Fix memtest hook: bad ver_arg format in apt-get download

ver_arg was set to "=memtest86+=VERSION" making the command
"apt-get download memtest86+=memtest86+=VERSION" (invalid).
Fixed to build pkg_spec directly as "memtest86+=VERSION".
Also add apt-get update retry if initial download fails.

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

audit/internal/platform/benchmark.go

@@ -49,6 +49,10 @@ type benchmarkPowerCalibrationResult struct {
 	Derated            bool
 	Completed          bool
 	Notes              []string
+	// CoolingWarning is set when the GPU throttled thermally with a clock drop
+	// ≥20% while server fans were below 100% duty cycle — a signal that the
+	// cooling system may not be correctly configured for full GPU load.
+	CoolingWarning string
 }
 
 type benchmarkBurnProfile struct {
@@ -344,6 +348,9 @@ func (s *System) RunNvidiaBenchmark(ctx context.Context, baseDir string, opts Nv
 				gpuResult.PowerCalibrationTries = calib.Attempts
 				gpuResult.PowerLimitDerated = calib.Derated
 				gpuResult.Notes = append(gpuResult.Notes, calib.Notes...)
+				if calib.CoolingWarning != "" {
+					gpuResult.CoolingWarning = calib.CoolingWarning
+				}
 			}
 			if norm := findBenchmarkNormalization(result.Normalization.GPUs, idx); norm != nil {
 				gpuResult.LockedGraphicsClockMHz = norm.GPUClockLockMHz
@@ -1625,7 +1632,15 @@ func buildBenchmarkFindings(result NvidiaBenchmarkResult) []string {
 			case "power_capped":
 				findings = append(findings, fmt.Sprintf("GPU %d spent measurable time under SW power cap.", gpu.Index))
 			case "thermal_limited":
-				findings = append(findings, fmt.Sprintf("GPU %d reported thermal slowdown during steady state.", gpu.Index))
+				msg := fmt.Sprintf("GPU %d reported thermal slowdown during steady state.", gpu.Index)
+				if result.Cooling != nil && result.Cooling.FanDutyCycleAvailable &&
+					result.Cooling.P95FanDutyCyclePct < 98 && gpu.Steady.ClockDriftPct >= 20 {
+					msg += fmt.Sprintf(
+						" Fans peaked at %.0f%% duty cycle (not at maximum) while clocks dropped %.0f%% — possible cooling misconfiguration; rerun the benchmark with fan speed manually fixed at 100%%.",
+						result.Cooling.P95FanDutyCyclePct, gpu.Steady.ClockDriftPct,
+					)
+				}
+				findings = append(findings, msg)
 			case "sync_boost_limited":
 				findings = append(findings, fmt.Sprintf("GPU %d was limited by sync boost behaviour.", gpu.Index))
 			case "low_sm_clock_vs_target":
@@ -1642,6 +1657,12 @@ func buildBenchmarkFindings(result NvidiaBenchmarkResult) []string {
 				findings = append(findings, fmt.Sprintf("GPU %d reported %d corrected ECC error(s) — possible DRAM degradation.", gpu.Index, gpu.ECC.Corrected))
 			}
 		}
+		if gpu.CoolingWarning != "" {
+			findings = append(findings, fmt.Sprintf(
+				"GPU %d: %s. Operator action: rerun the benchmark with fan speed manually fixed at 100%% to confirm actual thermal headroom.",
+				gpu.Index, gpu.CoolingWarning,
+			))
+		}
 		if len(gpu.PrecisionFailures) > 0 {
 			findings = append(findings, fmt.Sprintf("GPU %d had incomplete precision coverage: %s.", gpu.Index, strings.Join(gpu.PrecisionFailures, ", ")))
 		}
@@ -2044,6 +2065,9 @@ func runNvidiaBenchmarkParallel(
 			r.PowerCalibrationTries = calib.Attempts
 			r.PowerLimitDerated = calib.Derated
 			r.Notes = append(r.Notes, calib.Notes...)
+			if calib.CoolingWarning != "" {
+				r.CoolingWarning = calib.CoolingWarning
+			}
 		}
 		if norm := findBenchmarkNormalization(result.Normalization.GPUs, idx); norm != nil {
 			r.LockedGraphicsClockMHz = norm.GPUClockLockMHz
@@ -2448,8 +2472,10 @@ func runBenchmarkPowerCalibration(
 	logFunc func(string),
 ) (map[int]benchmarkPowerCalibrationResult, []benchmarkRestoreAction) {
 	const calibDurationSec = 120
-	const derateStepW = 25
 	const maxDerateW = 150
+	// calibSearchTolerance is the binary-search convergence threshold in watts.
+	// When hi-lo ≤ this, the highest verified-stable limit (lo) is used.
+	const calibSearchTolerance = 10
 	// dcgmResourceBusyMaxDelaySec caps the exponential back-off when DCGM
 	// returns DCGM_ST_IN_USE (exit 222). The sequence is 1 s, 2 s, 4 s, …
 	// doubling each retry until it would exceed the cap, at which point the
@@ -2472,8 +2498,25 @@ func runBenchmarkPowerCalibration(
 		err  error
 	}
 
+
+	// gpuCalibState holds per-GPU binary search state during parallel calibration.
+	type gpuCalibState struct {
+		idx            int
+		info           benchmarkGPUInfo
+		originalLimitW int
+		appliedLimitW  int
+		minLimitW      int
+		lo             int // highest verified-stable limit (assumed: minLimitW)
+		hi             int // lowest verified-unstable limit (exclusive sentinel above start)
+		calib          benchmarkPowerCalibrationResult
+		converged      bool
+	}
+
 	results := make(map[int]benchmarkPowerCalibrationResult, len(gpuIndices))
 	var restore []benchmarkRestoreAction
+
+	// Initialise per-GPU state.
+	states := make([]*gpuCalibState, 0, len(gpuIndices))
 	for _, idx := range gpuIndices {
 		info := infoByIndex[idx]
 		originalLimitW := int(math.Round(info.PowerLimitW))
@@ -2499,15 +2542,20 @@ func runBenchmarkPowerCalibration(
 		case appliedLimitW > 0:
 			minLimitW = appliedLimitW - maxDerateW
 		}
-		if minLimitW < derateStepW {
+		if minLimitW < calibSearchTolerance {
-			minLimitW = derateStepW
+			minLimitW = calibSearchTolerance
 		}
-
+		s := &gpuCalibState{
-		calib := benchmarkPowerCalibrationResult{
+			idx:            idx,
-			AppliedPowerLimitW: float64(appliedLimitW),
+			info:           info,
+			originalLimitW: originalLimitW,
+			appliedLimitW:  appliedLimitW,
+			minLimitW:      minLimitW,
+			lo:             minLimitW,
+			hi:             appliedLimitW + 1, // not yet tested, not yet confirmed unstable
+			calib:          benchmarkPowerCalibrationResult{AppliedPowerLimitW: float64(appliedLimitW)},
 		}
-		busyRetries := 0
+		states = append(states, s)
-		busyDelaySec := 1 // exponential back-off seed; doubles each retry up to dcgmResourceBusyMaxDelaySec
 		if canDerate && originalLimitW > 0 {
 			idxCopy := idx
 			orig := originalLimitW
@@ -2518,125 +2566,243 @@ func runBenchmarkPowerCalibration(
 				},
 			})
 		}
+	}
 
-	calibLoop:
+	// Shared DCGM resource-busy back-off state (single diagnostic session).
+	busyRetries := 0
+	busyDelaySec := 1
+	sharedAttempt := 0
+
+	type sharedAttemptResult struct {
+		out  []byte
+		rows []GPUMetricRow
+		err  error
+	}
+
+calibDone:
+	for {
+		// Collect non-converged GPUs.
+		var active []*gpuCalibState
+		for _, s := range states {
+			if !s.converged {
+				active = append(active, s)
+			}
+		}
+		if len(active) == 0 || ctx.Err() != nil {
+			break
+		}
+
+		sharedAttempt++
+		for _, s := range active {
+			s.calib.Attempts++
+			logFunc(fmt.Sprintf("power calibration: GPU %d targeted_power attempt %d at %d W for %ds", s.idx, s.calib.Attempts, s.appliedLimitW, calibDurationSec))
+		}
+
+		// Snapshot throttle counters for all active GPUs before the run.
+		beforeThrottle := make(map[int]BenchmarkThrottleCounters, len(active))
+		for _, s := range active {
+			beforeThrottle[s.idx], _ = queryThrottleCounters(s.idx)
+		}
+
+		// Run targeted_power for ALL gpuIndices simultaneously so every card
+		// is under load during calibration — this reflects real server thermals.
+		logName := fmt.Sprintf("power-calibration-attempt-%d.log", sharedAttempt)
+		cmd := nvidiaDCGMNamedDiagCommand("targeted_power", calibDurationSec, gpuIndices)
+		attemptCtx, cancelAttempt := context.WithCancel(ctx)
+		doneCh := make(chan sharedAttemptResult, 1)
+		go func() {
+			out, rows, err := runBenchmarkCommandWithMetrics(attemptCtx, verboseLog, logName, cmd, nil, gpuIndices, logFunc)
+			doneCh <- sharedAttemptResult{out: out, rows: rows, err: err}
+		}()
+
+		ticker := time.NewTicker(time.Second)
+		throttleReasons := make(map[int]string, len(active))
+		var ar sharedAttemptResult
+
+	attemptLoop:
 		for {
-			calib.Attempts++
+			select {
-			logFunc(fmt.Sprintf("power calibration: GPU %d targeted_power attempt %d at %d W for %ds", idx, calib.Attempts, appliedLimitW, calibDurationSec))
+			case ar = <-doneCh:
-
+				break attemptLoop
-			beforeThrottle, _ := queryThrottleCounters(idx)
+			case <-ticker.C:
-			attemptCtx, cancel := context.WithCancel(ctx)
+				// Poll throttle counters for each active GPU independently.
-			doneCh := make(chan calibrationAttemptResult, 1)
+				for _, s := range active {
-			logName := fmt.Sprintf("power-calibration-gpu-%d-attempt-%d.log", idx, calib.Attempts)
+					if throttleReasons[s.idx] != "" {
-			cmd := nvidiaDCGMNamedDiagCommand("targeted_power", calibDurationSec, []int{idx})
+						continue // already detected for this GPU
-			go func() {
+					}
-				out, rows, err := runBenchmarkCommandWithMetrics(attemptCtx, verboseLog, logName, cmd, nil, []int{idx}, logFunc)
+					after, err := queryThrottleCounters(s.idx)
-				doneCh <- calibrationAttemptResult{out: out, rows: rows, err: err}
-			}()
-
-			ticker := time.NewTicker(time.Second)
-			var (
-				attempt        calibrationAttemptResult
-				throttleReason string
-			)
-		attemptLoop:
-			for {
-				select {
-				case attempt = <-doneCh:
-					break attemptLoop
-				case <-ticker.C:
-					afterThrottle, err := queryThrottleCounters(idx)
 					if err != nil {
 						continue
 					}
-					if reason := benchmarkCalibrationThrottleReason(beforeThrottle, afterThrottle); reason != "" {
+					// Record throttle but do NOT cancel — let dcgmi finish so
-						throttleReason = reason
+					// nv-hostengine releases the slot cleanly before the next attempt.
-						cancel()
+					if reason := benchmarkCalibrationThrottleReason(beforeThrottle[s.idx], after); reason != "" {
+						throttleReasons[s.idx] = reason
+						logFunc(fmt.Sprintf("power calibration: GPU %d detected %s throttle at %d W, waiting for run to finish", s.idx, reason, s.appliedLimitW))
 					}
-				case <-ctx.Done():
-					cancel()
-					attempt = <-doneCh
-					break attemptLoop
 				}
+			case <-ctx.Done():
+				cancelAttempt()
+				ar = <-doneCh
+				break attemptLoop
 			}
-			ticker.Stop()
-			cancel()
-			_ = os.WriteFile(filepath.Join(runDir, logName), attempt.out, 0644)
-
-			perGPU := filterRowsByGPU(attempt.rows, idx)
-			summary := summarizeBenchmarkTelemetry(perGPU)
-			if throttleReason == "" && attempt.err == nil && summary.P95PowerW > 0 {
-				calib.Summary = summary
-				calib.Completed = true
-				calib.AppliedPowerLimitW = float64(appliedLimitW)
-				logFunc(fmt.Sprintf("power calibration: GPU %d stable at %d W, p95=%.0f W p95_temp=%.1f C (%d samples)", idx, appliedLimitW, summary.P95PowerW, summary.P95TempC, summary.Samples))
-				break
-			}
-
-			// If DCGM reports the resource is in use, nv-hostengine has not yet
-			// released the diagnostic slot from the previous attempt. Do not
-			// derate: wait with exponential back-off and retry at the same
-			// power limit. Once the back-off delay would exceed
-			// dcgmResourceBusyMaxDelaySec, fail — the slot is persistently
-			// held by something else.
-			if attempt.err != nil && isDCGMResourceBusy(attempt.err) {
-				if busyDelaySec > dcgmResourceBusyMaxDelaySec {
-					calib.Notes = append(calib.Notes, fmt.Sprintf("DCGM resource busy after %d retries, giving up", busyRetries))
-					logFunc(fmt.Sprintf("power calibration: GPU %d DCGM resource persistently busy after %d retries, stopping", idx, busyRetries))
-					break
-				}
-				busyRetries++
-				logFunc(fmt.Sprintf("power calibration: GPU %d DCGM resource busy (attempt %d), retrying in %ds", idx, calib.Attempts, busyDelaySec))
-				select {
-				case <-ctx.Done():
-					break calibLoop
-				case <-time.After(time.Duration(busyDelaySec) * time.Second):
-				}
-				next := busyDelaySec * 2
-				if next > dcgmResourceBusyMaxDelaySec {
-					next = dcgmResourceBusyMaxDelaySec + 1 // sentinel: next busy → fail
-				}
-				busyDelaySec = next
-				continue calibLoop
-			}
-			busyRetries = 0    // reset on any non-busy outcome
-			busyDelaySec = 1   // reset back-off
-
-			switch {
-			case throttleReason != "":
-				calib.Notes = append(calib.Notes, fmt.Sprintf("targeted_power was canceled on attempt %d after %s throttling at %d W", calib.Attempts, throttleReason, appliedLimitW))
-				logFunc(fmt.Sprintf("power calibration: GPU %d throttled (%s) at %d W, reducing power limit", idx, throttleReason, appliedLimitW))
-			case attempt.err != nil:
-				calib.Notes = append(calib.Notes, fmt.Sprintf("targeted_power attempt %d failed at %d W: %v", calib.Attempts, appliedLimitW, attempt.err))
-				logFunc(fmt.Sprintf("power calibration: GPU %d targeted_power failed at %d W: %v", idx, appliedLimitW, attempt.err))
-			default:
-				calib.Notes = append(calib.Notes, fmt.Sprintf("targeted_power attempt %d at %d W produced no valid power telemetry", calib.Attempts, appliedLimitW))
-				logFunc(fmt.Sprintf("power calibration: GPU %d attempt %d at %d W produced no valid telemetry", idx, calib.Attempts, appliedLimitW))
-			}
-
-			if !canDerate || appliedLimitW <= 0 {
-				break
-			}
-			nextLimitW := appliedLimitW - derateStepW
-			if nextLimitW < minLimitW {
-				calib.Notes = append(calib.Notes, fmt.Sprintf("could not find a stable targeted_power limit within %d W of the default/current limit", maxDerateW))
-				break
-			}
-			if err := setBenchmarkPowerLimit(ctx, verboseLog, idx, nextLimitW); err != nil {
-				calib.Notes = append(calib.Notes, "failed to lower power limit: "+err.Error())
-				logFunc(fmt.Sprintf("power calibration: GPU %d failed to set reduced power limit %d W: %v", idx, nextLimitW, err))
-				break
-			}
-			appliedLimitW = nextLimitW
-			calib.AppliedPowerLimitW = float64(appliedLimitW)
-			calib.Derated = true
-			info.PowerLimitW = float64(appliedLimitW)
-			infoByIndex[idx] = info
-			calib.Notes = append(calib.Notes, fmt.Sprintf("reduced power limit to %d W and restarted targeted_power from the beginning", appliedLimitW))
 		}
+		ticker.Stop()
+		cancelAttempt()
+		_ = os.WriteFile(filepath.Join(runDir, logName), ar.out, 0644)
 
-		if calib.Completed || calib.Attempts > 0 || len(calib.Notes) > 0 {
+		// Resource busy: retry with exponential back-off (shared — one DCGM session).
-			results[idx] = calib
+		if ar.err != nil && isDCGMResourceBusy(ar.err) {
+			if busyDelaySec > dcgmResourceBusyMaxDelaySec {
+				for _, s := range active {
+					s.calib.Notes = append(s.calib.Notes, fmt.Sprintf("DCGM resource busy after %d retries, giving up", busyRetries))
+					s.converged = true
+				}
+				logFunc(fmt.Sprintf("power calibration: DCGM resource persistently busy after %d retries, stopping", busyRetries))
+				break calibDone
+			}
+			busyRetries++
+			// Undo attempt counter: busy retries don't count as real attempts.
+			for _, s := range active {
+				s.calib.Attempts--
+			}
+			logFunc(fmt.Sprintf("power calibration: DCGM resource busy (attempt %d), retrying in %ds", sharedAttempt, busyDelaySec))
+			select {
+			case <-ctx.Done():
+				break calibDone
+			case <-time.After(time.Duration(busyDelaySec) * time.Second):
+			}
+			next := busyDelaySec * 2
+			if next > dcgmResourceBusyMaxDelaySec {
+				next = dcgmResourceBusyMaxDelaySec + 1
+			}
+			busyDelaySec = next
+			sharedAttempt-- // retry same logical attempt number
+			continue
+		}
+		busyRetries = 0
+		busyDelaySec = 1
+
+		// Per-GPU analysis and binary search update.
+		for _, s := range active {
+			perGPU := filterRowsByGPU(ar.rows, s.idx)
+			summary := summarizeBenchmarkTelemetry(perGPU)
+			throttle := throttleReasons[s.idx]
+
+			// Cooling warning: thermal throttle with fans not at maximum.
+			if strings.Contains(throttle, "thermal") && s.calib.CoolingWarning == "" {
+				clocks := make([]float64, 0, len(perGPU))
+				var fanDutyValues []float64
+				fanDutyAvail := false
+				for _, r := range perGPU {
+					if r.ClockMHz > 0 {
+						clocks = append(clocks, r.ClockMHz)
+					}
+					if r.FanDutyCycleAvailable {
+						fanDutyAvail = true
+						fanDutyValues = append(fanDutyValues, r.FanDutyCyclePct)
+					}
+				}
+				dropPct := benchmarkClockDrift(clocks)
+				p95FanDuty := benchmarkPercentile(fanDutyValues, 95)
+				if dropPct >= 20 && fanDutyAvail && p95FanDuty < 98 {
+					s.calib.CoolingWarning = fmt.Sprintf(
+						"thermal throttle (%s) caused a %.0f%% clock drop while fans were at %.0f%% duty cycle — server cooling may not be configured for full GPU load",
+						throttle, dropPct, p95FanDuty,
+					)
+					logFunc(fmt.Sprintf("power calibration: GPU %d cooling warning: %s", s.idx, s.calib.CoolingWarning))
+				}
+			}
+
+			if throttle == "" && ar.err == nil && summary.P95PowerW > 0 {
+				// Stable at current limit — update lo and binary-search upward.
+				s.calib.Summary = summary
+				s.calib.Completed = true
+				s.calib.AppliedPowerLimitW = float64(s.appliedLimitW)
+				logFunc(fmt.Sprintf("power calibration: GPU %d stable at %d W, p95=%.0f W p95_temp=%.1f C (%d samples)", s.idx, s.appliedLimitW, summary.P95PowerW, summary.P95TempC, summary.Samples))
+				s.lo = s.appliedLimitW
+				if canDerate && s.hi-s.lo > calibSearchTolerance {
+					next := roundTo5W((s.lo + s.hi) / 2)
+					if next > s.lo && next < s.hi {
+						if err := setBenchmarkPowerLimit(ctx, verboseLog, s.idx, next); err == nil {
+							s.appliedLimitW = next
+							s.calib.AppliedPowerLimitW = float64(next)
+							s.calib.Completed = false // keep searching
+							s.calib.Notes = append(s.calib.Notes, fmt.Sprintf("binary search: stable at %d W, trying %d W (lo=%d hi=%d)", s.lo, next, s.lo, s.hi))
+							logFunc(fmt.Sprintf("power calibration: GPU %d binary search up: stable at %d W, trying %d W", s.idx, s.lo, next))
+							continue // next GPU in active list
+						}
+					}
+				}
+				s.converged = true
+				continue
+			}
+
+			// Failed or throttled — log and binary-search downward.
+			switch {
+			case throttle != "":
+				s.calib.Notes = append(s.calib.Notes, fmt.Sprintf("targeted_power attempt %d: %s throttle at %d W", s.calib.Attempts, throttle, s.appliedLimitW))
+				logFunc(fmt.Sprintf("power calibration: GPU %d throttled (%s) at %d W, reducing power limit", s.idx, throttle, s.appliedLimitW))
+			case ar.err != nil:
+				s.calib.Notes = append(s.calib.Notes, fmt.Sprintf("targeted_power attempt %d failed at %d W: %v", s.calib.Attempts, s.appliedLimitW, ar.err))
+				logFunc(fmt.Sprintf("power calibration: GPU %d targeted_power failed at %d W: %v", s.idx, s.appliedLimitW, ar.err))
+			default:
+				s.calib.Notes = append(s.calib.Notes, fmt.Sprintf("targeted_power attempt %d at %d W: no valid power telemetry", s.calib.Attempts, s.appliedLimitW))
+				logFunc(fmt.Sprintf("power calibration: GPU %d attempt %d at %d W: no valid telemetry", s.idx, s.calib.Attempts, s.appliedLimitW))
+			}
+
+			if !canDerate || s.appliedLimitW <= 0 {
+				s.converged = true
+				continue
+			}
+			s.hi = s.appliedLimitW
+
+			if s.hi-s.lo <= calibSearchTolerance {
+				if s.lo > s.minLimitW {
+					s.calib.Notes = append(s.calib.Notes, fmt.Sprintf("binary search converged: using %d W (lo=%d hi=%d)", s.lo, s.lo, s.hi))
+					if err := setBenchmarkPowerLimit(ctx, verboseLog, s.idx, s.lo); err == nil {
+						s.appliedLimitW = s.lo
+						s.calib.AppliedPowerLimitW = float64(s.lo)
+						s.calib.Derated = s.lo < s.originalLimitW
+					}
+				} else {
+					s.calib.Notes = append(s.calib.Notes, fmt.Sprintf("could not find a stable targeted_power limit within %d W of the default", maxDerateW))
+				}
+				s.converged = true
+				continue
+			}
+
+			next := roundTo5W((s.lo + s.hi) / 2)
+			if next <= s.lo {
+				next = s.lo + calibSearchTolerance
+			}
+			if next >= s.hi {
+				next = (s.lo + s.hi) / 2
+			}
+			if next < s.minLimitW {
+				s.calib.Notes = append(s.calib.Notes, fmt.Sprintf("could not find a stable targeted_power limit within %d W of the default", maxDerateW))
+				s.converged = true
+				continue
+			}
+			if err := setBenchmarkPowerLimit(ctx, verboseLog, s.idx, next); err != nil {
+				s.calib.Notes = append(s.calib.Notes, "failed to set power limit: "+err.Error())
+				logFunc(fmt.Sprintf("power calibration: GPU %d failed to set power limit %d W: %v", s.idx, next, err))
+				s.converged = true
+				continue
+			}
+			s.appliedLimitW = next
+			s.calib.AppliedPowerLimitW = float64(next)
+			s.calib.Derated = next < s.originalLimitW
+			s.info.PowerLimitW = float64(next)
+			infoByIndex[s.idx] = s.info
+			s.calib.Notes = append(s.calib.Notes, fmt.Sprintf("binary search: trying %d W (lo=%d hi=%d)", next, s.lo, s.hi))
+			logFunc(fmt.Sprintf("power calibration: GPU %d binary search: trying %d W (lo=%d hi=%d)", s.idx, next, s.lo, s.hi))
+		}
+	}
+
+	for _, s := range states {
+		if s.calib.Completed || s.calib.Attempts > 0 || len(s.calib.Notes) > 0 {
+			results[s.idx] = s.calib
 		}
 	}
 	return results, restore
@@ -2649,6 +2815,11 @@ func isDCGMResourceBusy(err error) bool {
 	return errors.As(err, &exitErr) && exitErr.ExitCode() == 222
 }
 
+// roundTo5W rounds w to the nearest 5 W boundary.
+func roundTo5W(w int) int {
+	return ((w + 2) / 5) * 5
+}
+
 func powerBenchDurationSec(profile string) int {
 	switch strings.TrimSpace(strings.ToLower(profile)) {
 	case NvidiaBenchmarkProfileStability:
@@ -2660,15 +2831,6 @@ func powerBenchDurationSec(profile string) int {
 	}
 }
 
-func occupiedSlots(indices []int, current int) []int {
-	out := make([]int, 0, len(indices))
-	for _, idx := range indices {
-		if idx != current {
-			out = append(out, idx)
-		}
-	}
-	return out
-}
 
 func cloneBenchmarkGPUInfoMap(src map[int]benchmarkGPUInfo) map[int]benchmarkGPUInfo {
 	out := make(map[int]benchmarkGPUInfo, len(src))
@@ -2716,9 +2878,7 @@ func renderPowerBenchReport(result NvidiaPowerBenchResult) string {
 	b.WriteString("\n")
 	for _, gpu := range result.GPUs {
 		fmt.Fprintf(&b, "### GPU %d — %s\n\n", gpu.Index, gpu.Name)
-		if gpu.OccupiedSlotsNote != "" {
+
-			fmt.Fprintf(&b, "- %s\n", gpu.OccupiedSlotsNote)
-		}
 		for _, note := range gpu.Notes {
 			fmt.Fprintf(&b, "- %s\n", note)
 		}
@@ -2784,10 +2944,24 @@ func (s *System) RunNvidiaPowerBench(ctx context.Context, baseDir string, opts N
 	}
 	durationSec := powerBenchDurationSec(opts.Profile)
 	_ = durationSec
-	calibByIndex, restoreActions := runBenchmarkPowerCalibration(ctx, verboseLog, runDir, selected, infoByIndex, logFunc)
+	// Phase 1: calibrate each GPU individually (sequentially, one at a time) to
+	// establish a true single-card power baseline unaffected by neighbour heat.
+	calibByIndex := make(map[int]benchmarkPowerCalibrationResult, len(selected))
+	var allRestoreActions []benchmarkRestoreAction
+	for _, idx := range selected {
+		singleDir := filepath.Join(runDir, fmt.Sprintf("single-%02d", idx))
+		_ = os.MkdirAll(singleDir, 0755)
+		singleInfo := cloneBenchmarkGPUInfoMap(infoByIndex)
+		logFunc(fmt.Sprintf("power calibration: GPU %d single-card baseline", idx))
+		c, restore := runBenchmarkPowerCalibration(ctx, verboseLog, singleDir, []int{idx}, singleInfo, logFunc)
+		allRestoreActions = append(allRestoreActions, restore...)
+		if r, ok := c[idx]; ok {
+			calibByIndex[idx] = r
+		}
+	}
 	defer func() {
-		for i := len(restoreActions) - 1; i >= 0; i-- {
+		for i := len(allRestoreActions) - 1; i >= 0; i-- {
-			restoreActions[i].fn()
+			allRestoreActions[i].fn()
 		}
 	}()
 	gpus := make([]NvidiaPowerBenchGPU, 0, len(selected))
@@ -2804,11 +2978,6 @@ func (s *System) RunNvidiaPowerBench(ctx context.Context, baseDir string, opts N
 				result.OverallStatus = "PARTIAL"
 			}
 		}
-		occupied := occupiedSlots(selected, idx)
-		note := ""
-		if len(occupied) > 0 {
-			note = fmt.Sprintf("Slot recommendation was measured while slots %s were populated; airflow in a different chassis fill pattern may differ.", joinIndexList(occupied))
-		}
 		gpus = append(gpus, NvidiaPowerBenchGPU{
 			Index:               idx,
 			Name:                info.Name,
@@ -2820,9 +2989,8 @@ func (s *System) RunNvidiaPowerBench(ctx context.Context, baseDir string, opts N
 			CalibrationAttempts: calib.Attempts,
 			Derated:             calib.Derated,
 			Status:              status,
-			OccupiedSlots:       occupied,
-			OccupiedSlotsNote:   note,
 			Notes:               append([]string(nil), calib.Notes...),
+			CoolingWarning:      calib.CoolingWarning,
 		})
 	}
 	sort.Slice(gpus, func(i, j int) bool {
@@ -2849,19 +3017,37 @@ func (s *System) RunNvidiaPowerBench(ctx context.Context, baseDir string, opts N
 		if gpu.Derated {
 			result.Findings = append(result.Findings, fmt.Sprintf("GPU %d required reduced power limit %.0f W to complete targeted_power.", gpu.Index, gpu.AppliedPowerLimitW))
 		}
+		if gpu.CoolingWarning != "" {
+			result.Findings = append(result.Findings, fmt.Sprintf(
+				"GPU %d: %s. Operator action: rerun the benchmark with fan speed manually fixed at 100%% to confirm actual thermal headroom.",
+				gpu.Index, gpu.CoolingWarning,
+			))
+		}
 	}
 	singleByIndex := make(map[int]NvidiaPowerBenchGPU, len(gpus))
 	for _, gpu := range gpus {
 		singleByIndex[gpu.Index] = gpu
 	}
+
+	// Phase 2: ramp — add one GPU per step and calibrate the growing subset
+	// simultaneously. Step 1 reuses single-card results; steps 2..N run fresh
+	// targeted_power with derating if degradation is detected.
 	for step := 1; step <= len(result.RecommendedSlotOrder); step++ {
 		subset := append([]int(nil), result.RecommendedSlotOrder[:step]...)
 		stepDir := filepath.Join(runDir, fmt.Sprintf("step-%02d", step))
 		_ = os.MkdirAll(stepDir, 0755)
-		stepInfo := cloneBenchmarkGPUInfoMap(infoByIndex)
+		var stepCalib map[int]benchmarkPowerCalibrationResult
-		stepCalib, stepRestore := runBenchmarkPowerCalibration(ctx, verboseLog, stepDir, subset, stepInfo, logFunc)
+		if step == 1 {
-		for i := len(stepRestore) - 1; i >= 0; i-- {
+			// Single-GPU step — already measured in phase 1; reuse directly.
-			stepRestore[i].fn()
+			stepCalib = calibByIndex
+			logFunc(fmt.Sprintf("power ramp: step 1/%d — reusing single-card calibration for GPU %d", len(result.RecommendedSlotOrder), subset[0]))
+		} else {
+			stepInfo := cloneBenchmarkGPUInfoMap(infoByIndex)
+			var stepRestore []benchmarkRestoreAction
+			stepCalib, stepRestore = runBenchmarkPowerCalibration(ctx, verboseLog, stepDir, subset, stepInfo, logFunc)
+			for i := len(stepRestore) - 1; i >= 0; i-- {
+				stepRestore[i].fn()
+			}
 		}
 		ramp := NvidiaPowerBenchStep{
 			StepIndex:  step,

audit/internal/platform/benchmark_types.go

@@ -131,6 +131,9 @@ type BenchmarkGPUResult struct {
 	Scores             BenchmarkScorecard         `json:"scores"`
 	DegradationReasons []string                   `json:"degradation_reasons,omitempty"`
 	Notes              []string                   `json:"notes,omitempty"`
+	// CoolingWarning is non-empty when a thermal throttle event occurred with
+	// a clock drop ≥20% while server fans were not at 100% duty cycle.
+	CoolingWarning string `json:"cooling_warning,omitempty"`
 }
 
 type BenchmarkTelemetrySummary struct {
@@ -277,9 +280,9 @@ type NvidiaPowerBenchGPU struct {
 	CalibrationAttempts int      `json:"calibration_attempts,omitempty"`
 	Derated             bool     `json:"derated,omitempty"`
 	Status              string   `json:"status"`
-	OccupiedSlots       []int    `json:"occupied_slots,omitempty"`
-	OccupiedSlotsNote   string   `json:"occupied_slots_note,omitempty"`
 	Notes               []string `json:"notes,omitempty"`
+	// CoolingWarning mirrors BenchmarkGPUResult.CoolingWarning for the power workflow.
+	CoolingWarning string `json:"cooling_warning,omitempty"`
 }
 
 type NvidiaPowerBenchStep struct {

audit/internal/platform/sat.go

@@ -552,9 +552,13 @@ func (s *System) RunMemoryAcceptancePack(ctx context.Context, baseDir string, si
 	if passes <= 0 {
 		passes = 1
 	}
+	// Bound memtester with a hard wall-clock timeout: ~2.5 min per 100 MB per
+	// pass, plus a fixed 2-minute buffer. Without this, a stuck memory
+	// controller can cause memtester to spin forever on a single subtest.
+	timeoutSec := sizeMB*passes*150/100 + 120
 	return runAcceptancePackCtx(ctx, baseDir, "memory", []satJob{
 		{name: "01-free-before.log", cmd: []string{"free", "-h"}},
-		{name: "02-memtester.log", cmd: []string{"memtester", fmt.Sprintf("%dM", sizeMB), fmt.Sprintf("%d", passes)}},
+		{name: "02-memtester.log", cmd: []string{"timeout", fmt.Sprintf("%d", timeoutSec), "memtester", fmt.Sprintf("%dM", sizeMB), fmt.Sprintf("%d", passes)}},
 		{name: "03-free-after.log", cmd: []string{"free", "-h"}},
 	}, logFunc)
 }

audit/internal/webui/api.go

@@ -1529,6 +1529,11 @@ func (h *handler) handleAPINetworkRollback(w http.ResponseWriter, _ *http.Reques
 	writeJSON(w, map[string]string{"status": "rolled back"})
 }
 
+func (h *handler) handleAPIBenchmarkResults(w http.ResponseWriter, r *http.Request) {
+	w.Header().Set("Content-Type", "text/html; charset=utf-8")
+	fmt.Fprint(w, renderBenchmarkResultsCard(h.opts.ExportDir))
+}
+
 func (h *handler) rollbackPendingNetworkChange() error {
 	h.pendingNetMu.Lock()
 	pnc := h.pendingNet

audit/internal/webui/pages.go

@@ -2002,7 +2002,7 @@ func renderBenchmark(opts HandlerOptions) string {
   </div>
 </div>
 
-` + renderBenchmarkResultsCard(opts.ExportDir) + `
+`+`<div id="benchmark-results-section">`+renderBenchmarkResultsCard(opts.ExportDir)+`</div>`+`
 
 <div id="benchmark-output" style="display:none;margin-top:16px" class="card">
   <div class="card-head">Benchmark Output <span id="benchmark-title"></span></div>
@@ -2188,7 +2188,9 @@ function runNvidiaBenchmark(kind) {
         if (e.data) failures += 1;
         term.textContent += (e.data ? '\nERROR: ' + e.data : '\nCompleted.') + '\n';
         term.scrollTop = term.scrollHeight;
+        const isLast = (idx + 1 >= taskIds.length);
         streamNext(idx + 1, failures);
+        if (isLast) { benchmarkRefreshResults(); }
       });
       benchmarkES.onerror = function() {
         if (benchmarkES) {
@@ -2208,18 +2210,30 @@ function runNvidiaBenchmark(kind) {
 }
 
 benchmarkLoadGPUs();
+
+function benchmarkRefreshResults() {
+  fetch('/api/benchmark/results')
+    .then(function(r) { return r.text(); })
+    .then(function(html) {
+      const el = document.getElementById('benchmark-results-section');
+      if (el) el.innerHTML = html;
+    })
+    .catch(function() {});
+}
 </script>`
 }
 
 func renderBenchmarkResultsCard(exportDir string) string {
 	maxIdx, runs := loadBenchmarkHistory(exportDir)
-	return renderBenchmarkResultsCardFromRuns(
+	perf := renderBenchmarkResultsCardFromRuns(
-		"Perf Results",
+		"Performance Results",
 		"Composite score by saved benchmark run and GPU.",
-		"No saved benchmark runs yet.",
+		"No saved performance benchmark runs yet.",
 		maxIdx,
 		runs,
 	)
+	power := renderPowerBenchmarkResultsCard(exportDir)
+	return perf + "\n" + power
 }
 
 func renderBenchmarkResultsCardFromRuns(title, description, emptyMessage string, maxGPUIndex int, runs []benchmarkHistoryRun) string {
@@ -2299,6 +2313,126 @@ func loadBenchmarkHistoryFromPaths(paths []string) (int, []benchmarkHistoryRun)
 	return maxGPUIndex, runs
 }
 
+func renderPowerBenchmarkResultsCard(exportDir string) string {
+	baseDir := app.DefaultBeeBenchPowerDir
+	if strings.TrimSpace(exportDir) != "" {
+		baseDir = filepath.Join(exportDir, "bee-bench", "power")
+	}
+	paths, err := filepath.Glob(filepath.Join(baseDir, "power-*", "result.json"))
+	if err != nil || len(paths) == 0 {
+		return `<div class="card" style="margin-top:16px"><div class="card-head">Power / Thermal Fit Results</div><div class="card-body"><p style="color:var(--muted);font-size:13px">No saved power benchmark runs yet.</p></div></div>`
+	}
+	sort.Strings(paths)
+
+	type powerRun struct {
+		generatedAt time.Time
+		displayTime string
+		result      platform.NvidiaPowerBenchResult
+	}
+	var runs []powerRun
+	for _, path := range paths {
+		raw, err := os.ReadFile(path)
+		if err != nil {
+			continue
+		}
+		var r platform.NvidiaPowerBenchResult
+		if err := json.Unmarshal(raw, &r); err != nil {
+			continue
+		}
+		runs = append(runs, powerRun{
+			generatedAt: r.GeneratedAt,
+			displayTime: r.GeneratedAt.Local().Format("2006-01-02 15:04:05"),
+			result:      r,
+		})
+	}
+	sort.Slice(runs, func(i, j int) bool {
+		return runs[i].generatedAt.After(runs[j].generatedAt)
+	})
+
+	// Show only the most recent run's GPU slot table, plus a run history summary.
+	var b strings.Builder
+	b.WriteString(`<div class="card" style="margin-top:16px"><div class="card-head">Power / Thermal Fit Results</div><div class="card-body">`)
+
+	latest := runs[0].result
+	b.WriteString(`<p style="font-size:12px;color:var(--muted);margin-bottom:10px">Latest run: ` + html.EscapeString(runs[0].displayTime))
+	if latest.Hostname != "" {
+		b.WriteString(` — ` + html.EscapeString(latest.Hostname))
+	}
+	if latest.OverallStatus != "" {
+		statusColor := "var(--ok)"
+		if latest.OverallStatus != "OK" {
+			statusColor = "var(--warn)"
+		}
+		b.WriteString(` — <span style="color:` + statusColor + `;font-weight:600">` + html.EscapeString(latest.OverallStatus) + `</span>`)
+	}
+	b.WriteString(`</p>`)
+
+	if len(latest.GPUs) > 0 {
+		b.WriteString(`<div style="overflow-x:auto"><table><thead><tr>`)
+		b.WriteString(`<th>GPU</th><th>Model</th><th>Nominal W</th><th>Achieved W</th><th>P95 Observed W</th><th>Status</th>`)
+		b.WriteString(`</tr></thead><tbody>`)
+		for _, gpu := range latest.GPUs {
+			derated := gpu.Derated || (gpu.DefaultPowerLimitW > 0 && gpu.AppliedPowerLimitW < gpu.DefaultPowerLimitW-1)
+			rowStyle := ""
+			achievedStyle := ""
+			if derated {
+				rowStyle = ` style="background:rgba(255,180,0,0.08)"`
+				achievedStyle = ` style="color:#e6a000;font-weight:600"`
+			}
+			statusLabel := gpu.Status
+			if statusLabel == "" {
+				statusLabel = "OK"
+			}
+			statusColor := "var(--ok)"
+			if statusLabel != "OK" {
+				statusColor = "var(--warn)"
+			}
+			nominalStr := "-"
+			if gpu.DefaultPowerLimitW > 0 {
+				nominalStr = fmt.Sprintf("%.0f", gpu.DefaultPowerLimitW)
+			}
+			achievedStr := "-"
+			if gpu.AppliedPowerLimitW > 0 {
+				achievedStr = fmt.Sprintf("%.0f", gpu.AppliedPowerLimitW)
+			}
+			p95Str := "-"
+			if gpu.MaxObservedPowerW > 0 {
+				p95Str = fmt.Sprintf("%.0f", gpu.MaxObservedPowerW)
+			}
+			b.WriteString(`<tr` + rowStyle + `>`)
+			b.WriteString(`<td>` + strconv.Itoa(gpu.Index) + `</td>`)
+			b.WriteString(`<td>` + html.EscapeString(gpu.Name) + `</td>`)
+			b.WriteString(`<td>` + nominalStr + `</td>`)
+			b.WriteString(`<td` + achievedStyle + `>` + achievedStr + `</td>`)
+			b.WriteString(`<td>` + p95Str + `</td>`)
+			b.WriteString(`<td style="color:` + statusColor + `;font-weight:600">` + html.EscapeString(statusLabel) + `</td>`)
+			b.WriteString(`</tr>`)
+		}
+		b.WriteString(`</tbody></table></div>`)
+	}
+
+	if len(runs) > 1 {
+		b.WriteString(`<details style="margin-top:12px"><summary style="font-size:12px;color:var(--muted);cursor:pointer">` + strconv.Itoa(len(runs)) + ` runs total</summary>`)
+		b.WriteString(`<div style="overflow-x:auto;margin-top:8px"><table><thead><tr><th>#</th><th>Time</th><th>GPUs</th><th>Status</th></tr></thead><tbody>`)
+		for i, run := range runs {
+			statusColor := "var(--ok)"
+			if run.result.OverallStatus != "OK" {
+				statusColor = "var(--warn)"
+			}
+			b.WriteString(`<tr>`)
+			b.WriteString(`<td>#` + strconv.Itoa(i+1) + `</td>`)
+			b.WriteString(`<td>` + html.EscapeString(run.displayTime) + `</td>`)
+			b.WriteString(`<td>` + strconv.Itoa(len(run.result.GPUs)) + `</td>`)
+			b.WriteString(`<td style="color:` + statusColor + `;font-weight:600">` + html.EscapeString(run.result.OverallStatus) + `</td>`)
+			b.WriteString(`</tr>`)
+		}
+		b.WriteString(`</tbody></table></div></details>`)
+	}
+
+	b.WriteString(`</div></div>`)
+	return b.String()
+}
+
 // ── Burn ──────────────────────────────────────────────────────────────────────
 
 func renderBurn() string {

audit/internal/webui/server.go

@@ -263,6 +263,7 @@ func NewHandler(opts HandlerOptions) http.Handler {
 	mux.HandleFunc("POST /api/sat/abort", h.handleAPISATAbort)
 	mux.HandleFunc("POST /api/bee-bench/nvidia/perf/run", h.handleAPIBenchmarkNvidiaRunKind("nvidia-bench-perf"))
 	mux.HandleFunc("POST /api/bee-bench/nvidia/power/run", h.handleAPIBenchmarkNvidiaRunKind("nvidia-bench-power"))
+	mux.HandleFunc("GET /api/benchmark/results", h.handleAPIBenchmarkResults)
 
 	// Tasks
 	mux.HandleFunc("GET /api/tasks", h.handleAPITasksList)

iso/builder/VERSIONS

@@ -21,3 +21,4 @@ HIPBLASLT_VERSION=0.10.0.60304-76~22.04
 COMGR_VERSION=2.8.0.60304-76~22.04
 GO_VERSION=1.24.0
 AUDIT_VERSION=1.0.0
+MEMTEST_VERSION=6.10-4

iso/builder/auto/config

@@ -23,9 +23,9 @@ lb config noauto \
     --bootloaders "grub-efi,syslinux" \
     --debian-installer none \
     --archive-areas "main contrib non-free non-free-firmware" \
-    --mirror-bootstrap "https://deb.debian.org/debian" \
+    --mirror-bootstrap "http://mirror.mephi.ru/debian/" \
-    --mirror-chroot "https://deb.debian.org/debian" \
+    --mirror-chroot "http://mirror.mephi.ru/debian/" \
-    --mirror-binary "https://deb.debian.org/debian" \
+    --mirror-binary "http://mirror.mephi.ru/debian/" \
     --security true \
     --linux-flavours "amd64" \
     --linux-packages "${LB_LINUX_PACKAGES}" \

iso/builder/build-in-container.sh

@@ -161,6 +161,7 @@ run_variant() {
             -e GOMODCACHE=/cache/go-mod \
             -e TMPDIR=/cache/tmp \
             -e BEE_CACHE_DIR=/cache/bee \
+            -e BEE_REQUIRE_MEMTEST=1 \
             -w /work \
             "${IMAGE_REF}" \
             sh /work/iso/builder/build.sh --variant "${_v}" \
@@ -175,6 +176,7 @@ run_variant() {
             -e GOMODCACHE=/cache/go-mod \
             -e TMPDIR=/cache/tmp \
             -e BEE_CACHE_DIR=/cache/bee \
+            -e BEE_REQUIRE_MEMTEST=1 \
             -w /work \
             "${IMAGE_REF}" \
             sh /work/iso/builder/build.sh --variant "${_v}"

iso/builder/build.sh

@@ -57,6 +57,7 @@ OVERLAY_STAGE_DIR="${DIST_DIR}/overlay-stage-${BUILD_VARIANT}"
 export BEE_GPU_VENDOR BEE_NVIDIA_MODULE_FLAVOR BUILD_VARIANT
 
 . "${BUILDER_DIR}/VERSIONS"
+export MEMTEST_VERSION
 export PATH="$PATH:/usr/local/go/bin"
 : "${BEE_REQUIRE_MEMTEST:=0}"
 
@@ -775,6 +776,7 @@ run_optional_step_sh() {
         return 0
     fi
 
+    mkdir -p "${LOG_DIR}" 2>/dev/null || true
     step_log="${LOG_DIR}/${step_slug}.log"
     echo ""
     echo "=== optional step: ${step_name} ==="
@@ -798,13 +800,14 @@ start_build_log
 # install them on the fly so NVIDIA modules and ISO kernel always match.
 if [ -z "${DEBIAN_KERNEL_ABI}" ] || [ "${DEBIAN_KERNEL_ABI}" = "auto" ]; then
     echo "=== refreshing apt index to detect current kernel ABI ==="
-    apt-get update -qq
+    apt-get update -qq || echo "WARNING: apt-get update failed, trying cached index"
     DEBIAN_KERNEL_ABI=$(apt-cache depends linux-image-amd64 2>/dev/null \
         | awk '/Depends:.*linux-image-[0-9]/{print $2}' \
         | grep -oE '[0-9]+\.[0-9]+\.[0-9]+-[0-9]+' \
         | head -1)
     if [ -z "${DEBIAN_KERNEL_ABI}" ]; then
         echo "ERROR: could not auto-detect kernel ABI from apt-cache" >&2
+        echo "Hint: set DEBIAN_KERNEL_ABI=x.y.z-N in iso/builder/VERSIONS to skip auto-detection" >&2
         exit 1
     fi
     echo "=== kernel ABI: ${DEBIAN_KERNEL_ABI} ==="

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

@@ -5,6 +5,8 @@ set -e
 
 : "${BEE_REQUIRE_MEMTEST:=0}"
 
+# memtest86+ 6.x uses memtest86+.bin (no x64 suffix) for the BIOS binary,
+# while 5.x used memtest86+x64.bin. We normalise both to x64 names in the ISO.
 MEMTEST_FILES="memtest86+x64.bin memtest86+x64.efi"
 BINARY_BOOT_DIR="binary/boot"
 GRUB_CFG="binary/boot/grub/grub.cfg"
@@ -26,13 +28,13 @@ fail_or_warn() {
 
 copy_memtest_file() {
     src="$1"
-    base="$(basename "$src")"
+    dst_name="${2:-$(basename "$src")}"
-    dst="${BINARY_BOOT_DIR}/${base}"
+    dst="${BINARY_BOOT_DIR}/${dst_name}"
 
     [ -f "$src" ] || return 1
     mkdir -p "${BINARY_BOOT_DIR}"
     cp "$src" "$dst"
-    log "copied ${base} from ${src}"
+    log "copied ${dst_name} from ${src}"
 }
 
 extract_memtest_from_deb() {
@@ -41,14 +43,44 @@ extract_memtest_from_deb() {
 
     log "extracting memtest payload from ${deb}"
     dpkg-deb -x "$deb" "$tmpdir"
-    for f in ${MEMTEST_FILES}; do
+
-        if [ -f "${tmpdir}/boot/${f}" ]; then
+    # EFI binary: both 5.x and 6.x use memtest86+x64.efi
-            copy_memtest_file "${tmpdir}/boot/${f}"
+    if [ -f "${tmpdir}/boot/memtest86+x64.efi" ]; then
-        fi
+        copy_memtest_file "${tmpdir}/boot/memtest86+x64.efi"
-    done
+    fi
+
+    # BIOS binary: 5.x = memtest86+x64.bin, 6.x = memtest86+.bin
+    if [ -f "${tmpdir}/boot/memtest86+x64.bin" ]; then
+        copy_memtest_file "${tmpdir}/boot/memtest86+x64.bin"
+    elif [ -f "${tmpdir}/boot/memtest86+.bin" ]; then
+        copy_memtest_file "${tmpdir}/boot/memtest86+.bin" "memtest86+x64.bin"
+    fi
+
     rm -rf "$tmpdir"
 }
 
+download_and_extract_memtest() {
+    tmpdl="$(mktemp -d)"
+    if [ -n "${MEMTEST_VERSION:-}" ]; then
+        pkg_spec="memtest86+=${MEMTEST_VERSION}"
+    else
+        pkg_spec="memtest86+"
+    fi
+    log "downloading ${pkg_spec} from apt"
+    if ! ( cd "$tmpdl" && apt-get download "$pkg_spec" 2>/dev/null ); then
+        log "apt download failed, retrying after apt-get update"
+        apt-get update -qq >/dev/null 2>&1 || true
+        ( cd "$tmpdl" && apt-get download "$pkg_spec" 2>/dev/null ) || true
+    fi
+    deb="$(find "$tmpdl" -maxdepth 1 -type f -name 'memtest86+*.deb' 2>/dev/null | head -1)"
+    if [ -n "$deb" ]; then
+        extract_memtest_from_deb "$deb"
+    else
+        log "apt download of memtest86+ failed"
+    fi
+    rm -rf "$tmpdl"
+}
+
 ensure_memtest_binaries() {
     missing=0
     for f in ${MEMTEST_FILES}; do
@@ -56,10 +88,15 @@ ensure_memtest_binaries() {
     done
     [ "$missing" -eq 1 ] || return 0
 
+    # 1. Try files already placed by lb binary_memtest or chroot
     for root in chroot/boot /boot; do
         for f in ${MEMTEST_FILES}; do
             [ -f "${BINARY_BOOT_DIR}/${f}" ] || copy_memtest_file "${root}/${f}" || true
         done
+        # 6.x BIOS binary may lack x64 in name — copy with normalised name
+        if [ ! -f "${BINARY_BOOT_DIR}/memtest86+x64.bin" ]; then
+            copy_memtest_file "${root}/memtest86+.bin" "memtest86+x64.bin" || true
+        fi
     done
 
     missing=0
@@ -68,6 +105,7 @@ ensure_memtest_binaries() {
     done
     [ "$missing" -eq 1 ] || return 0
 
+    # 2. Try apt package cache (may be empty if lb binary_memtest already purged)
     for root in cache chroot/var/cache/apt/archives /var/cache/apt/archives; do
         [ -d "$root" ] || continue
         deb="$(find "$root" -type f \( -name 'memtest86+_*.deb' -o -name 'memtest86+*.deb' \) 2>/dev/null | head -1)"
@@ -76,6 +114,15 @@ ensure_memtest_binaries() {
         break
     done
 
+    missing=0
+    for f in ${MEMTEST_FILES}; do
+        [ -f "${BINARY_BOOT_DIR}/${f}" ] || missing=1
+    done
+    [ "$missing" -eq 1 ] || return 0
+
+    # 3. Fallback: download fresh from apt (lb binary_memtest purges the cache)
+    download_and_extract_memtest
+
     missing=0
     for f in ${MEMTEST_FILES}; do
         if [ ! -f "${BINARY_BOOT_DIR}/${f}" ]; then