
+			continue
+		}
+		sensorName := strings.TrimSpace(parts[0])
+		if existing, ok := bySlot[slot]; !ok || w > existing.powerW {
+			bySlot[slot] = entry{name: sensorName, powerW: w}
+		}
+	}
+	if len(bySlot) == 0 {
+		return nil
+	}
+	slots := make([]int, 0, len(bySlot))
+	for s := range bySlot {
+		slots = append(slots, s)
+	}
+	sort.Ints(slots)
+	psus := make([]PSUReading, 0, len(slots))
+	for _, s := range slots {
+		e := bySlot[s]
+		psus = append(psus, PSUReading{Slot: s, Name: e.name, PowerW: e.powerW})
+	}
+	return psus
+}

audit/internal/platform/sat.go

@@ -20,6 +20,54 @@ import (
 	"time"
 )
 
+// Estimated wall-clock durations for each SAT/validate test, derived from real
+// production logs in _benchmark/_v8/.
+//
+// Rule: whenever the commands, timeout parameters, or number of sub-jobs inside
+// the corresponding Run*Pack function change, re-measure the wall-clock duration
+// from actual task logs and update the matching constant here.
+//
+// Sources:
+//   - SATEstimatedCPUValidateSec:                 xFusion v8.6 — 62 s
+//   - SATEstimatedMemoryValidateSec:               xFusion v8.6 — 68 s
+//   - SATEstimatedNvidiaGPUValidatePerGPUSec:      xFusion v8.6/v8.22 — 77–87 s/GPU
+//   - SATEstimatedNvidiaGPUStressPerGPUSec:        xFusion v8.6/v8.22 — 444–448 s/GPU
+//   - SATEstimatedNvidiaTargetedStressPerGPUSec:   xFusion v8.6/v8.22 — 347–348 s/GPU (300 s default + overhead)
+//   - SATEstimatedNvidiaTargetedPowerPerGPUSec:    MSI v8.22 / xFusion v8.6 — 346–351 s/GPU
+//   - SATEstimatedNvidiaPulseTestSec:              xFusion v8.6 — 4 926 s / 8 GPU (all simultaneous)
+//   - SATEstimatedNvidiaInterconnectSec:           xFusion v8.6/v8.22 — 210–384 s / 8 GPU (all simultaneous)
+//   - SATEstimatedNvidiaBandwidthSec:              xFusion v8.6/v8.22 — 2 664–2 688 s / 8 GPU (all simultaneous)
+const (
+	// CPU stress: stress-ng 60 s + lscpu/sensors overhead.
+	SATEstimatedCPUValidateSec = 65
+	// CPU stress: stress-ng 1800 s (stress mode default).
+	SATEstimatedCPUStressSec = 1800
+
+	// RAM: memtester 256 MB / 1 pass.
+	SATEstimatedMemoryValidateSec = 70
+	// RAM: memtester 512 MB / 1 pass (extrapolated from validate timing, linear with size).
+	SATEstimatedMemoryStressSec = 140
+
+	// NVIDIA dcgmi diag Level 2 (medium), per GPU, sequential.
+	SATEstimatedNvidiaGPUValidatePerGPUSec = 85
+	// NVIDIA dcgmi diag Level 3 (targeted stress), per GPU, sequential.
+	SATEstimatedNvidiaGPUStressPerGPUSec = 450
+
+	// NVIDIA dcgmi targeted_stress 300 s + overhead, per GPU, sequential.
+	SATEstimatedNvidiaTargetedStressPerGPUSec = 350
+	// NVIDIA dcgmi targeted_power 300 s + overhead, per GPU, sequential.
+	SATEstimatedNvidiaTargetedPowerPerGPUSec = 350
+
+	// NVIDIA dcgmi pulse_test, all GPUs simultaneously (not per-GPU).
+	SATEstimatedNvidiaPulseTestSec = 5000
+
+	// NCCL all_reduce_perf, all GPUs simultaneously.
+	SATEstimatedNvidiaInterconnectSec = 300
+	// nvbandwidth, all GPUs simultaneously. Tool runs all built-in tests
+	// without a user-configurable time limit; duration is determined by nvbandwidth itself.
+	SATEstimatedNvidiaBandwidthSec = 2700
+)
+
 var (
 	satExecCommand  = exec.Command
 	satLookPath     = exec.LookPath

audit/internal/webui/api_test.go

@@ -178,16 +178,20 @@ func TestHandleAPIBenchmarkPowerFitRampQueuesBenchmarkPowerFitTasks(t *testing.T
 	}
 	globalQueue.mu.Lock()
 	defer globalQueue.mu.Unlock()
-	if len(globalQueue.tasks) != 3 {
-		t.Fatalf("tasks=%d want 3", len(globalQueue.tasks))
+	// Ramp-up mode creates a single task that handles the 1→N GPU ramp internally
+	// (spawning N separate tasks would redundantly repeat all earlier ramp steps).
+	if len(globalQueue.tasks) != 1 {
+		t.Fatalf("tasks=%d want 1 (ramp-up uses single task)", len(globalQueue.tasks))
 	}
-	for i, task := range globalQueue.tasks {
-		if task.Target != "nvidia-bench-power" {
-			t.Fatalf("task[%d] target=%q", i, task.Target)
-		}
-		if task.Priority != taskPriorityBenchmark {
-			t.Fatalf("task[%d] priority=%d want %d", i, task.Priority, taskPriorityBenchmark)
-		}
+	task := globalQueue.tasks[0]
+	if task.Target != "nvidia-bench-power" {
+		t.Fatalf("task target=%q want nvidia-bench-power", task.Target)
+	}
+	if task.Priority != taskPriorityBenchmark {
+		t.Fatalf("task priority=%d want %d", task.Priority, taskPriorityBenchmark)
+	}
+	if task.params.RampTotal != 3 {
+		t.Fatalf("task RampTotal=%d want 3", task.params.RampTotal)
 	}
 }
 

audit/internal/webui/charts_svg.go

@@ -462,6 +462,127 @@ func synthesizeChartTimes(times []time.Time, count int) []time.Time {
 	return out
 }
 
+// renderStackedMetricChartSVG renders a stacked area chart where each dataset
+// is visually "stacked" on top of the previous one. Intended for multi-PSU
+// power charts where the filled area of each PSU shows its individual
+// contribution and the total height equals the combined draw.
+func renderStackedMetricChartSVG(title string, labels []string, times []time.Time, datasets [][]float64, names []string, yMax *float64, canvasHeight int, timeline []chartTimelineSegment) ([]byte, error) {
+	pointCount := len(labels)
+	if len(times) > pointCount {
+		pointCount = len(times)
+	}
+	if pointCount == 0 {
+		pointCount = 1
+		labels = []string{""}
+		times = []time.Time{{}}
+	}
+	if len(labels) < pointCount {
+		padded := make([]string, pointCount)
+		copy(padded, labels)
+		labels = padded
+	}
+	if len(times) < pointCount {
+		times = synthesizeChartTimes(times, pointCount)
+	}
+	for i := range datasets {
+		if len(datasets[i]) == 0 {
+			datasets[i] = make([]float64, pointCount)
+		}
+	}
+
+	times, datasets = downsampleTimeSeries(times, datasets, 1400)
+	pointCount = len(times)
+
+	// Build cumulative sums per time point.
+	cumulative := make([][]float64, len(datasets)+1)
+	for i := range cumulative {
+		cumulative[i] = make([]float64, pointCount)
+	}
+	for i, ds := range datasets {
+		for j, v := range ds {
+			cumulative[i+1][j] = cumulative[i][j] + v
+		}
+	}
+
+	// Scale is based on the total (top cumulative row).
+	total := cumulative[len(cumulative)-1]
+	yMin := floatPtr(0)
+	if yMax == nil {
+		yMax = autoMax120(total)
+	}
+	scale := singleAxisChartScale([][]float64{total}, yMin, yMax)
+
+	legendItems := make([]metricChartSeries, len(datasets))
+	for i, name := range names {
+		color := metricChartPalette[i%len(metricChartPalette)]
+		legendItems[i] = metricChartSeries{Name: name, Color: color, Values: datasets[i]}
+	}
+
+	// Stats label from totals.
+	statsLabel := chartStatsLabel([][]float64{total})
+
+	layout := singleAxisChartLayout(canvasHeight, len(legendItems))
+	start, end := chartTimeBounds(times)
+
+	var b strings.Builder
+	writeSVGOpen(&b, layout.Width, layout.Height)
+	writeChartFrame(&b, title, statsLabel, layout.Width, layout.Height)
+	writeTimelineIdleSpans(&b, layout, start, end, timeline)
+	writeVerticalGrid(&b, layout, times, pointCount, 8)
+	writeHorizontalGrid(&b, layout, scale)
+	writeTimelineBoundaries(&b, layout, start, end, timeline)
+	writePlotBorder(&b, layout)
+	writeSingleAxisY(&b, layout, scale)
+	writeXAxisLabels(&b, layout, times, labels, start, end, 8)
+
+	// Draw stacked areas from top to bottom so lower layers are visible.
+	for i := len(datasets) - 1; i >= 0; i-- {
+		writeStackedArea(&b, layout, times, start, end, cumulative[i], cumulative[i+1], scale, legendItems[i].Color)
+	}
+	// Draw border polylines on top.
+	for i := len(datasets) - 1; i >= 0; i-- {
+		writeSeriesPolyline(&b, layout, times, start, end, cumulative[i+1], scale, legendItems[i].Color)
+	}
+
+	writeLegend(&b, layout, legendItems)
+	writeSVGClose(&b)
+	return []byte(b.String()), nil
+}
+
+// writeStackedArea draws a filled polygon between two cumulative value arrays
+// (baseline and top), using the given color at 55% opacity.
+func writeStackedArea(b *strings.Builder, layout chartLayout, times []time.Time, start, end time.Time, baseline, top []float64, scale chartScale, color string) {
+	n := len(top)
+	if n == 0 {
+		return
+	}
+	if len(baseline) < n {
+		baseline = make([]float64, n)
+	}
+
+	// Forward path along top values, then backward along baseline values.
+	var points strings.Builder
+	for i := 0; i < n; i++ {
+		x := chartXForTime(chartPointTime(times, i), start, end, layout.PlotLeft, layout.PlotRight)
+		y := chartYForValue(valueClamp(top[i], scale), scale, layout.PlotTop, layout.PlotBottom)
+		if i > 0 {
+			points.WriteByte(' ')
+		}
+		points.WriteString(strconv.FormatFloat(x, 'f', 1, 64))
+		points.WriteByte(',')
+		points.WriteString(strconv.FormatFloat(y, 'f', 1, 64))
+	}
+	for i := n - 1; i >= 0; i-- {
+		x := chartXForTime(chartPointTime(times, i), start, end, layout.PlotLeft, layout.PlotRight)
+		y := chartYForValue(valueClamp(baseline[i], scale), scale, layout.PlotTop, layout.PlotBottom)
+		points.WriteByte(' ')
+		points.WriteString(strconv.FormatFloat(x, 'f', 1, 64))
+		points.WriteByte(',')
+		points.WriteString(strconv.FormatFloat(y, 'f', 1, 64))
+	}
+	fmt.Fprintf(b, `<polygon points="%s" fill="%s" fill-opacity="0.55" stroke="none"/>`+"\n", points.String(), color)
+}
+
 func writeSVGOpen(b *strings.Builder, width, height int) {
 	fmt.Fprintf(b, `<svg xmlns="http://www.w3.org/2000/svg" width="%d" height="%d" viewBox="0 0 %d %d">`+"\n", width, height, width, height)
 }

audit/internal/webui/jobs.go

@@ -1,6 +1,9 @@
 package webui
 
 import (
+	"bufio"
+	"fmt"
+	"io"
 	"os"
 	"strings"
 	"sync"
@@ -17,6 +20,25 @@ type jobState struct {
 	cancel       func() // optional cancel function; nil if job is not cancellable
 	logPath      string
 	serialPrefix string
+	logFile      *os.File    // kept open for the task lifetime to avoid per-line open/close
+	logBuf       *bufio.Writer
+}
+
+// readTaskLogFile reads a task log, refusing files over 50 MB.
+func readTaskLogFile(path string) ([]byte, error) {
+	f, err := os.Open(path)
+	if err != nil {
+		return nil, err
+	}
+	defer f.Close()
+	data, err := io.ReadAll(io.LimitReader(f, 50<<20+1))
+	if err != nil {
+		return nil, err
+	}
+	if int64(len(data)) > 50<<20 {
+		return nil, fmt.Errorf("task log %s too large (exceeds 50 MB)", path)
+	}
+	return data, nil
 }
 
 // abort cancels the job if it has a cancel function and is not yet done.
@@ -35,7 +57,7 @@ func (j *jobState) append(line string) {
 	defer j.mu.Unlock()
 	j.lines = append(j.lines, line)
 	if j.logPath != "" {
-		appendJobLog(j.logPath, line)
+		j.writeLogLineLocked(line)
 	}
 	if j.serialPrefix != "" {
 		taskSerialWriteLine(j.serialPrefix + line)
@@ -48,6 +70,35 @@ func (j *jobState) append(line string) {
 	}
 }
 
+// writeLogLineLocked writes a line to the persistent log file, opening it lazily.
+// Must be called with j.mu held. Uses a buffered writer kept open for the task
+// lifetime — avoids thousands of open/close syscalls during high-frequency logs.
+func (j *jobState) writeLogLineLocked(line string) {
+	if j.logFile == nil {
+		f, err := os.OpenFile(j.logPath, os.O_CREATE|os.O_APPEND|os.O_WRONLY, 0644)
+		if err != nil {
+			return
+		}
+		j.logFile = f
+		j.logBuf = bufio.NewWriterSize(f, 64*1024)
+	}
+	_, _ = j.logBuf.WriteString(line + "\n")
+}
+
+// closeLog flushes and closes the log file. Called after all task output is done.
+func (j *jobState) closeLog() {
+	j.mu.Lock()
+	defer j.mu.Unlock()
+	if j.logBuf != nil {
+		_ = j.logBuf.Flush()
+	}
+	if j.logFile != nil {
+		_ = j.logFile.Close()
+		j.logFile = nil
+		j.logBuf = nil
+	}
+}
+
 func (j *jobState) finish(errMsg string) {
 	j.mu.Lock()
 	defer j.mu.Unlock()
@@ -119,7 +170,7 @@ func newTaskJobState(logPath string, serialPrefix ...string) *jobState {
 	if logPath == "" {
 		return j
 	}
-	data, err := os.ReadFile(logPath)
+	data, err := readTaskLogFile(logPath)
 	if err != nil || len(data) == 0 {
 		return j
 	}

audit/internal/webui/metricsdb.go

@@ -161,6 +161,56 @@ func (m *MetricsDB) Write(s platform.LiveMetricSample) error {
 	return tx.Commit()
 }
 
+// Downsample reduces density of old metrics rows to 1 sample per minute.
+// Only rows in the half-open window [deleteOlderThan, downsampleBefore) are
+// affected — rows newer than downsampleBefore keep full 5-second resolution.
+// For each 60-second bucket the row with the smallest ts is kept; the rest
+// are deleted. This trims ~92 % of rows in that window while preserving
+// the overall shape of every chart.
+//
+// Called hourly by the metrics collector background goroutine.
+func (m *MetricsDB) Downsample(downsampleBefore, deleteOlderThan time.Time) error {
+	if m == nil || m.db == nil {
+		return nil
+	}
+	start := deleteOlderThan.Unix()
+	end := downsampleBefore.Unix()
+	if end <= start {
+		return nil
+	}
+	// For each table: delete rows in [start, end) whose ts is NOT the minimum
+	// ts in its 60-second bucket (ts/60 integer division = bucket ID).
+	for _, table := range []string{"sys_metrics", "gpu_metrics", "fan_metrics", "temp_metrics"} {
+		_, err := m.db.Exec(`
+DELETE FROM `+table+` WHERE ts >= ? AND ts < ?
+  AND ts NOT IN (
+    SELECT MIN(ts) FROM `+table+`
+    WHERE ts >= ? AND ts < ?
+    GROUP BY ts / 60
+  )`, start, end, start, end)
+		if err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
+// Prune deletes all rows older than the given cutoff from every metrics table.
+// Called hourly by the metrics collector to keep the DB size bounded.
+func (m *MetricsDB) Prune(before time.Time) error {
+	if m == nil || m.db == nil {
+		return nil
+	}
+	cutTS := before.Unix()
+	for _, table := range []string{"sys_metrics", "gpu_metrics", "fan_metrics", "temp_metrics"} {
+		if _, err := m.db.Exec("DELETE FROM "+table+" WHERE ts < ?", cutTS); err != nil {
+			return err
+		}
+	}
+	_, _ = m.db.Exec("PRAGMA wal_checkpoint(TRUNCATE)")
+	return nil
+}
+
 // LoadRecent returns up to n samples in chronological order (oldest first).
 func (m *MetricsDB) LoadRecent(n int) ([]platform.LiveMetricSample, error) {
 	return m.loadSamples(`SELECT ts,cpu_load_pct,mem_load_pct,power_w FROM (SELECT ts,cpu_load_pct,mem_load_pct,power_w FROM sys_metrics ORDER BY ts DESC LIMIT ?) ORDER BY ts`, n)

audit/internal/webui/pages.go

@@ -1378,15 +1378,64 @@ setInterval(loadMetricsLayout, 5000);
 // ── Validate (Acceptance Tests) ───────────────────────────────────────────────
 
 type validateInventory struct {
-	CPU     string
-	Memory  string
-	Storage string
-	NVIDIA  string
-	AMD     string
+	CPU           string
+	Memory        string
+	Storage       string
+	NVIDIA        string
+	AMD           string
+	NvidiaGPUCount int
+	AMDGPUCount    int
+}
+
+// validateFmtDur formats a duration in seconds as a human-readable "~N min" or "~N s" string.
+func validateFmtDur(secs int) string {
+	if secs < 120 {
+		return fmt.Sprintf("~%d s", secs)
+	}
+	mins := (secs + 29) / 60
+	return fmt.Sprintf("~%d min", mins)
+}
+
+// validateTotalValidateSec returns the estimated wall-clock duration of
+// "Validate one by one" in Validate mode for n NVIDIA GPUs.
+func validateTotalValidateSec(n int) int {
+	if n < 0 {
+		n = 0
+	}
+	total := platform.SATEstimatedCPUValidateSec +
+		platform.SATEstimatedMemoryValidateSec +
+		n*platform.SATEstimatedNvidiaGPUValidatePerGPUSec +
+		platform.SATEstimatedNvidiaInterconnectSec +
+		platform.SATEstimatedNvidiaBandwidthSec
+	return total
+}
+
+// validateTotalStressSec returns the estimated wall-clock duration of
+// "Validate one by one" in Stress mode for n NVIDIA GPUs.
+func validateTotalStressSec(n int) int {
+	if n < 0 {
+		n = 0
+	}
+	total := platform.SATEstimatedCPUStressSec +
+		platform.SATEstimatedMemoryStressSec +
+		n*platform.SATEstimatedNvidiaGPUStressPerGPUSec +
+		n*platform.SATEstimatedNvidiaTargetedStressPerGPUSec +
+		n*platform.SATEstimatedNvidiaTargetedPowerPerGPUSec +
+		platform.SATEstimatedNvidiaPulseTestSec +
+		platform.SATEstimatedNvidiaInterconnectSec +
+		platform.SATEstimatedNvidiaBandwidthSec
+	return total
 }
 
 func renderValidate(opts HandlerOptions) string {
 	inv := loadValidateInventory(opts)
+	n := inv.NvidiaGPUCount
+	validateTotalStr := validateFmtDur(validateTotalValidateSec(n))
+	stressTotalStr := validateFmtDur(validateTotalStressSec(n))
+	gpuNote := ""
+	if n > 0 {
+		gpuNote = fmt.Sprintf(" (%d GPU)", n)
+	}
 	return `<div class="alert alert-info" style="margin-bottom:16px"><strong>Non-destructive:</strong> Validate tests collect diagnostics only. They do not write to disks, do not run sustained load, and do not increment hardware wear counters.</div>
 <p style="color:var(--muted);font-size:13px;margin-bottom:16px">Tasks continue in the background — view progress in <a href="/tasks">Tasks</a>.</p>
 
@@ -1396,10 +1445,10 @@ func renderValidate(opts HandlerOptions) string {
 	    <div class="validate-profile-col">
 	      <div class="form-row" style="margin:12px 0 0"><label>Mode</label></div>
 	      <label class="cb-row"><input type="radio" name="sat-mode" id="sat-mode-validate" value="validate" checked onchange="satModeChanged()"><span>Validate — quick non-destructive check</span></label>
-	      <label class="cb-row"><input type="radio" name="sat-mode" id="sat-mode-stress" value="stress" onchange="satModeChanged()"><span>Stress — thorough load test (~30–60 min)</span></label>
+	      <label class="cb-row"><input type="radio" name="sat-mode" id="sat-mode-stress" value="stress" onchange="satModeChanged()"><span>Stress — thorough load test (` + stressTotalStr + gpuNote + `)</span></label>
 	    </div>
 	    <div class="validate-profile-col validate-profile-action">
-	      <p style="color:var(--muted);font-size:12px;margin:0 0 10px">Runs validate modules sequentially with the selected cycle count and mode. Validate is quick (~5–15 min total); Stress is thorough (~30–60 min total).</p>
+	      <p style="color:var(--muted);font-size:12px;margin:0 0 10px">Runs validate modules sequentially. Validate: ` + validateTotalStr + gpuNote + `; Stress: ` + stressTotalStr + gpuNote + `. Estimates are based on real log data and scale with GPU count.</p>
 	      <button type="button" class="btn btn-primary" onclick="runAllSAT()">Validate one by one</button>
 	      <div style="margin-top:12px">
 	        <span id="sat-all-status" style="font-size:12px;color:var(--muted)"></span>
@@ -1413,19 +1462,19 @@ func renderValidate(opts HandlerOptions) string {
 		inv.CPU,
 		`Collects CPU inventory and temperatures, then runs a bounded CPU stress pass.`,
 		`<code>lscpu</code>, <code>sensors</code>, <code>stress-ng</code>`,
-		`60s in Validate, 30 min in Stress.`,
+		validateFmtDur(platform.SATEstimatedCPUValidateSec)+` in Validate (stress-ng 60 s). `+validateFmtDur(platform.SATEstimatedCPUStressSec)+` in Stress (stress-ng 30 min).`,
 	)) +
 		renderSATCard("memory", "Memory", "runSAT('memory')", "", renderValidateCardBody(
 			inv.Memory,
 			`Runs a RAM validation pass and records memory state around the test.`,
 			`<code>free</code>, <code>memtester</code>`,
-			`256 MB / 1 pass in Validate, 512 MB / 1 pass in Stress.`,
+			validateFmtDur(platform.SATEstimatedMemoryValidateSec)+` in Validate (256 MB × 1 pass). `+validateFmtDur(platform.SATEstimatedMemoryStressSec)+` in Stress (512 MB × 1 pass).`,
 		)) +
 		renderSATCard("storage", "Storage", "runSAT('storage')", "", renderValidateCardBody(
 			inv.Storage,
 			`Scans all storage devices and runs the matching health or self-test path for each device type.`,
 			`<code>lsblk</code>; NVMe: <code>nvme</code>; SATA/SAS: <code>smartctl</code>`,
-			`Short self-test in Validate, extended self-test in Stress.`,
+			`Seconds in Validate (NVMe: instant device query; SATA/SAS: short self-test). Up to ~1 h per device in Stress (extended self-test, device-dependent).`,
 		)) +
 		`</div>
 <div style="height:1px;background:var(--border);margin:16px 0"></div>
@@ -1450,14 +1499,33 @@ func renderValidate(opts HandlerOptions) string {
 		inv.NVIDIA,
 		`Runs NVIDIA diagnostics and board inventory checks.`,
 		`<code>nvidia-smi</code>, <code>dmidecode</code>, <code>dcgmi diag</code>`,
-		`Level 2 in Validate, Level 3 in Stress. Runs one GPU at a time on the selected NVIDIA GPUs.`,
+		func() string {
+			perV := platform.SATEstimatedNvidiaGPUValidatePerGPUSec
+			perS := platform.SATEstimatedNvidiaGPUStressPerGPUSec
+			if n > 0 {
+				return fmt.Sprintf("Validate: %s/GPU × %d = %s (Level 2, sequential). Stress: %s/GPU × %d = %s (Level 3, sequential).",
+					validateFmtDur(perV), n, validateFmtDur(perV*n),
+					validateFmtDur(perS), n, validateFmtDur(perS*n))
+			}
+			return fmt.Sprintf("Validate: %s/GPU (Level 2, sequential). Stress: %s/GPU (Level 3, sequential).",
+				validateFmtDur(perV), validateFmtDur(perS))
+		}(),
 	)) +
 		`<div id="sat-card-nvidia-targeted-stress">` +
 		renderSATCard("nvidia-targeted-stress", "NVIDIA GPU Targeted Stress", "runNvidiaValidateSet('nvidia-targeted-stress')", "", renderValidateCardBody(
 			inv.NVIDIA,
 			`Runs a controlled NVIDIA DCGM load to check stability under moderate stress.`,
 			`<code>dcgmi diag targeted_stress</code>`,
-			`Skipped in Validate mode. Runs after dcgmi diag in Stress mode. Runs one GPU at a time on the selected NVIDIA GPUs.<p id="sat-ts-mode-hint" style="color:var(--warn-fg);font-size:12px;margin:8px 0 0">Only runs in Stress mode. Switch mode above to enable in Run All.</p>`,
+			func() string {
+				per := platform.SATEstimatedNvidiaTargetedStressPerGPUSec
+				s := "Skipped in Validate. "
+				if n > 0 {
+					s += fmt.Sprintf("Stress: %s/GPU × %d = %s sequential.", validateFmtDur(per), n, validateFmtDur(per*n))
+				} else {
+					s += fmt.Sprintf("Stress: %s/GPU sequential.", validateFmtDur(per))
+				}
+				return s + `<p id="sat-ts-mode-hint" style="color:var(--warn-fg);font-size:12px;margin:8px 0 0">Only runs in Stress mode. Switch mode above to enable in Run All.</p>`
+			}(),
 		)) +
 		`</div>` +
 		`<div id="sat-card-nvidia-targeted-power">` +
@@ -1465,7 +1533,16 @@ func renderValidate(opts HandlerOptions) string {
 			inv.NVIDIA,
 			`Checks that the GPU can sustain its declared power delivery envelope. Pass/fail determined by DCGM.`,
 			`<code>dcgmi diag targeted_power</code>`,
-			`Skipped in Validate mode. Runs in Stress mode only. Runs one GPU at a time.<p id="sat-tp-mode-hint" style="color:var(--warn-fg);font-size:12px;margin:8px 0 0">Only runs in Stress mode. Switch mode above to enable in Run All.</p>`,
+			func() string {
+				per := platform.SATEstimatedNvidiaTargetedPowerPerGPUSec
+				s := "Skipped in Validate. "
+				if n > 0 {
+					s += fmt.Sprintf("Stress: %s/GPU × %d = %s sequential.", validateFmtDur(per), n, validateFmtDur(per*n))
+				} else {
+					s += fmt.Sprintf("Stress: %s/GPU sequential.", validateFmtDur(per))
+				}
+				return s + `<p id="sat-tp-mode-hint" style="color:var(--warn-fg);font-size:12px;margin:8px 0 0">Only runs in Stress mode. Switch mode above to enable in Run All.</p>`
+			}(),
 		)) +
 		`</div>` +
 		`<div id="sat-card-nvidia-pulse">` +
@@ -1473,7 +1550,7 @@ func renderValidate(opts HandlerOptions) string {
 			inv.NVIDIA,
 			`Tests power supply transient response by pulsing all GPUs simultaneously between idle and full load. Synchronous pulses across all GPUs create worst-case PSU load spikes — running per-GPU would miss PSU-level failures.`,
 			`<code>dcgmi diag pulse_test</code>`,
-			`Skipped in Validate mode. Runs in Stress mode only. Runs all selected GPUs simultaneously — synchronous pulsing is required to stress the PSU.<p id="sat-pt-mode-hint" style="color:var(--warn-fg);font-size:12px;margin:8px 0 0">Only runs in Stress mode. Switch mode above to enable in Run All.</p>`,
+			`Skipped in Validate. Stress: `+validateFmtDur(platform.SATEstimatedNvidiaPulseTestSec)+` (all GPUs simultaneously; measured on 8-GPU system).`+`<p id="sat-pt-mode-hint" style="color:var(--warn-fg);font-size:12px;margin:8px 0 0">Only runs in Stress mode. Switch mode above to enable in Run All.</p>`,
 		)) +
 		`</div>` +
 		`<div id="sat-card-nvidia-interconnect">` +
@@ -1481,7 +1558,7 @@ func renderValidate(opts HandlerOptions) string {
 			inv.NVIDIA,
 			`Verifies NVLink/NVSwitch fabric bandwidth using NCCL all_reduce_perf across all selected GPUs. Pass/fail based on achieved bandwidth vs. theoretical.`,
 			`<code>all_reduce_perf</code> (NCCL tests)`,
-			`Runs in Validate and Stress. Uses all selected GPUs simultaneously (requires ≥2) and is kept short so it fits the Validate flow.`,
+			`Validate and Stress: `+validateFmtDur(platform.SATEstimatedNvidiaInterconnectSec)+` (all GPUs simultaneously, requires ≥2).`,
 		)) +
 		`</div>` +
 		`<div id="sat-card-nvidia-bandwidth">` +
@@ -1489,7 +1566,7 @@ func renderValidate(opts HandlerOptions) string {
 			inv.NVIDIA,
 			`Validates GPU memory copy and peer-to-peer bandwidth paths using NVBandwidth.`,
 			`<code>nvbandwidth</code>`,
-			`Runs in Validate and Stress across all selected GPUs simultaneously. Intended to stay short enough for Validate.`,
+			`Validate and Stress: `+validateFmtDur(platform.SATEstimatedNvidiaBandwidthSec)+` (all GPUs simultaneously; nvbandwidth runs all built-in tests without a time limit — duration set by the tool).`,
 		)) +
 		`</div>` +
 		`</div>
@@ -1922,6 +1999,8 @@ func loadValidateInventory(opts HandlerOptions) validateInventory {
 	out.Storage = formatValidateDeviceSummary(storageTotal, storageCounts, "device")
 	out.NVIDIA = formatValidateDeviceSummary(nvidiaTotal, nvidiaCounts, "GPU")
 	out.AMD = formatValidateDeviceSummary(amdTotal, amdCounts, "GPU")
+	out.NvidiaGPUCount = nvidiaTotal
+	out.AMDGPUCount = amdTotal
 	return out
 }
 
@@ -2031,9 +2110,9 @@ func renderBenchmark(opts HandlerOptions) string {
       <div class="form-row">
         <label>Profile</label>
         <select id="benchmark-profile">
-          <option value="standard" selected>Standard — about 15 minutes</option>
-          <option value="stability">Stability — 1 to 2 hours</option>
-          <option value="overnight">Overnight — 8 hours</option>
+          <option value="standard" selected>Standard — Perf ` + validateFmtDur(platform.BenchmarkEstimatedPerfStandardSec) + ` / Power Fit ` + validateFmtDur(platform.BenchmarkEstimatedPowerStandardSec) + `</option>
+          <option value="stability">Stability — Perf ` + validateFmtDur(platform.BenchmarkEstimatedPerfStabilitySec) + ` / Power Fit ` + validateFmtDur(platform.BenchmarkEstimatedPowerStabilitySec) + `</option>
+          <option value="overnight">Overnight — Perf ` + validateFmtDur(platform.BenchmarkEstimatedPerfOvernightSec) + ` / Power Fit ` + validateFmtDur(platform.BenchmarkEstimatedPowerOvernightSec) + `</option>
         </select>
       </div>
       <div class="form-row">
@@ -2073,11 +2152,11 @@ func renderBenchmark(opts HandlerOptions) string {
     <div class="card-body">
       <p style="font-size:13px;color:var(--muted);margin-bottom:10px">The benchmark page now exposes two fundamentally different test families so compute score and server power-fit are not mixed into one number.</p>
       <table>
-        <tr><th>Run Type</th><th>Engine</th><th>Question</th></tr>
-        <tr><td>Performance Benchmark</td><td><code>bee-gpu-burn</code></td><td>How much isolated compute performance does the GPU realize in this server?</td></tr>
-        <tr><td>Power / Thermal Fit</td><td><code>dcgmi targeted_power</code></td><td>How much power per GPU can this server sustain as GPU count ramps up?</td></tr>
+        <tr><th>Run Type</th><th>Engine</th><th>Question</th><th>Standard</th><th>Stability</th></tr>
+        <tr><td>Performance Benchmark</td><td><code>bee-gpu-burn</code></td><td>How much isolated compute performance does the GPU realize in this server?</td><td>` + validateFmtDur(platform.BenchmarkEstimatedPerfStandardSec) + `</td><td>` + validateFmtDur(platform.BenchmarkEstimatedPerfStabilitySec) + `</td></tr>
+        <tr><td>Power / Thermal Fit</td><td><code>dcgmi targeted_power</code></td><td>How much power per GPU can this server sustain as GPU count ramps up?</td><td>` + validateFmtDur(platform.BenchmarkEstimatedPowerStandardSec) + `</td><td>` + validateFmtDur(platform.BenchmarkEstimatedPowerStabilitySec) + `</td></tr>
       </table>
-      <p style="font-size:12px;color:var(--muted);margin-top:10px">Use ramp-up mode for capacity work: it creates 1 GPU → 2 GPU → … → all selected steps so analysis software can derive server total score and watts-per-GPU curves.</p>
+      <p style="font-size:12px;color:var(--muted);margin-top:10px">Timings are per full ramp-up run (1 GPU → all selected), measured on 4–8 GPU servers. Use ramp-up mode for capacity work: it creates 1 GPU → 2 GPU → … → all selected steps so analysis software can derive server total score and watts-per-GPU curves.</p>
     </div>
   </div>
 </div>
@@ -2306,7 +2385,7 @@ function benchmarkRefreshResults() {
 func renderBenchmarkResultsCard(exportDir string) string {
 	maxIdx, runs := loadBenchmarkHistory(exportDir)
 	perf := renderBenchmarkResultsCardFromRuns(
-		"Performance Results",
+		"Perf Results",
 		"Composite score by saved benchmark run and GPU.",
 		"No saved performance benchmark runs yet.",
 		maxIdx,
@@ -2566,13 +2645,13 @@ func renderBurn() string {
   <div class="card-body burn-profile-body">
     <div class="burn-profile-col">
       <div class="form-row" style="margin:0 0 8px"><label>Preset</label></div>
-      <label class="cb-row"><input type="radio" name="burn-profile" value="smoke" checked><span>Smoke — quick check (~5 min)</span></label>
-      <label class="cb-row"><input type="radio" name="burn-profile" value="acceptance"><span>Acceptance — 1 hour</span></label>
-      <label class="cb-row"><input type="radio" name="burn-profile" value="overnight"><span>Overnight — 8 hours</span></label>
+      <label class="cb-row"><input type="radio" name="burn-profile" value="smoke" checked><span>Smoke — 5 min/GPU (sequential) or 5 min (parallel)</span></label>
+      <label class="cb-row"><input type="radio" name="burn-profile" value="acceptance"><span>Acceptance — 1 h/GPU (sequential) or 1 h (parallel)</span></label>
+      <label class="cb-row"><input type="radio" name="burn-profile" value="overnight"><span>Overnight — 8 h/GPU (sequential) or 8 h (parallel)</span></label>
     </div>
     <div class="burn-profile-col burn-profile-action">
       <button type="button" class="btn btn-primary" onclick="runAllBurnTasks()">Burn one by one</button>
-      <p>Run checked tests one by one. Tests run without cooldown. Each test duration is determined by the Burn Profile. Total test duration is the sum of all selected tests multiplied by the Burn Profile duration.</p>
+      <p>Runs checked tests as separate sequential tasks. In sequential GPU mode, total time = profile duration × N GPU. In parallel mode, all selected GPUs burn simultaneously for one profile duration.</p>
     </div>
     <div class="burn-profile-col burn-profile-action">
       <button type="button" class="btn btn-secondary" onclick="runPlatformStress()">Thermal Cycling</button>

audit/internal/webui/server.go

@@ -135,6 +135,14 @@ type namedMetricsRing struct {
 // At metricsCollectInterval = 5 s this covers 30 minutes of live history.
 const metricsChartWindow = 360
 
+// metricsDownsampleAge is the age after which old metrics rows are downsampled
+// to 1 sample per minute. Data fresher than this is kept at full resolution.
+const metricsDownsampleAge = 2 * time.Hour
+
+// metricsRetainWindow is the total retention period for metrics rows.
+// Rows older than this are deleted entirely by the background compactor.
+const metricsRetainWindow = 48 * time.Hour
+
 var metricsCollectInterval = 5 * time.Second
 
 // pendingNetChange tracks a network state change awaiting confirmation.
@@ -335,13 +343,24 @@ func (h *handler) startMetricsCollector() {
 	goRecoverLoop("metrics collector", 2*time.Second, func() {
 		ticker := time.NewTicker(metricsCollectInterval)
 		defer ticker.Stop()
-		for range ticker.C {
-			sample := platform.SampleLiveMetrics()
-			if h.metricsDB != nil {
-				_ = h.metricsDB.Write(sample)
+		pruneTicker := time.NewTicker(time.Hour)
+		defer pruneTicker.Stop()
+		for {
+			select {
+			case <-ticker.C:
+				sample := platform.SampleLiveMetrics()
+				if h.metricsDB != nil {
+					_ = h.metricsDB.Write(sample)
+				}
+				h.feedRings(sample)
+				h.setLatestMetric(sample)
+			case <-pruneTicker.C:
+				if h.metricsDB != nil {
+					now := time.Now().UTC()
+					_ = h.metricsDB.Downsample(now.Add(-metricsDownsampleAge), now.Add(-metricsRetainWindow))
+					_ = h.metricsDB.Prune(now.Add(-metricsRetainWindow))
+				}
 			}
-			h.feedRings(sample)
-			h.setLatestMetric(sample)
 		}
 	})
 }
@@ -575,12 +594,14 @@ func (h *handler) handleMetricsChartSVG(w http.ResponseWriter, r *http.Request)
 	}
 	timeline := metricsTimelineSegments(samples, time.Now())
 	if idx, sub, ok := parseGPUChartPath(path); ok && sub == "overview" {
-		buf, ok, err := renderGPUOverviewChartSVG(idx, samples, timeline)
+		var overviewOk bool
+		var buf []byte
+		buf, overviewOk, err = renderGPUOverviewChartSVG(idx, samples, timeline)
 		if err != nil {
 			http.Error(w, err.Error(), http.StatusInternalServerError)
 			return
 		}
-		if !ok {
+		if !overviewOk {
 			http.Error(w, "metrics history unavailable", http.StatusServiceUnavailable)
 			return
 		}
@@ -589,23 +610,37 @@ func (h *handler) handleMetricsChartSVG(w http.ResponseWriter, r *http.Request)
 		_, _ = w.Write(buf)
 		return
 	}
-	datasets, names, labels, title, yMin, yMax, ok := chartDataFromSamples(path, samples)
+	datasets, names, labels, title, yMin, yMax, stacked, ok := chartDataFromSamples(path, samples)
 	if !ok {
 		http.Error(w, "metrics history unavailable", http.StatusServiceUnavailable)
 		return
 	}
 
-	buf, err := renderMetricChartSVG(
-		title,
-		labels,
-		sampleTimes(samples),
-		datasets,
-		names,
-		yMin,
-		yMax,
-		chartCanvasHeightForPath(path, len(names)),
-		timeline,
-	)
+	var buf []byte
+	if stacked {
+		buf, err = renderStackedMetricChartSVG(
+			title,
+			labels,
+			sampleTimes(samples),
+			datasets,
+			names,
+			yMax,
+			chartCanvasHeightForPath(path, len(names)),
+			timeline,
+		)
+	} else {
+		buf, err = renderMetricChartSVG(
+			title,
+			labels,
+			sampleTimes(samples),
+			datasets,
+			names,
+			yMin,
+			yMax,
+			chartCanvasHeightForPath(path, len(names)),
+			timeline,
+		)
+	}
 	if err != nil {
 		http.Error(w, err.Error(), http.StatusInternalServerError)
 		return
@@ -615,12 +650,8 @@ func (h *handler) handleMetricsChartSVG(w http.ResponseWriter, r *http.Request)
 	_, _ = w.Write(buf)
 }
 
-func chartDataFromSamples(path string, samples []platform.LiveMetricSample) ([][]float64, []string, []string, string, *float64, *float64, bool) {
-	var datasets [][]float64
-	var names []string
-	var title string
-	var yMin, yMax *float64
-	labels := sampleTimeLabels(samples)
+func chartDataFromSamples(path string, samples []platform.LiveMetricSample) (datasets [][]float64, names []string, labels []string, title string, yMin, yMax *float64, stacked bool, ok bool) {
+	labels = sampleTimeLabels(samples)
 
 	switch {
 	case path == "server-load":
@@ -656,15 +687,41 @@ func chartDataFromSamples(path string, samples []platform.LiveMetricSample) ([][
 
 	case path == "server-power":
 		title = "System Power"
-		power := make([]float64, len(samples))
-		for i, s := range samples {
-			power[i] = s.PowerW
+		// Use per-PSU stacked chart when PSU SDR data is available.
+		// Collect the union of PSU slots seen across all samples.
+		psuSlots := psuSlotsFromSamples(samples)
+		if len(psuSlots) > 1 {
+			// Build one dataset per PSU slot.
+			psuDatasets := make([][]float64, len(psuSlots))
+			psuNames := make([]string, len(psuSlots))
+			for si, slot := range psuSlots {
+				ds := make([]float64, len(samples))
+				for i, s := range samples {
+					for _, psu := range s.PSUs {
+						if psu.Slot == slot {
+							ds[i] = psu.PowerW
+							break
+						}
+					}
+				}
+				psuDatasets[si] = normalizePowerSeries(ds)
+				psuNames[si] = fmt.Sprintf("PSU %d", slot)
+			}
+			datasets = psuDatasets
+			names = psuNames
+			stacked = true
+			yMax = autoMax120(psuStackedTotal(psuDatasets))
+		} else {
+			power := make([]float64, len(samples))
+			for i, s := range samples {
+				power[i] = s.PowerW
+			}
+			power = normalizePowerSeries(power)
+			datasets = [][]float64{power}
+			names = []string{"Power W"}
+			yMin = floatPtr(0)
+			yMax = autoMax120(power)
 		}
-		power = normalizePowerSeries(power)
-		datasets = [][]float64{power}
-		names = []string{"Power W"}
-		yMin = floatPtr(0)
-		yMax = autoMax120(power)
 
 	case path == "server-fans":
 		title = "Fan RPM"
@@ -707,7 +764,7 @@ func chartDataFromSamples(path string, samples []platform.LiveMetricSample) ([][
 	case strings.HasPrefix(path, "gpu/"):
 		idx, sub, ok := parseGPUChartPath(path)
 		if !ok {
-			return nil, nil, nil, "", nil, nil, false
+			return nil, nil, nil, "", nil, nil, false, false
 		}
 		switch sub {
 		case "load":
@@ -715,7 +772,7 @@ func chartDataFromSamples(path string, samples []platform.LiveMetricSample) ([][
 			util := gpuDatasetByIndex(samples, idx, func(g platform.GPUMetricRow) float64 { return g.UsagePct })
 			mem := gpuDatasetByIndex(samples, idx, func(g platform.GPUMetricRow) float64 { return g.MemUsagePct })
 			if util == nil && mem == nil {
-				return nil, nil, nil, "", nil, nil, false
+				return nil, nil, nil, "", nil, nil, false, false
 			}
 			datasets = [][]float64{coalesceDataset(util, len(samples)), coalesceDataset(mem, len(samples))}
 			names = []string{"Load %", "Mem %"}
@@ -725,7 +782,7 @@ func chartDataFromSamples(path string, samples []platform.LiveMetricSample) ([][
 			title = gpuDisplayLabel(idx) + " Temperature"
 			temp := gpuDatasetByIndex(samples, idx, func(g platform.GPUMetricRow) float64 { return g.TempC })
 			if temp == nil {
-				return nil, nil, nil, "", nil, nil, false
+				return nil, nil, nil, "", nil, nil, false, false
 			}
 			datasets = [][]float64{temp}
 			names = []string{"Temp °C"}
@@ -735,7 +792,7 @@ func chartDataFromSamples(path string, samples []platform.LiveMetricSample) ([][
 			title = gpuDisplayLabel(idx) + " Core Clock"
 			clock := gpuDatasetByIndex(samples, idx, func(g platform.GPUMetricRow) float64 { return g.ClockMHz })
 			if clock == nil {
-				return nil, nil, nil, "", nil, nil, false
+				return nil, nil, nil, "", nil, nil, false, false
 			}
 			datasets = [][]float64{clock}
 			names = []string{"Core Clock MHz"}
@@ -744,7 +801,7 @@ func chartDataFromSamples(path string, samples []platform.LiveMetricSample) ([][
 			title = gpuDisplayLabel(idx) + " Memory Clock"
 			clock := gpuDatasetByIndex(samples, idx, func(g platform.GPUMetricRow) float64 { return g.MemClockMHz })
 			if clock == nil {
-				return nil, nil, nil, "", nil, nil, false
+				return nil, nil, nil, "", nil, nil, false, false
 			}
 			datasets = [][]float64{clock}
 			names = []string{"Memory Clock MHz"}
@@ -753,7 +810,7 @@ func chartDataFromSamples(path string, samples []platform.LiveMetricSample) ([][
 			title = gpuDisplayLabel(idx) + " Power"
 			power := gpuDatasetByIndex(samples, idx, func(g platform.GPUMetricRow) float64 { return g.PowerW })
 			if power == nil {
-				return nil, nil, nil, "", nil, nil, false
+				return nil, nil, nil, "", nil, nil, false, false
 			}
 			datasets = [][]float64{power}
 			names = []string{"Power W"}
@@ -761,10 +818,10 @@ func chartDataFromSamples(path string, samples []platform.LiveMetricSample) ([][
 		}
 
 	default:
-		return nil, nil, nil, "", nil, nil, false
+		return nil, nil, nil, "", nil, nil, false, false
 	}
 
-	return datasets, names, labels, title, yMin, yMax, len(datasets) > 0
+	return datasets, names, labels, title, yMin, yMax, stacked, len(datasets) > 0
 }
 
 func parseGPUChartPath(path string) (idx int, sub string, ok bool) {
@@ -930,6 +987,37 @@ func normalizePowerSeries(ds []float64) []float64 {
 	return out
 }
 
+// psuSlotsFromSamples returns the sorted list of PSU slot numbers seen across samples.
+func psuSlotsFromSamples(samples []platform.LiveMetricSample) []int {
+	seen := map[int]struct{}{}
+	for _, s := range samples {
+		for _, p := range s.PSUs {
+			seen[p.Slot] = struct{}{}
+		}
+	}
+	slots := make([]int, 0, len(seen))
+	for s := range seen {
+		slots = append(slots, s)
+	}
+	sort.Ints(slots)
+	return slots
+}
+
+// psuStackedTotal returns the point-by-point sum of all PSU datasets (for scale calculation).
+func psuStackedTotal(datasets [][]float64) []float64 {
+	if len(datasets) == 0 {
+		return nil
+	}
+	n := len(datasets[0])
+	total := make([]float64, n)
+	for _, ds := range datasets {
+		for i, v := range ds {
+			total[i] += v
+		}
+	}
+	return total
+}
+
 func normalizeFanSeries(ds []float64) []float64 {
 	if len(ds) == 0 {
 		return nil

audit/internal/webui/server_test.go

@@ -120,7 +120,7 @@ func TestChartDataFromSamplesUsesFullHistory(t *testing.T) {
 		},
 	}
 
-	datasets, names, labels, title, _, _, ok := chartDataFromSamples("gpu-all-power", samples)
+	datasets, names, labels, title, _, _, _, ok := chartDataFromSamples("gpu-all-power", samples)
 	if !ok {
 		t.Fatal("chartDataFromSamples returned ok=false")
 	}
@@ -164,7 +164,7 @@ func TestChartDataFromSamplesKeepsStableGPUSeriesOrder(t *testing.T) {
 		},
 	}
 
-	datasets, names, _, title, _, _, ok := chartDataFromSamples("gpu-all-power", samples)
+	datasets, names, _, title, _, _, _, ok := chartDataFromSamples("gpu-all-power", samples)
 	if !ok {
 		t.Fatal("chartDataFromSamples returned ok=false")
 	}
@@ -209,7 +209,7 @@ func TestChartDataFromSamplesIncludesGPUClockCharts(t *testing.T) {
 		},
 	}
 
-	datasets, names, _, title, _, _, ok := chartDataFromSamples("gpu-all-clock", samples)
+	datasets, names, _, title, _, _, _, ok := chartDataFromSamples("gpu-all-clock", samples)
 	if !ok {
 		t.Fatal("gpu-all-clock returned ok=false")
 	}
@@ -754,9 +754,9 @@ func TestValidatePageRendersNvidiaFabricCardsInValidateMode(t *testing.T) {
 	body := rec.Body.String()
 	for _, needle := range []string{
 		`NVIDIA Interconnect (NCCL)`,
-		`Runs in Validate and Stress.`,
+		`Validate and Stress:`,
 		`NVIDIA Bandwidth (NVBandwidth)`,
-		`Intended to stay short enough for Validate.`,
+		`nvbandwidth runs all built-in tests without a time limit`,
 	} {
 		if !strings.Contains(body, needle) {
 			t.Fatalf("validate page missing %q: %s", needle, body)

audit/internal/webui/stability.go

@@ -7,14 +7,43 @@ import (
 	"time"
 )
 
+const (
+	recoverLoopMaxDelay   = 60 * time.Second
+	recoverLoopResetAfter = 30 * time.Second
+)
+
+// goRecoverLoop starts fn in a goroutine, restarting after panics.
+// restartDelay is the initial delay; successive panics double it up to
+// recoverLoopMaxDelay. The delay resets to restartDelay once fn runs
+// successfully for recoverLoopResetAfter without panicking.
 func goRecoverLoop(name string, restartDelay time.Duration, fn func()) {
 	go func() {
+		delay := restartDelay
+		consecutive := 0
 		for {
-			if !runRecoverable(name, fn) {
+			start := time.Now()
+			panicked := runRecoverable(name, fn)
+			if !panicked {
 				return
 			}
-			if restartDelay > 0 {
-				time.Sleep(restartDelay)
+			consecutive++
+			if time.Since(start) >= recoverLoopResetAfter {
+				delay = restartDelay
+				consecutive = 1
+			}
+			slog.Warn("goroutine restarting after panic",
+				"component", name,
+				"consecutive_panics", consecutive,
+				"next_delay", delay,
+			)
+			if delay > 0 {
+				time.Sleep(delay)
+			}
+			if delay < recoverLoopMaxDelay {
+				delay *= 2
+				if delay > recoverLoopMaxDelay {
+					delay = recoverLoopMaxDelay
+				}
 			}
 		}
 	}()

audit/internal/webui/task_report.go

@@ -171,21 +171,17 @@ func renderTaskChartSVG(path string, samples []platform.LiveMetricSample, timeli
 		}
 		return gpuDisplayLabel(idx) + " Overview", buf, true
 	}
-	datasets, names, labels, title, yMin, yMax, ok := chartDataFromSamples(path, samples)
+	datasets, names, labels, title, yMin, yMax, stacked, ok := chartDataFromSamples(path, samples)
 	if !ok {
 		return "", nil, false
 	}
-	buf, err := renderMetricChartSVG(
-		title,
-		labels,
-		sampleTimes(samples),
-		datasets,
-		names,
-		yMin,
-		yMax,
-		chartCanvasHeightForPath(path, len(names)),
-		timeline,
-	)
+	var buf []byte
+	var err error
+	if stacked {
+		buf, err = renderStackedMetricChartSVG(title, labels, sampleTimes(samples), datasets, names, yMax, chartCanvasHeightForPath(path, len(names)), timeline)
+	} else {
+		buf, err = renderMetricChartSVG(title, labels, sampleTimes(samples), datasets, names, yMin, yMax, chartCanvasHeightForPath(path, len(names)), timeline)
+	}
 	if err != nil {
 		return "", nil, false
 	}

audit/internal/webui/tasks.go

@@ -585,6 +585,7 @@ func (q *taskQueue) finalizeTaskRun(t *Task, j *jobState) {
 	if err := writeTaskReportArtifacts(t); err != nil {
 		appendJobLog(t.LogPath, "WARN: task report generation failed: "+err.Error())
 	}
+	j.closeLog()
 	if t.ErrMsg != "" {
 		taskSerialEvent(t, "finished with status="+t.Status+" error="+t.ErrMsg)
 		return

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

@@ -110,8 +110,12 @@ nvidia-smi / lspci (audit collection)
 
 ---
 
-## What Needs Fixing
+## Fixed Issues
 
-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)
+All previously open items are resolved:
+
+1. **NVIDIA PCIe Model** — `enrichPCIeWithNVIDIAData()` sets `dev.Model = &v` (`nvidia.go:78`).
+2. **Fallback consistency** — `sat.go` and `benchmark_report.go` both use `"Unknown GPU"`.
+3. **`tops_per_sm_per_ghz`** — computed in `benchmark.go` and stored in `BenchmarkGPUScore.TOPSPerSMPerGHz`.
+4. **`MultiprocessorCount`, `PowerLimitW`, `DefaultPowerLimitW`** — present in `benchmark_types.go`.
+5. **Old benchmark JSONs** — no fix possible for already-saved results with missing names (display-only issue).

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

@@ -15,6 +15,41 @@ This applies to:
 - `iso/builder/config/package-lists/*.list.chroot`
 - Any package referenced in bootloader configs, hooks, or overlay scripts
 
+## Bootloader sync rule
+
+The ISO has two independent bootloader configs that must be kept in sync manually:
+
+| File | Used by |
+|------|---------|
+| `config/bootloaders/grub-efi/grub.cfg` | UEFI (all modern servers) |
+| `config/bootloaders/isolinux/live.cfg.in` | CSM / legacy BIOS (syslinux) |
+
+live-build does NOT derive one from the other. Any new boot entry, kernel parameter
+change, or new mode added to one file must be manually mirrored in the other.
+
+**Canonical entry list** (both files must have all of these):
+
+| Label | Key params |
+|-------|-----------|
+| normal (default) | `nomodeset bee.nvidia.mode=normal` + full param set |
+| load to RAM | `toram nomodeset bee.nvidia.mode=normal` + full param set |
+| GSP=off | `nomodeset bee.nvidia.mode=gsp-off` + full param set |
+| KMS | no `nomodeset`, `bee.nvidia.mode=normal` + full param set |
+| KMS + GSP=off | no `nomodeset`, `bee.nvidia.mode=gsp-off` + full param set |
+| fail-safe | `nomodeset bee.nvidia.mode=gsp-off noapic noapm nodma nomce nolapic nosmp` |
+
+**Full standard param set** (append after `@APPEND_LIVE@` / `nomodeset` flags):
+```
+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
+```
+(fail-safe is the exception — it deliberately uses minimal params.)
+
+**Historical note:** `grub-pc/` was mistakenly used instead of `grub-efi/` until v8.25.
+live-build reads `config/bootloaders/grub-efi/` for UEFI because the build is
+configured with `--bootloaders "grub-efi,syslinux"`. Directory `grub-pc` is ignored.
+
 ## Memtest rule
 
 Do not assume live-build's built-in memtest integration is sufficient for `bee`.

iso/builder/build.sh

@@ -203,7 +203,7 @@ dump_memtest_debug() {
 
         echo "-- source bootloader templates --"
         for cfg in \
-            "${BUILDER_DIR}/config/bootloaders/grub-pc/grub.cfg" \
+            "${BUILDER_DIR}/config/bootloaders/grub-efi/grub.cfg" \
             "${BUILDER_DIR}/config/bootloaders/isolinux/live.cfg.in"; do
             if [ -f "$cfg" ]; then
                 echo "  file: $cfg"
@@ -954,86 +954,6 @@ elif [ -d "${LB_PKG_CACHE}" ] && [ "$(ls -A "${LB_PKG_CACHE}" 2>/dev/null)" ]; t
     rsync -a "${LB_PKG_CACHE}/" "${BUILD_WORK_DIR}/cache/packages.chroot/"
 fi
 
-if [ "$BEE_GPU_VENDOR" != "nvidia" ] || [ "$BEE_NVIDIA_MODULE_FLAVOR" != "proprietary" ]; then
-    cat > "${BUILD_WORK_DIR}/config/bootloaders/grub-pc/grub.cfg" <<'EOF'
-source /boot/grub/config.cfg
-
-echo ""
-echo "  ███████╗ █████╗ ███████╗██╗   ██╗      ██████╗ ███████╗███████╗"
-echo "  ██╔════╝██╔══██╗██╔════╝╚██╗ ██╔╝      ██╔══██╗██╔════╝██╔════╝"
-echo "  █████╗  ███████║███████╗ ╚████╔╝ █████╗██████╔╝█████╗  █████╗"
-echo "  ██╔══╝  ██╔══██║╚════██║  ╚██╔╝  ╚════╝██╔══██╗██╔══╝  ██╔══╝"
-echo "  ███████╗██║  ██║███████║   ██║         ██████╔╝███████╗███████╗"
-echo "  ╚══════╝╚═╝  ╚═╝╚══════╝   ╚═╝         ╚═════╝ ╚══════╝╚══════╝"
-echo "  Hardware Audit LiveCD"
-echo ""
-
-menuentry "EASY-BEE" {
-    linux   @KERNEL_LIVE@ @APPEND_LIVE@ nomodeset net.ifnames=0 biosdevname=0 mitigations=off transparent_hugepage=always numa_balancing=disable nowatchdog nosoftlockup
-    initrd  @INITRD_LIVE@
-}
-
-submenu "EASY-BEE (advanced options) -->" {
-    menuentry "EASY-BEE — KMS (no nomodeset)" {
-        linux   @KERNEL_LIVE@ @APPEND_LIVE@ net.ifnames=0 biosdevname=0 mitigations=off transparent_hugepage=always numa_balancing=disable nowatchdog nosoftlockup
-        initrd  @INITRD_LIVE@
-    }
-
-    menuentry "EASY-BEE — fail-safe" {
-        linux   @KERNEL_LIVE@ @APPEND_LIVE@ nomodeset noapic noapm nodma nomce nolapic nosmp vga=normal net.ifnames=0 biosdevname=0
-        initrd  @INITRD_LIVE@
-    }
-}
-
-if [ "${grub_platform}" = "efi" ]; then
-    menuentry "Memory Test (memtest86+)" {
-        chainloader /boot/memtest86+x64.efi
-    }
-else
-    menuentry "Memory Test (memtest86+)" {
-        linux16 /boot/memtest86+x64.bin
-    }
-fi
-
-if [ "${grub_platform}" = "efi" ]; then
-    menuentry "UEFI Firmware Settings" {
-        fwsetup
-    }
-fi
-EOF
-
-    cat > "${BUILD_WORK_DIR}/config/bootloaders/isolinux/live.cfg.in" <<'EOF'
-label live-@FLAVOUR@-normal
-    menu label ^EASY-BEE
-    menu default
-    linux @LINUX@
-    initrd @INITRD@
-    append @APPEND_LIVE@
-
-label live-@FLAVOUR@-kms
-    menu label EASY-BEE (^graphics/KMS)
-    linux @LINUX@
-    initrd @INITRD@
-    append @APPEND_LIVE@ bee.display=kms
-
-label live-@FLAVOUR@-toram
-    menu label EASY-BEE (^load to RAM)
-    linux @LINUX@
-    initrd @INITRD@
-    append @APPEND_LIVE@ toram
-
-label live-@FLAVOUR@-failsafe
-    menu label EASY-BEE (^fail-safe)
-    linux @LINUX@
-    initrd @INITRD@
-    append @APPEND_LIVE@ memtest noapic noapm nodma nomce nolapic nosmp vga=normal
-
-label memtest
-    menu label ^Memory Test (memtest86+)
-    linux /boot/memtest86+x64.bin
-EOF
-fi
-
 rsync -a "${OVERLAY_DIR}/" "${OVERLAY_STAGE_DIR}/"
 rm -f \
     "${OVERLAY_STAGE_DIR}/etc/bee-ssh-password-fallback" \

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

@@ -31,6 +31,11 @@ submenu "EASY-BEE (advanced options) -->" {
         initrd  @INITRD_LIVE@
     }
 
+    menuentry "EASY-BEE — KMS + GSP=off" {
+        linux   @KERNEL_LIVE@ @APPEND_LIVE@ 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 — fail-safe" {
         linux   @KERNEL_LIVE@ @APPEND_LIVE@ nomodeset bee.nvidia.mode=gsp-off noapic noapm nodma nomce nolapic nosmp vga=normal net.ifnames=0 biosdevname=0
         initrd  @INITRD_LIVE@

iso/builder/config/bootloaders/grub-efi/live-theme/theme.txt

@@ -5,6 +5,15 @@ title-text: ""
 message-font: "Unifont Regular 16"
 terminal-font: "Unifont Regular 16"
 
+#bee logo — centered, upper third of screen
++ image {
+        top = 4%
+        left = 50%-200
+        width = 400
+        height = 400
+        file = "bee-logo.png"
+}
+
 #help bar at the bottom
 + label {
         top = 100%-50
@@ -21,8 +30,8 @@ terminal-font: "Unifont Regular 16"
 + boot_menu {
         left = 20%
         width = 60%
-        top = 62%
-        height = 38%-80
+        top = 65%
+        height = 35%-80
         item_color = "#c88000"
         item_font = "Unifont Regular 16"
         selected_item_color= "#f5a800"

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

@@ -3,37 +3,37 @@ label live-@FLAVOUR@-normal
     menu default
     linux @LINUX@
     initrd @INITRD@
-    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 pcie_aspm=off intel_idle.max_cstate=1 processor.max_cstate=1
+    append @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
 
 label live-@FLAVOUR@-toram
     menu label EASY-BEE (^load to RAM)
     linux @LINUX@
     initrd @INITRD@
-    append @APPEND_LIVE@ toram bee.nvidia.mode=normal pcie_aspm=off intel_idle.max_cstate=1 processor.max_cstate=1
+    append @APPEND_LIVE@ toram 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
 
 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 pcie_aspm=off intel_idle.max_cstate=1 processor.max_cstate=1
+    append @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
 
-label live-@FLAVOUR@-kms-gsp-off
-    menu label EASY-BEE (g^raphics/KMS, GSP=off)
+label live-@FLAVOUR@-kms
+    menu label EASY-BEE (^KMS, no nomodeset)
     linux @LINUX@
     initrd @INITRD@
-    append @APPEND_LIVE@ bee.display=kms bee.nvidia.mode=gsp-off pcie_aspm=off intel_idle.max_cstate=1 processor.max_cstate=1
+    append @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
+
+label live-@FLAVOUR@-kms-gsp-off
+    menu label EASY-BEE (KMS, ^GSP=off)
+    linux @LINUX@
+    initrd @INITRD@
+    append @APPEND_LIVE@ 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
 
 label live-@FLAVOUR@-failsafe
     menu label EASY-BEE (^fail-safe)
     linux @LINUX@
     initrd @INITRD@
-    append @APPEND_LIVE@ bee.nvidia.mode=gsp-off memtest noapic noapm nodma nomce nolapic nosmp vga=normal
+    append @APPEND_LIVE@ nomodeset bee.nvidia.mode=gsp-off noapic noapm nodma nomce nolapic nosmp vga=normal net.ifnames=0 biosdevname=0
 
 label memtest
     menu label ^Memory Test (memtest86+)

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

@@ -10,6 +10,7 @@ RestartSec=3
 StandardOutput=journal
 StandardError=journal
 LimitMEMLOCK=infinity
+MemoryMax=3G
 # Keep the web server responsive during GPU/CPU stress (children inherit nice+10
 # via Setpriority in runCmdJob, but the bee-web parent stays at 0).
 Nice=0