iso: improve burn-in, export, and live boot

console=tty0 sent kernel messages to the active VT (tty1), overwriting
the TUI. Changed to console=tty2 so kernel logs land on a dedicated
console. tty1 is now clean; operator can press Alt+F2 to inspect kernel
messages and Alt+F3 for an extra shell.

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

audit/internal/app/app.go

@@ -231,8 +231,11 @@ func (a *App) ExportLatestAudit(target platform.RemovableTarget) (string, error)
 
 func (a *App) ExportLatestAuditResult(target platform.RemovableTarget) (ActionResult, error) {
 	path, err := a.ExportLatestAudit(target)
-	body := "Audit exported."
+	body := "Audit export failed."
-	if path != "" {
+	if err == nil {
+		body = "Audit exported."
+	}
+	if err == nil && path != "" {
 		body = "Audit exported to " + path
 	}
 	return ActionResult{Title: "Export audit", Body: body}, err
@@ -249,8 +252,11 @@ func (a *App) ExportSupportBundle(target platform.RemovableTarget) (string, erro
 
 func (a *App) ExportSupportBundleResult(target platform.RemovableTarget) (ActionResult, error) {
 	path, err := a.ExportSupportBundle(target)
-	body := "Support bundle exported. USB target unmounted and safe to remove."
+	body := "Support bundle export failed."
-	if path != "" {
+	if err == nil {
+		body = "Support bundle exported. USB target unmounted and safe to remove."
+	}
+	if err == nil && path != "" {
 		body = "Support bundle exported to " + path + ".\n\nUSB target unmounted and safe to remove."
 	}
 	return ActionResult{Title: "Export support bundle", Body: body}, err

audit/internal/app/app_test.go

@@ -470,6 +470,41 @@ func TestExportSupportBundleResultMentionsUnmountedUSB(t *testing.T) {
 	}
 }
 
+func TestExportSupportBundleResultDoesNotPretendSuccessOnError(t *testing.T) {
+	t.Parallel()
+
+	tmp := t.TempDir()
+	oldExportDir := DefaultExportDir
+	DefaultExportDir = tmp
+	t.Cleanup(func() { DefaultExportDir = oldExportDir })
+
+	if err := os.WriteFile(filepath.Join(tmp, "bee-audit.json"), []byte("{}\n"), 0644); err != nil {
+		t.Fatalf("write bee-audit.json: %v", err)
+	}
+	if err := os.WriteFile(filepath.Join(tmp, "bee-audit.log"), []byte("audit ok\n"), 0644); err != nil {
+		t.Fatalf("write bee-audit.log: %v", err)
+	}
+
+	a := &App{
+		exports: fakeExports{
+			exportToTargetFn: func(string, platform.RemovableTarget) (string, error) {
+				return "", errors.New("mount /dev/sda1: exFAT support is missing in this ISO build")
+			},
+		},
+	}
+
+	result, err := a.ExportSupportBundleResult(platform.RemovableTarget{Device: "/dev/sda1", FSType: "exfat"})
+	if err == nil {
+		t.Fatal("expected export error")
+	}
+	if contains(result.Body, "exported to") {
+		t.Fatalf("body should not claim success:\n%s", result.Body)
+	}
+	if result.Body != "Support bundle export failed." {
+		t.Fatalf("body=%q want %q", result.Body, "Support bundle export failed.")
+	}
+}
+
 func TestRunNvidiaAcceptancePackResult(t *testing.T) {
 	t.Parallel()
 

audit/internal/platform/export.go

@@ -11,8 +11,48 @@ import (
 
 var exportExecCommand = exec.Command
 
+func formatMountTargetError(target RemovableTarget, raw string, err error) error {
+	msg := strings.TrimSpace(raw)
+	fstype := strings.ToLower(strings.TrimSpace(target.FSType))
+	if fstype == "exfat" && strings.Contains(strings.ToLower(msg), "unknown filesystem type 'exfat'") {
+		return fmt.Errorf("mount %s: exFAT support is missing in this ISO build: %w", target.Device, err)
+	}
+	if msg == "" {
+		return err
+	}
+	return fmt.Errorf("%s: %w", msg, err)
+}
+
+func removableTargetReadOnly(fields map[string]string) bool {
+	if fields["RO"] == "1" {
+		return true
+	}
+	switch strings.ToLower(strings.TrimSpace(fields["FSTYPE"])) {
+	case "iso9660", "squashfs":
+		return true
+	default:
+		return false
+	}
+}
+
+func ensureWritableMountpoint(mountpoint string) error {
+	probe, err := os.CreateTemp(mountpoint, ".bee-write-test-*")
+	if err != nil {
+		return fmt.Errorf("target filesystem is not writable: %w", err)
+	}
+	name := probe.Name()
+	if closeErr := probe.Close(); closeErr != nil {
+		_ = os.Remove(name)
+		return closeErr
+	}
+	if err := os.Remove(name); err != nil {
+		return err
+	}
+	return nil
+}
+
 func (s *System) ListRemovableTargets() ([]RemovableTarget, error) {
-	raw, err := exportExecCommand("lsblk", "-P", "-o", "NAME,TYPE,PKNAME,RM,FSTYPE,MOUNTPOINT,SIZE,LABEL,MODEL").Output()
+	raw, err := exportExecCommand("lsblk", "-P", "-o", "NAME,TYPE,PKNAME,RM,RO,FSTYPE,MOUNTPOINT,SIZE,LABEL,MODEL").Output()
 	if err != nil {
 		return nil, err
 	}
@@ -36,7 +76,7 @@ func (s *System) ListRemovableTargets() ([]RemovableTarget, error) {
 				}
 			}
 		}
-		if !removable || fields["FSTYPE"] == "" {
+		if !removable || fields["FSTYPE"] == "" || removableTargetReadOnly(fields) {
 			continue
 		}
 
@@ -72,7 +112,7 @@ func (s *System) ExportFileToTarget(src string, target RemovableTarget) (dst str
 		}
 		if raw, err := exportExecCommand("mount", target.Device, mountpoint).CombinedOutput(); err != nil {
 			_ = os.Remove(mountpoint)
-			return string(raw), err
+			return "", formatMountTargetError(target, string(raw), err)
 		}
 		mountedHere = true
 		mounted = true
@@ -95,6 +135,10 @@ func (s *System) ExportFileToTarget(src string, target RemovableTarget) (dst str
 		}
 	}()
 
+	if err := ensureWritableMountpoint(mountpoint); err != nil {
+		return "", err
+	}
+
 	filename := filepath.Base(src)
 	dst = filepath.Join(mountpoint, filename)
 	data, err := os.ReadFile(src)

audit/internal/platform/export_test.go

@@ -4,12 +4,11 @@ import (
 	"os"
 	"os/exec"
 	"path/filepath"
+	"strings"
 	"testing"
 )
 
 func TestExportFileToTargetUnmountsExistingMountpoint(t *testing.T) {
-	t.Parallel()
-
 	tmp := t.TempDir()
 	src := filepath.Join(tmp, "bundle.tar.gz")
 	mountpoint := filepath.Join(tmp, "mnt")
@@ -54,3 +53,60 @@ func TestExportFileToTargetUnmountsExistingMountpoint(t *testing.T) {
 		t.Fatalf("expected umount %q call, got %#v", mountpoint, calls)
 	}
 }
+
+func TestExportFileToTargetRejectsNonWritableMountpoint(t *testing.T) {
+	tmp := t.TempDir()
+	src := filepath.Join(tmp, "bundle.tar.gz")
+	mountpoint := filepath.Join(tmp, "mnt")
+	if err := os.MkdirAll(mountpoint, 0755); err != nil {
+		t.Fatalf("mkdir mountpoint: %v", err)
+	}
+	if err := os.WriteFile(src, []byte("bundle"), 0644); err != nil {
+		t.Fatalf("write src: %v", err)
+	}
+	if err := os.Chmod(mountpoint, 0555); err != nil {
+		t.Fatalf("chmod mountpoint: %v", err)
+	}
+
+	oldExec := exportExecCommand
+	exportExecCommand = func(name string, args ...string) *exec.Cmd {
+		return exec.Command("sh", "-c", "exit 0")
+	}
+	t.Cleanup(func() { exportExecCommand = oldExec })
+
+	s := &System{}
+	_, err := s.ExportFileToTarget(src, RemovableTarget{
+		Device:     "/dev/sdb1",
+		Mountpoint: mountpoint,
+	})
+	if err == nil {
+		t.Fatal("expected error for non-writable mountpoint")
+	}
+	if !strings.Contains(err.Error(), "target filesystem is not writable") {
+		t.Fatalf("err=%q want writable message", err)
+	}
+}
+
+func TestListRemovableTargetsSkipsReadOnlyMedia(t *testing.T) {
+	oldExec := exportExecCommand
+	lsblkOut := `NAME="sda1" TYPE="part" PKNAME="sda" RM="1" RO="1" FSTYPE="iso9660" MOUNTPOINT="/run/live/medium" SIZE="3.7G" LABEL="BEE" MODEL=""
+NAME="sdb1" TYPE="part" PKNAME="sdb" RM="1" RO="0" FSTYPE="vfat" MOUNTPOINT="/media/bee/USB" SIZE="29.8G" LABEL="USB" MODEL=""`
+	exportExecCommand = func(name string, args ...string) *exec.Cmd {
+		cmd := exec.Command("sh", "-c", "printf '%s\n' \"$LSBLK_OUT\"")
+		cmd.Env = append(os.Environ(), "LSBLK_OUT="+lsblkOut)
+		return cmd
+	}
+	t.Cleanup(func() { exportExecCommand = oldExec })
+
+	s := &System{}
+	targets, err := s.ListRemovableTargets()
+	if err != nil {
+		t.Fatalf("ListRemovableTargets error: %v", err)
+	}
+	if len(targets) != 1 {
+		t.Fatalf("len(targets)=%d want 1 (%+v)", len(targets), targets)
+	}
+	if got := targets[0].Device; got != "/dev/sdb1" {
+		t.Fatalf("device=%q want /dev/sdb1", got)
+	}
+}

audit/internal/platform/gpu_metrics.go

@@ -69,6 +69,11 @@ func parseGPUFloat(s string) float64 {
 	return v
 }
 
+// SampleGPUMetrics runs nvidia-smi once and returns current metrics for each GPU.
+func SampleGPUMetrics(gpuIndices []int) ([]GPUMetricRow, error) {
+	return sampleGPUMetrics(gpuIndices)
+}
+
 // WriteGPUMetricsCSV writes collected rows as a CSV file.
 func WriteGPUMetricsCSV(path string, rows []GPUMetricRow) error {
 	var b bytes.Buffer
@@ -370,6 +375,162 @@ func RenderGPUTerminalChart(rows []GPUMetricRow) string {
 	return strings.TrimRight(b.String(), "\n")
 }
 
+// RenderGPULiveChart renders all GPU metrics on a single combined chart per GPU.
+// Each series is normalised to its own min–max and drawn in a different colour.
+// chartWidth controls the width of the plot area (Y-axis label uses 5 extra chars).
+func RenderGPULiveChart(rows []GPUMetricRow, chartWidth int) string {
+	if chartWidth < 20 {
+		chartWidth = 70
+	}
+	const chartHeight = 14
+
+	seen := make(map[int]bool)
+	var order []int
+	gpuMap := make(map[int][]GPUMetricRow)
+	for _, r := range rows {
+		if !seen[r.GPUIndex] {
+			seen[r.GPUIndex] = true
+			order = append(order, r.GPUIndex)
+		}
+		gpuMap[r.GPUIndex] = append(gpuMap[r.GPUIndex], r)
+	}
+
+	type seriesDef struct {
+		label string
+		color string
+		unit  string
+		fn    func(GPUMetricRow) float64
+	}
+	defs := []seriesDef{
+		{"Usage", ansiBlue, "%", func(r GPUMetricRow) float64 { return r.UsagePct }},
+		{"Temp", ansiRed, "°C", func(r GPUMetricRow) float64 { return r.TempC }},
+		{"Power", ansiGreen, "W", func(r GPUMetricRow) float64 { return r.PowerW }},
+	}
+
+	var b strings.Builder
+	for _, gpuIdx := range order {
+		gr := gpuMap[gpuIdx]
+		if len(gr) == 0 {
+			continue
+		}
+		elapsed := gr[len(gr)-1].ElapsedSec
+
+		// Build value slices for each series.
+		type seriesData struct {
+			seriesDef
+			vals []float64
+			mn   float64
+			mx   float64
+		}
+		var series []seriesData
+		for _, d := range defs {
+			vals := extractGPUField(gr, d.fn)
+			mn, mx := gpuMinMax(vals)
+			if mn == mx {
+				mx = mn + 1
+			}
+			series = append(series, seriesData{d, vals, mn, mx})
+		}
+
+		// Shared character grid: row 0 = top (max), row chartHeight = bottom (min).
+		type cell struct {
+			ch    rune
+			color string
+		}
+		grid := make([][]cell, chartHeight+1)
+		for r := range grid {
+			grid[r] = make([]cell, chartWidth)
+			for c := range grid[r] {
+				grid[r][c] = cell{' ', ""}
+			}
+		}
+
+		// Plot each series onto the shared grid.
+		for _, s := range series {
+			w := chartWidth
+			if len(s.vals) < w {
+				w = len(s.vals)
+			}
+			data := gpuDownsample(s.vals, w)
+			prevRow := -1
+			for x, v := range data {
+				row := chartHeight - int(math.Round((v-s.mn)/(s.mx-s.mn)*float64(chartHeight)))
+				if row < 0 {
+					row = 0
+				}
+				if row > chartHeight {
+					row = chartHeight
+				}
+				if prevRow < 0 || prevRow == row {
+					grid[row][x] = cell{'─', s.color}
+				} else {
+					lo, hi := prevRow, row
+					if lo > hi {
+						lo, hi = hi, lo
+					}
+					for y := lo + 1; y < hi; y++ {
+						grid[y][x] = cell{'│', s.color}
+					}
+					if prevRow < row {
+						grid[prevRow][x] = cell{'╮', s.color}
+						grid[row][x] = cell{'╰', s.color}
+					} else {
+						grid[prevRow][x] = cell{'╯', s.color}
+						grid[row][x] = cell{'╭', s.color}
+					}
+				}
+				prevRow = row
+			}
+		}
+
+		// Render: Y axis + data rows.
+		fmt.Fprintf(&b, "GPU %d  (%.0fs)  each series normalised to its range\n", gpuIdx, elapsed)
+		for r := 0; r <= chartHeight; r++ {
+			// Y axis label: 100% at top, 50% in middle, 0% at bottom.
+			switch r {
+			case 0:
+				fmt.Fprintf(&b, "%4s┤", "100%")
+			case chartHeight / 2:
+				fmt.Fprintf(&b, "%4s┤", "50%")
+			case chartHeight:
+				fmt.Fprintf(&b, "%4s┤", "0%")
+			default:
+				fmt.Fprintf(&b, "%4s│", "")
+			}
+			for c := 0; c < chartWidth; c++ {
+				cl := grid[r][c]
+				if cl.color != "" {
+					b.WriteString(cl.color)
+					b.WriteRune(cl.ch)
+					b.WriteString(ansiReset)
+				} else {
+					b.WriteRune(' ')
+				}
+			}
+			b.WriteRune('\n')
+		}
+		// Bottom axis.
+		b.WriteString("     └")
+		b.WriteString(strings.Repeat("─", chartWidth))
+		b.WriteRune('\n')
+
+		// Legend with current (last) values.
+		b.WriteString("     ")
+		for i, s := range series {
+			last := s.vals[len(s.vals)-1]
+			b.WriteString(s.color)
+			fmt.Fprintf(&b, "▐ %s: %.0f%s", s.label, last, s.unit)
+			b.WriteString(ansiReset)
+			if i < len(series)-1 {
+				b.WriteString("   ")
+			}
+		}
+		b.WriteRune('\n')
+	}
+
+	return strings.TrimRight(b.String(), "\n")
+}
+
 // renderLineChart draws a single time-series line chart using box-drawing characters.
 // Produces output in the style of asciigraph: ╭─╮ │ ╰─╯ with a Y axis and caption.
 func renderLineChart(vals []float64, color, caption string, height, width int) string {

audit/internal/tui/forms.go

@@ -151,8 +151,10 @@ func (m model) confirmCancelTarget() screen {
 	switch m.pendingAction {
 	case actionExportBundle:
 		return screenExportTargets
-	case actionRunAll, actionRunMemorySAT, actionRunStorageSAT, actionRunCPUSAT, actionRunAMDGPUSAT, actionRunFanStress:
+	case actionRunAll, actionRunMemorySAT, actionRunStorageSAT, actionRunCPUSAT, actionRunAMDGPUSAT:
 		return screenHealthCheck
+	case actionRunFanStress:
+		return screenBurnInTests
 	default:
 		return screenMain
 	}
@@ -182,12 +184,13 @@ func hcFanStressOpts(hcMode int, application interface {
 		}
 	}
 
-	// Use minimum GPU memory size to fit all GPUs.
+	// Use nearly full GPU memory on the smallest GPU (leave 512 MB for driver overhead).
 	sizeMB := 64
 	if gpus, err := application.ListNvidiaGPUs(); err == nil {
 		for _, g := range gpus {
-			if g.MemoryMB > 0 && (sizeMB == 64 || g.MemoryMB < sizeMB) {
+			free := g.MemoryMB - 512
-				sizeMB = g.MemoryMB / 16 // allocate 1/16 of VRAM per GPU
+			if free > 0 && (sizeMB == 64 || free < sizeMB) {
+				sizeMB = free
 			}
 		}
 	}

audit/internal/tui/messages.go

@@ -50,3 +50,8 @@ type gpuStressDoneMsg struct {
 	body  string
 	err   error
 }
+
+type gpuLiveTickMsg struct {
+	rows    []platform.GPUMetricRow
+	indices []int
+}

audit/internal/tui/screen_burn_in.go

@@ -0,0 +1,117 @@
+package tui
+
+import (
+	"fmt"
+	"strings"
+
+	tea "github.com/charmbracelet/bubbletea"
+)
+
+const (
+	burnCurGPUStress = 0
+	burnCurModeQuick = 1
+	burnCurModeStd   = 2
+	burnCurModeExpr  = 3
+	burnCurRun       = 4
+	burnCurTotal     = 5
+)
+
+func (m model) enterBurnInTests() (tea.Model, tea.Cmd) {
+	m.screen = screenBurnInTests
+	m.cursor = 0
+	if !m.burnInitialized {
+		m.burnMode = 0
+		m.burnCursor = 0
+		m.burnInitialized = true
+	}
+	return m, nil
+}
+
+func (m model) updateBurnInTests(msg tea.KeyMsg) (tea.Model, tea.Cmd) {
+	switch msg.String() {
+	case "up", "k":
+		if m.burnCursor > 0 {
+			m.burnCursor--
+		}
+	case "down", "j":
+		if m.burnCursor < burnCurTotal-1 {
+			m.burnCursor++
+		}
+	case " ":
+		switch m.burnCursor {
+		case burnCurModeQuick, burnCurModeStd, burnCurModeExpr:
+			m.burnMode = m.burnCursor - burnCurModeQuick
+		}
+	case "enter":
+		switch m.burnCursor {
+		case burnCurGPUStress, burnCurRun:
+			return m.burnRunSelected()
+		case burnCurModeQuick, burnCurModeStd, burnCurModeExpr:
+			m.burnMode = m.burnCursor - burnCurModeQuick
+		}
+	case "f", "F", "r", "R":
+		return m.burnRunSelected()
+	case "1":
+		m.burnMode = 0
+	case "2":
+		m.burnMode = 1
+	case "3":
+		m.burnMode = 2
+	case "esc":
+		m.screen = screenMain
+		m.cursor = 1
+	case "q", "ctrl+c":
+		return m, tea.Quit
+	}
+	return m, nil
+}
+
+func (m model) burnRunSelected() (tea.Model, tea.Cmd) {
+	return m.hcRunFanStress()
+}
+
+func renderBurnInTests(m model) string {
+	var b strings.Builder
+
+	fmt.Fprintln(&b, "BURN-IN TESTS")
+	fmt.Fprintln(&b)
+	fmt.Fprintln(&b, "  Stress tests:")
+	fmt.Fprintln(&b)
+
+	pfx := "  "
+	if m.burnCursor == burnCurGPUStress {
+		pfx = "> "
+	}
+	fmt.Fprintf(&b, "%s[ GPU PLATFORM STRESS TEST [F] ]   (thermal cycling, fan lag, throttle check)\n", pfx)
+
+	fmt.Fprintln(&b)
+	fmt.Fprintln(&b, "  Mode:")
+	modes := []struct{ label, key string }{
+		{"Quick", "1"},
+		{"Standard", "2"},
+		{"Express", "3"},
+	}
+	for i, mode := range modes {
+		pfx := "  "
+		if m.burnCursor == burnCurModeQuick+i {
+			pfx = "> "
+		}
+		radio := "( )"
+		if m.burnMode == i {
+			radio = "(*)"
+		}
+		fmt.Fprintf(&b, "%s%s  %-10s  [%s]\n", pfx, radio, mode.label, mode.key)
+	}
+
+	fmt.Fprintln(&b)
+	pfx = "  "
+	if m.burnCursor == burnCurRun {
+		pfx = "> "
+	}
+	fmt.Fprintf(&b, "%s[ RUN SELECTED [R] ]\n", pfx)
+
+	fmt.Fprintln(&b)
+	fmt.Fprintln(&b, "─────────────────────────────────────────────────────────────────")
+	fmt.Fprint(&b, "[↑↓] move  [space/enter] select  [1/2/3] mode  [R/F] run  [Esc] back")
+	return b.String()
+}

audit/internal/tui/screen_export.go

@@ -4,7 +4,12 @@ import tea "github.com/charmbracelet/bubbletea"
 
 func (m model) handleExportTargetsMenu() (tea.Model, tea.Cmd) {
 	if len(m.targets) == 0 {
-		return m, resultCmd("Export support bundle", "No removable filesystems found", nil, screenMain)
+		return m, resultCmd(
+			"Export support bundle",
+			"No writable removable filesystems found.\n\nRead-only or boot media are hidden from this list.",
+			nil,
+			screenMain,
+		)
 	}
 	target := m.targets[m.cursor]
 	m.selectedTarget = &target

audit/internal/tui/screen_health_check.go

@@ -3,8 +3,10 @@ package tui
 import (
 	"context"
 	"fmt"
-	"os/exec"
 	"strings"
+	"time"
+
+	"bee/audit/internal/platform"
 
 	tea "github.com/charmbracelet/bubbletea"
 )
@@ -28,8 +30,7 @@ const (
 	hcCurModeStd   = 6
 	hcCurModeExpr  = 7
 	hcCurRunAll    = 8
-	hcCurFanStress  = 9
+	hcCurTotal     = 9
-	hcCurTotal      = 10
 )
 
 // hcModeDurations maps mode index (0=Quick,1=Standard,2=Express) to GPU stress seconds.
@@ -84,8 +85,6 @@ func (m model) updateHealthCheck(msg tea.KeyMsg) (tea.Model, tea.Cmd) {
 			m.hcMode = m.hcCursor - hcCurModeQuick
 		case hcCurRunAll:
 			return m.hcRunAll()
-		case hcCurFanStress:
-			return m.hcRunFanStress()
 		}
 	case "g", "G":
 		return m.hcRunSingle(hcGPU)
@@ -97,8 +96,6 @@ func (m model) updateHealthCheck(msg tea.KeyMsg) (tea.Model, tea.Cmd) {
 		return m.hcRunSingle(hcCPU)
 	case "r", "R":
 		return m.hcRunAll()
-	case "f", "F":
-		return m.hcRunFanStress()
 	case "a", "A":
 		allOn := m.hcSel[0] && m.hcSel[1] && m.hcSel[2] && m.hcSel[3]
 		for i := range m.hcSel {
@@ -156,14 +153,16 @@ func (m model) hcRunFanStress() (tea.Model, tea.Cmd) {
 	return m, nil
 }
 
-// startGPUStressTest launches the GPU Platform Stress Test and nvtop concurrently.
+// startGPUStressTest launches the GPU Platform Stress Test with a live in-TUI chart.
-// nvtop occupies the full terminal as a live chart; the stress test runs in background.
 func (m model) startGPUStressTest() (tea.Model, tea.Cmd) {
-	opts := hcFanStressOpts(m.hcMode, m.app)
+	opts := hcFanStressOpts(m.burnMode, m.app)
 
 	ctx, cancel := context.WithCancel(context.Background())
 	m.gpuStressCancel = cancel
 	m.gpuStressAborted = false
+	m.gpuLiveRows = nil
+	m.gpuLiveIndices = opts.GPUIndices
+	m.gpuLiveStart = time.Now()
 	m.screen = screenGPUStressRunning
 	m.nvidiaSATCursor = 0
 
@@ -172,37 +171,29 @@ func (m model) startGPUStressTest() (tea.Model, tea.Cmd) {
 		return gpuStressDoneMsg{title: result.Title, body: result.Body, err: err}
 	}
 
-	nvtopPath, lookErr := exec.LookPath("nvtop")
+	return m, tea.Batch(stressCmd, pollGPULive(opts.GPUIndices))
-	if lookErr != nil {
+}
-		return m, stressCmd
-	}
 
-	return m, tea.Batch(
+// pollGPULive samples nvidia-smi once after one second and returns a gpuLiveTickMsg.
-		stressCmd,
+// The update handler reschedules it to achieve continuous 1s polling.
-		tea.ExecProcess(exec.Command(nvtopPath), func(_ error) tea.Msg {
+func pollGPULive(indices []int) tea.Cmd {
-			return nvtopClosedMsg{}
+	return tea.Tick(time.Second, func(_ time.Time) tea.Msg {
-		}),
+		rows, _ := platform.SampleGPUMetrics(indices)
-	)
+		return gpuLiveTickMsg{rows: rows, indices: indices}
+	})
 }
 
 // updateGPUStressRunning handles keys on the GPU stress running screen.
 func (m model) updateGPUStressRunning(msg tea.KeyMsg) (tea.Model, tea.Cmd) {
 	switch msg.String() {
-	case "o", "O":
-		nvtopPath, err := exec.LookPath("nvtop")
-		if err != nil {
-			return m, nil
-		}
-		return m, tea.ExecProcess(exec.Command(nvtopPath), func(_ error) tea.Msg {
-			return nvtopClosedMsg{}
-		})
 	case "a", "A":
 		if m.gpuStressCancel != nil {
 			m.gpuStressCancel()
 			m.gpuStressCancel = nil
 		}
 		m.gpuStressAborted = true
-		m.screen = screenHealthCheck
+		m.screen = screenBurnInTests
+		m.burnCursor = burnCurGPUStress
 		m.cursor = 0
 	case "ctrl+c":
 		return m, tea.Quit
@@ -210,8 +201,22 @@ func (m model) updateGPUStressRunning(msg tea.KeyMsg) (tea.Model, tea.Cmd) {
 	return m, nil
 }
 
-func renderGPUStressRunning() string {
+func renderGPUStressRunning(m model) string {
-	return "GPU PLATFORM STRESS TEST\n\nTest is running...\n\n[o] Open nvtop  [a] Abort test  [ctrl+c] quit\n"
+	var b strings.Builder
+	fmt.Fprintln(&b, "GPU PLATFORM STRESS TEST")
+	fmt.Fprintln(&b)
+	if len(m.gpuLiveRows) == 0 {
+		fmt.Fprintln(&b, "Collecting metrics...")
+	} else {
+		chartWidth := m.width - 8
+		if chartWidth < 40 {
+			chartWidth = 70
+		}
+		b.WriteString(platform.RenderGPULiveChart(m.gpuLiveRows, chartWidth))
+	}
+	fmt.Fprintln(&b)
+	b.WriteString("[a] Abort test  [ctrl+c] quit")
+	return b.String()
 }
 
 func (m model) hcRunAll() (tea.Model, tea.Cmd) {
@@ -371,16 +376,8 @@ func renderHealthCheck(m model) string {
 		fmt.Fprintf(&b, "%s[ RUN ALL [R] ]\n", pfx)
 	}
 
-	{
-		pfx := "  "
-		if m.hcCursor == hcCurFanStress {
-			pfx = "> "
-		}
-		fmt.Fprintf(&b, "%s[ GPU PLATFORM STRESS TEST [F] ]   (thermal cycling, fan lag, throttle check)\n", pfx)
-	}
-
 	fmt.Fprintln(&b)
 	fmt.Fprintln(&b, "─────────────────────────────────────────────────────────────────")
-	fmt.Fprint(&b, "[↑↓] move  [space/enter] toggle  [letter] single test  [R] run all  [F] gpu stress  [Esc] back")
+	fmt.Fprint(&b, "[↑↓] move  [space/enter] toggle  [letter] single test  [R] run all  [Esc] back")
 	return b.String()
 }

audit/internal/tui/screen_main.go

@@ -8,7 +8,9 @@ func (m model) handleMainMenu() (tea.Model, tea.Cmd) {
 	switch m.cursor {
 	case 0: // Health Check
 		return m.enterHealthCheck()
-	case 1: // Export support bundle
+	case 1: // Burn-in tests
+		return m.enterBurnInTests()
+	case 2: // Export support bundle
 		m.pendingAction = actionExportBundle
 		m.busy = true
 		m.busyTitle = "Export support bundle"
@@ -16,11 +18,11 @@ func (m model) handleMainMenu() (tea.Model, tea.Cmd) {
 			targets, err := m.app.ListRemovableTargets()
 			return exportTargetsMsg{targets: targets, err: err}
 		}
-	case 2: // Settings
+	case 3: // Settings
 		m.screen = screenSettings
 		m.cursor = 0
 		return m, nil
-	case 3: // Exit
+	case 4: // Exit
 		return m, tea.Quit
 	}
 	return m, nil

audit/internal/tui/tui_test.go

@@ -54,9 +54,10 @@ func TestUpdateMainMenuEnterActions(t *testing.T) {
 		wantCmd    bool
 	}{
 		{name: "health_check", cursor: 0, wantScreen: screenHealthCheck, wantCmd: true},
-		{name: "export", cursor: 1, wantScreen: screenMain, wantBusy: true, wantCmd: true},
+		{name: "burn_in_tests", cursor: 1, wantScreen: screenBurnInTests, wantCmd: true},
-		{name: "settings", cursor: 2, wantScreen: screenSettings, wantCmd: true},
+		{name: "export", cursor: 2, wantScreen: screenMain, wantBusy: true, wantCmd: true},
-		{name: "exit", cursor: 3, wantScreen: screenMain, wantCmd: true},
+		{name: "settings", cursor: 3, wantScreen: screenSettings, wantCmd: true},
+		{name: "exit", cursor: 4, wantScreen: screenMain, wantCmd: true},
 	}
 
 	for _, test := range tests {
@@ -115,7 +116,8 @@ func TestMainMenuSimpleTransitions(t *testing.T) {
 		wantScreen screen
 	}{
 		{name: "health_check", cursor: 0, wantScreen: screenHealthCheck},
-		{name: "settings", cursor: 2, wantScreen: screenSettings},
+		{name: "burn_in_tests", cursor: 1, wantScreen: screenBurnInTests},
+		{name: "settings", cursor: 3, wantScreen: screenSettings},
 	}
 
 	for _, test := range tests {
@@ -146,7 +148,7 @@ func TestMainMenuExportSetsBusy(t *testing.T) {
 	t.Parallel()
 
 	m := newTestModel()
-	m.cursor = 1 // Export support bundle
+	m.cursor = 2 // Export support bundle
 
 	next, cmd := m.handleMainMenu()
 	got := next.(model)
@@ -163,12 +165,13 @@ func TestMainViewRendersTwoColumns(t *testing.T) {
 	t.Parallel()
 
 	m := newTestModel()
-	m.cursor = 1
+	m.cursor = 2
 
 	view := m.View()
 	for _, want := range []string{
 		"bee",
 		"Health Check",
+		"Burn-in tests",
 		"> Export support bundle",
 		"Settings",
 		"Exit",
@@ -400,6 +403,11 @@ func TestConfirmCancelTarget(t *testing.T) {
 		t.Fatalf("storage sat cancel target=%q want %q", got, screenHealthCheck)
 	}
 
+	m.pendingAction = actionRunFanStress
+	if got := m.confirmCancelTarget(); got != screenBurnInTests {
+		t.Fatalf("fan stress cancel target=%q want %q", got, screenBurnInTests)
+	}
+
 	m.pendingAction = actionNone
 	if got := m.confirmCancelTarget(); got != screenMain {
 		t.Fatalf("default cancel target=%q want %q", got, screenMain)
@@ -439,6 +447,68 @@ func TestViewBusyStateUsesBusyTitle(t *testing.T) {
 	}
 }
 
+func TestBurnInTestsEscReturnsToMain(t *testing.T) {
+	t.Parallel()
+
+	m := newTestModel()
+	m.screen = screenBurnInTests
+	m.burnCursor = 3
+
+	next, _ := m.updateBurnInTests(tea.KeyMsg{Type: tea.KeyEsc})
+	got := next.(model)
+
+	if got.screen != screenMain {
+		t.Fatalf("screen=%q want %q", got.screen, screenMain)
+	}
+	if got.cursor != 1 {
+		t.Fatalf("cursor=%d want 1", got.cursor)
+	}
+}
+
+func TestBurnInTestsRunOpensConfirm(t *testing.T) {
+	t.Parallel()
+
+	m := newTestModel()
+	m.screen = screenBurnInTests
+	m.burnInitialized = true
+	m.burnMode = 2
+
+	next, _ := m.burnRunSelected()
+	got := next.(model)
+
+	if got.screen != screenConfirm {
+		t.Fatalf("screen=%q want %q", got.screen, screenConfirm)
+	}
+	if got.pendingAction != actionRunFanStress {
+		t.Fatalf("pendingAction=%q want %q", got.pendingAction, actionRunFanStress)
+	}
+	if got.cursor != 0 {
+		t.Fatalf("cursor=%d want 0", got.cursor)
+	}
+}
+
+func TestViewBurnInTestsRendersGPUStressEntry(t *testing.T) {
+	t.Parallel()
+
+	m := newTestModel()
+	m.screen = screenBurnInTests
+
+	view := m.View()
+
+	for _, want := range []string{
+		"BURN-IN TESTS",
+		"GPU PLATFORM STRESS TEST",
+		"Quick",
+		"Standard",
+		"Express",
+		"[ RUN SELECTED [R] ]",
+	} {
+		if !strings.Contains(view, want) {
+			t.Fatalf("view missing %q\nview:\n%s", want, view)
+		}
+	}
+}
+
 func TestViewOutputScreenRendersBodyAndBackHint(t *testing.T) {
 	t.Parallel()
 
@@ -528,7 +598,7 @@ func TestViewExportTargetsRendersDeviceMetadata(t *testing.T) {
 
 	for _, want := range []string{
 		"Export support bundle",
-		"Select removable filesystem",
+		"Select writable removable filesystem (read-only/boot media hidden)",
 		"> /dev/sdb1 [vfat 29G] label=BEEUSB mounted=/media/bee",
 	} {
 		if !strings.Contains(view, want) {
@@ -537,6 +607,32 @@ func TestViewExportTargetsRendersDeviceMetadata(t *testing.T) {
 	}
 }
 
+func TestExportTargetsMsgEmptyShowsHiddenBootMediaHint(t *testing.T) {
+	t.Parallel()
+
+	m := newTestModel()
+	m.busy = true
+	m.busyTitle = "Export support bundle"
+
+	next, _ := m.Update(exportTargetsMsg{})
+	got := next.(model)
+
+	if got.screen != screenOutput {
+		t.Fatalf("screen=%q want %q", got.screen, screenOutput)
+	}
+	if got.title != "Export support bundle" {
+		t.Fatalf("title=%q want %q", got.title, "Export support bundle")
+	}
+	for _, want := range []string{
+		"No writable removable filesystems found.",
+		"Read-only or boot media are hidden from this list.",
+	} {
+		if !strings.Contains(got.body, want) {
+			t.Fatalf("body missing %q\nbody:\n%s", want, got.body)
+		}
+	}
+}
+
 func TestViewStaticFormRendersFields(t *testing.T) {
 	t.Parallel()
 

audit/internal/tui/types.go

@@ -16,6 +16,7 @@ type screen string
 const (
 	screenMain             screen = "main"
 	screenHealthCheck      screen = "health_check"
+	screenBurnInTests      screen = "burn_in_tests"
 	screenSettings         screen = "settings"
 	screenNetwork          screen = "network"
 	screenInterfacePick    screen = "interface_pick"
@@ -84,6 +85,11 @@ type model struct {
 	hcCursor      int
 	hcInitialized bool
 
+	// Burn-in tests screen
+	burnMode        int
+	burnCursor      int
+	burnInitialized bool
+
 	// NVIDIA SAT setup
 	nvidiaGPUs      []platform.NvidiaGPU
 	nvidiaGPUSel    []bool
@@ -97,6 +103,9 @@ type model struct {
 	// GPU Platform Stress Test running
 	gpuStressCancel  func()
 	gpuStressAborted bool
+	gpuLiveRows      []platform.GPUMetricRow
+	gpuLiveIndices   []int
+	gpuLiveStart     time.Time
 
 	// SAT verbose progress (CPU / Memory / Storage / AMD GPU)
 	progressLines  []string
@@ -129,6 +138,7 @@ func newModel(application *app.App, runtimeMode runtimeenv.Mode) model {
 		screen:      screenMain,
 		mainMenu: []string{
 			"Health Check",
+			"Burn-in tests",
 			"Export support bundle",
 			"Settings",
 			"Exit",
@@ -198,7 +208,7 @@ func (m model) confirmBody() (string, string) {
 		modes := []string{"Quick (2×2min)", "Standard (2×5min)", "Express (2×10min)"}
 		return "GPU Platform Stress Test", "Two-phase GPU thermal cycling test.\n" +
 			"Monitors fans, temps, power — detects throttling.\n" +
-			"Mode: " + modes[m.hcMode] + "\n\nAll NVIDIA GPUs will be stressed."
+			"Mode: " + modes[m.burnMode] + "\n\nAll NVIDIA GPUs will be stressed."
 	default:
 		return "Confirm", "Proceed?"
 	}

audit/internal/tui/update.go

@@ -3,6 +3,7 @@ package tui
 import (
 	"fmt"
 	"strings"
+	"time"
 
 	tea "github.com/charmbracelet/bubbletea"
 )
@@ -100,6 +101,13 @@ func (m model) Update(msg tea.Msg) (tea.Model, tea.Cmd) {
 			m.screen = screenOutput
 			return m, m.refreshSnapshotCmd()
 		}
+		if len(msg.targets) == 0 {
+			m.title = "Export support bundle"
+			m.body = "No writable removable filesystems found.\n\nRead-only or boot media are hidden from this list."
+			m.prevScreen = screenMain
+			m.screen = screenOutput
+			return m, m.refreshSnapshotCmd()
+		}
 		m.targets = msg.targets
 		m.screen = screenExportTargets
 		m.cursor = 0
@@ -116,7 +124,7 @@ func (m model) Update(msg tea.Msg) (tea.Model, tea.Cmd) {
 			m.gpuStressCancel()
 			m.gpuStressCancel = nil
 		}
-		m.prevScreen = screenHealthCheck
+		m.prevScreen = screenBurnInTests
 		m.screen = screenOutput
 		m.title = msg.title
 		if msg.err != nil {
@@ -130,6 +138,22 @@ func (m model) Update(msg tea.Msg) (tea.Model, tea.Cmd) {
 			m.body = msg.body
 		}
 		return m, m.refreshSnapshotCmd()
+	case gpuLiveTickMsg:
+		if m.screen == screenGPUStressRunning {
+			if len(msg.rows) > 0 {
+				elapsed := time.Since(m.gpuLiveStart).Seconds()
+				for i := range msg.rows {
+					msg.rows[i].ElapsedSec = elapsed
+				}
+				m.gpuLiveRows = append(m.gpuLiveRows, msg.rows...)
+				n := max(1, len(msg.indices))
+				if len(m.gpuLiveRows) > 60*n {
+					m.gpuLiveRows = m.gpuLiveRows[len(m.gpuLiveRows)-60*n:]
+				}
+			}
+			return m, pollGPULive(msg.indices)
+		}
+		return m, nil
 	case nvidiaSATDoneMsg:
 		if m.nvidiaSATAborted {
 			return m, nil
@@ -162,6 +186,8 @@ func (m model) updateKey(msg tea.KeyMsg) (tea.Model, tea.Cmd) {
 		return m.updateMain(msg)
 	case screenHealthCheck:
 		return m.updateHealthCheck(msg)
+	case screenBurnInTests:
+		return m.updateBurnInTests(msg)
 	case screenSettings:
 		return m.updateMenu(msg, len(m.settingsMenu), m.handleSettingsMenu)
 	case screenNetwork:

audit/internal/tui/view.go

@@ -57,6 +57,8 @@ func (m model) View() string {
 			body = renderTwoColumnMain(m)
 		case screenHealthCheck:
 			body = renderHealthCheck(m)
+		case screenBurnInTests:
+			body = renderBurnInTests(m)
 		case screenSettings:
 			body = renderMenu("Settings", "Select action", m.settingsMenu, m.cursor)
 		case screenNetwork:
@@ -66,7 +68,12 @@ func (m model) View() string {
 		case screenServiceAction:
 			body = renderMenu("Service: "+m.selectedService, "Select action", m.serviceMenu, m.cursor)
 		case screenExportTargets:
-			body = renderMenu("Export support bundle", "Select removable filesystem", renderTargetItems(m.targets), m.cursor)
+			body = renderMenu(
+				"Export support bundle",
+				"Select writable removable filesystem (read-only/boot media hidden)",
+				renderTargetItems(m.targets),
+				m.cursor,
+			)
 		case screenInterfacePick:
 			body = renderMenu("Interfaces", "Select interface", renderInterfaceItems(m.interfaces), m.cursor)
 		case screenStaticForm:
@@ -79,7 +86,7 @@ func (m model) View() string {
 		case screenNvidiaSATRunning:
 			body = renderNvidiaSATRunning()
 		case screenGPUStressRunning:
-			body = renderGPUStressRunning()
+			body = renderGPUStressRunning(m)
 		case screenOutput:
 			body = fmt.Sprintf("%s\n\n%s\n\n[enter/esc] back  [ctrl+c] quit\n", m.title, strings.TrimSpace(m.body))
 		default:

bible-local/architecture/runtime-flows.md

@@ -9,6 +9,8 @@ DHCP is used only for LAN (operator SSH access). Internet is NOT available.
 
 ## Boot sequence (single ISO)
 
+The live system is expected to boot with `toram`, so `live-boot` copies the full read-only medium into RAM before mounting the root filesystem. After that point, runtime must not depend on the original USB/BMC virtual media staying readable.
+
 `systemd` boot order:
 
 ```
@@ -25,6 +27,7 @@ local-fs.target
 ```
 
 **Critical invariants:**
+- The live ISO boots with `boot=live toram`. Runtime binaries must continue working even if the original boot media disappears after early boot.
 - OpenSSH MUST start without network. `bee-sshsetup.service` runs before `ssh.service`.
 - `bee-network.service` uses `dhclient -nw` (background) — network bring-up is best effort and non-blocking.
 - `bee-nvidia.service` loads modules via `insmod` with absolute paths — NOT `modprobe`.
@@ -71,24 +74,39 @@ build-in-container.sh [--authorized-keys /path/to/keys]
        d. build kernel modules against Debian headers
        e. create `libnvidia-ml.so.1` / `libcuda.so.1` symlinks in cache
        f. cache in `dist/nvidia-<version>-<kver>/`
-  7. inject NVIDIA `.ko` → staged `/usr/local/lib/nvidia/`
+  7. `build-cublas.sh`:
-  8. inject `nvidia-smi` → staged `/usr/local/bin/nvidia-smi`
+       a. download `libcublas`, `libcublasLt`, `libcudart` runtime + dev packages from the NVIDIA CUDA Debian repo
-  9. inject `libnvidia-ml` + `libcuda` → staged `/usr/lib/`
+       b. verify packages against repo `Packages.gz`
-  10. write staged `/etc/bee-release` (versions + git commit)
+       c. extract headers for `bee-gpu-stress` build
-  11. patch staged `motd` with build metadata
+       d. cache userspace libs in `dist/cublas-<version>+cuda<series>/`
-  12. copy `iso/builder/` into a temporary live-build workdir under `dist/`
+  8. build `bee-gpu-stress` against extracted cuBLASLt/cudart headers
-  13. sync staged overlay into workdir `config/includes.chroot/`
+  9. inject NVIDIA `.ko` → staged `/usr/local/lib/nvidia/`
-  14. run `lb config && lb build` inside the privileged builder container
+  10. inject `nvidia-smi` → staged `/usr/local/bin/nvidia-smi`
+  11. inject `libnvidia-ml` + `libcuda` + `libcublas` + `libcublasLt` + `libcudart` → staged `/usr/lib/`
+  12. write staged `/etc/bee-release` (versions + git commit)
+  13. patch staged `motd` with build metadata
+  14. copy `iso/builder/` into a temporary live-build workdir under `dist/`
+  15. sync staged overlay into workdir `config/includes.chroot/`
+  16. run `lb config && lb build` inside the privileged builder container
 ```
 
+Build host notes:
+- `build-in-container.sh` targets `linux/amd64` builder containers by default, including Docker Desktop on macOS / Apple Silicon.
+- Override with `BEE_BUILDER_PLATFORM=<os/arch>` only if you intentionally need a different container platform.
+- If the local builder image under the same tag was previously built for the wrong architecture, the script rebuilds it automatically.
+
 **Critical invariants:**
 - `DEBIAN_KERNEL_ABI` in `iso/builder/VERSIONS` pins the exact kernel ABI used in BOTH places:
   1. `build-in-container.sh` / `build-nvidia-module.sh` — Debian kernel headers for module build
   2. `auto/config` — `linux-image-${DEBIAN_KERNEL_ABI}` in the ISO
 - NVIDIA modules go to staged `usr/local/lib/nvidia/` — NOT to `/lib/modules/<kver>/extra/`.
+- `bee-gpu-stress` must be built against cached CUDA userspace headers from `build-cublas.sh`, not against random host-installed CUDA headers.
+- The live ISO must ship `libcublas`, `libcublasLt`, and `libcudart` together with `libcuda` so tensor-core stress works without internet or package installs at boot.
 - The source overlay in `iso/overlay/` is treated as immutable source. Build-time files are injected only into the staged overlay.
 - The live-build workdir under `dist/` is disposable; source files under `iso/builder/` stay clean.
 - Container build requires `--privileged` because `live-build` uses mounts/chroots/loop devices during ISO assembly.
+- On macOS / Docker Desktop, the builder still must run as `linux/amd64` so the shipped ISO binaries remain `amd64`.
+- Operators must provision enough RAM to hold the full compressed live medium plus normal runtime overhead, because `toram` copies the entire read-only ISO payload into memory before the system reaches steady state.
 
 ## Post-boot smoke test
 
@@ -131,10 +149,15 @@ Current validation state:
 Every collector returns `nil, nil` on tool-not-found. Errors are logged, never fatal.
 
 Acceptance flows:
-- `bee sat nvidia` → diagnostic archive with `nvidia-smi -q` + `nvidia-bug-report` + lightweight `bee-gpu-stress`
+- `bee sat nvidia` → diagnostic archive with `nvidia-smi -q` + `nvidia-bug-report` + mixed-precision `bee-gpu-stress`
 - `bee sat memory` → `memtester` archive
 - `bee sat storage` → SMART/NVMe diagnostic archive and short self-test trigger where supported
 - SAT `summary.txt` now includes `overall_status` and per-job `*_status` values (`OK`, `FAILED`, `UNSUPPORTED`)
+- `bee-gpu-stress` should prefer cuBLASLt GEMM load over the old integer/PTX burn path:
+  - Ampere: `fp16` + `fp32`/TF32 tensor-core load
+  - Ada / Hopper: add `fp8`
+  - Blackwell+: add `fp4`
+  - PTX fallback is only for missing cuBLASLt/userspace or unsupported narrow datatypes
 - Runtime overrides:
   - `BEE_GPU_STRESS_SECONDS`
   - `BEE_GPU_STRESS_SIZE_MB`

bible-local/architecture/system-overview.md

@@ -21,7 +21,8 @@ Fills gaps where Redfish/logpile is blind:
 - Read-only hardware inventory: board, CPU, memory, storage, PCIe, PSU, GPU, NIC, RAID
 - Machine-readable health summary derived from collector verdicts
 - Operator-triggered acceptance tests for NVIDIA, memory, and storage
-- NVIDIA SAT includes both diagnostic collection and lightweight GPU stress via `bee-gpu-stress`
+- NVIDIA SAT includes both diagnostic collection and mixed-precision GPU stress via `bee-gpu-stress`
+- `bee-gpu-stress` should exercise tensor/inference paths (`fp16`, `fp32`/TF32, `fp8`, `fp4` when supported by the GPU/userspace stack) and fall back to Driver API PTX burn only if cuBLASLt is unavailable
 - Automatic boot audit with operator-facing local console and SSH access
 - NVIDIA proprietary driver loaded at boot for GPU enrichment via `nvidia-smi`
 - SSH access (OpenSSH) always available for inspection and debugging
@@ -69,6 +70,7 @@ Fills gaps where Redfish/logpile is blind:
 | SSH | OpenSSH server |
 | NVIDIA driver | Proprietary `.run` installer, built against Debian kernel headers |
 | NVIDIA modules | Loaded via `insmod` from `/usr/local/lib/nvidia/` |
+| GPU stress backend | `bee-gpu-stress` + cuBLASLt/cuBLAS/cudart mixed-precision GEMM, with Driver API PTX fallback |
 | Builder | Debian 12 host/VM or Debian 12 container image |
 
 ## Operator UX
@@ -78,6 +80,7 @@ Fills gaps where Redfish/logpile is blind:
 - The TUI itself executes privileged actions as `root` via `sudo -n`
 - SSH remains available independently of the local console path
 - VM-oriented builds also include `qemu-guest-agent` and serial console support for debugging
+- The ISO boots with `toram`, so loss of the original USB/BMC virtual media after boot should not break already-installed runtime binaries
 
 ## Runtime split
 
@@ -85,6 +88,7 @@ Fills gaps where Redfish/logpile is blind:
 - Live-ISO-only responsibilities stay in `iso/` integration code
 - Live ISO launches the Go CLI with `--runtime livecd`
 - Local/manual runs use `--runtime auto` or `--runtime local`
+- Live ISO targets must have enough RAM for the full compressed live medium plus runtime working set because the boot medium is copied into memory at startup
 
 ## Key paths
 

iso/README.md

@@ -0,0 +1,58 @@
+# ISO Build
+
+`bee` ISO is built inside a Debian 12 builder container via `iso/builder/build-in-container.sh`.
+
+## Requirements
+
+- Docker Desktop or another Docker-compatible container runtime
+- Privileged containers enabled
+- Enough free disk space for builder cache, Debian live-build artifacts, NVIDIA driver cache, and CUDA userspace packages
+
+## Build On macOS
+
+From the repository root:
+
+```sh
+sh iso/builder/build-in-container.sh
+```
+
+The script defaults to `linux/amd64` builder containers, so it works on:
+
+- Intel Mac
+- Apple Silicon (`M1` / `M2` / `M3` / `M4`) via Docker Desktop's Linux VM
+
+You do not need to pass `--platform` manually for normal ISO builds.
+
+## Useful Options
+
+Build with explicit SSH keys baked into the ISO:
+
+```sh
+sh iso/builder/build-in-container.sh --authorized-keys ~/.ssh/id_ed25519.pub
+```
+
+Rebuild the builder image:
+
+```sh
+sh iso/builder/build-in-container.sh --rebuild-image
+```
+
+Use a custom cache directory:
+
+```sh
+sh iso/builder/build-in-container.sh --cache-dir /path/to/cache
+```
+
+## Notes
+
+- The builder image is automatically rebuilt if the local tag exists for the wrong architecture.
+- The live ISO boots with Debian `live-boot` `toram`, so the read-only medium is copied into RAM during boot and the runtime no longer depends on the original USB/BMC virtual media staying present.
+- Target systems need enough RAM for the full compressed live medium plus normal runtime overhead, or boot may fail before reaching the TUI.
+- Override the container platform only if you know why:
+
+```sh
+BEE_BUILDER_PLATFORM=linux/amd64 sh iso/builder/build-in-container.sh
+```
+
+- The shipped ISO is still `amd64`.
+- Output ISO artifacts are written under `dist/`.

iso/builder/VERSIONS

@@ -4,5 +4,7 @@ NVIDIA_DRIVER_VERSION=590.48.01
 NCCL_VERSION=2.28.9-1
 NCCL_CUDA_VERSION=13.0
 NCCL_SHA256=2e6faafd2c19cffc7738d9283976a3200ea9db9895907f337f0c7e5a25563186
+CUBLAS_VERSION=13.0.2.14-1
+CUDA_USERSPACE_VERSION=13.0.96-1
 GO_VERSION=1.24.0
 AUDIT_VERSION=1.0.0

iso/builder/auto/config

@@ -32,6 +32,6 @@ lb config noauto \
     --memtest none \
     --iso-volume "EASY-BEE" \
     --iso-application "EASY-BEE" \
-    --bootappend-live "boot=live components console=tty0 console=ttyS0,115200n8 loglevel=3 username=bee user-fullname=Bee modprobe.blacklist=nouveau" \
+    --bootappend-live "boot=live toram components console=tty2 console=ttyS0,115200n8 loglevel=7 username=bee user-fullname=Bee modprobe.blacklist=nouveau" \
     --apt-recommends false \
     "${@}"

iso/builder/build-cublas.sh

@@ -0,0 +1,170 @@
+#!/bin/sh
+# build-cublas.sh — download cuBLASLt/cuBLAS/cudart runtime + headers for bee-gpu-stress.
+#
+# Downloads .deb packages from NVIDIA's CUDA apt repository (Debian 12, x86_64),
+# verifies them against Packages.gz, and extracts the small subset we need:
+#   - headers for compiling bee-gpu-stress against cuBLASLt
+#   - runtime libs for libcublas, libcublasLt, libcudart inside the ISO
+
+set -e
+
+CUBLAS_VERSION="$1"
+CUDA_USERSPACE_VERSION="$2"
+CUDA_SERIES="$3"
+DIST_DIR="$4"
+
+[ -n "$CUBLAS_VERSION" ]        || { echo "usage: $0 <cublas-version> <cuda-userspace-version> <cuda-series> <dist-dir>"; exit 1; }
+[ -n "$CUDA_USERSPACE_VERSION" ] || { echo "usage: $0 <cublas-version> <cuda-userspace-version> <cuda-series> <dist-dir>"; exit 1; }
+[ -n "$CUDA_SERIES" ]            || { echo "usage: $0 <cublas-version> <cuda-userspace-version> <cuda-series> <dist-dir>"; exit 1; }
+[ -n "$DIST_DIR" ]               || { echo "usage: $0 <cublas-version> <cuda-userspace-version> <cuda-series> <dist-dir>"; exit 1; }
+
+CUDA_SERIES_DASH=$(printf '%s' "$CUDA_SERIES" | tr '.' '-')
+REPO_BASE="https://developer.download.nvidia.com/compute/cuda/repos/debian12/x86_64"
+CACHE_DIR="${DIST_DIR}/cublas-${CUBLAS_VERSION}+cuda${CUDA_SERIES}"
+CACHE_ROOT="${BEE_CACHE_DIR:-${DIST_DIR}/cache}"
+DOWNLOAD_CACHE_DIR="${CACHE_ROOT}/cublas-downloads"
+PACKAGES_GZ="${DOWNLOAD_CACHE_DIR}/Packages.gz"
+
+echo "=== cuBLAS ${CUBLAS_VERSION} / cudart ${CUDA_USERSPACE_VERSION} / CUDA ${CUDA_SERIES} ==="
+
+if [ -f "${CACHE_DIR}/include/cublasLt.h" ] && [ -f "${CACHE_DIR}/include/cuda_runtime_api.h" ] \
+    && [ "$(find "${CACHE_DIR}/lib" \( -name 'libcublas.so*' -o -name 'libcublasLt.so*' -o -name 'libcudart.so*' \) 2>/dev/null | wc -l)" -gt 0 ]; then
+    echo "=== cuBLAS cached, skipping download ==="
+    echo "cache: $CACHE_DIR"
+    exit 0
+fi
+
+mkdir -p "${DOWNLOAD_CACHE_DIR}" "${CACHE_DIR}/include" "${CACHE_DIR}/lib"
+
+echo "=== downloading Packages.gz ==="
+wget -q -O "${PACKAGES_GZ}" "${REPO_BASE}/Packages.gz"
+
+lookup_pkg() {
+    pkg="$1"
+    ver="$2"
+    gzip -dc "${PACKAGES_GZ}" | awk -v pkg="$pkg" -v ver="$ver" '
+        /^Package: / { cur_pkg=$2 }
+        /^Version: / { cur_ver=$2 }
+        /^Filename: / { cur_file=$2 }
+        /^SHA256: / { cur_sha=$2 }
+        /^$/ {
+            if (cur_pkg == pkg && cur_ver == ver) {
+                print cur_file " " cur_sha
+                exit
+            }
+            cur_pkg=""; cur_ver=""; cur_file=""; cur_sha=""
+        }
+        END {
+            if (cur_pkg == pkg && cur_ver == ver) {
+                print cur_file " " cur_sha
+            }
+        }'
+}
+
+download_verified_pkg() {
+    pkg="$1"
+    ver="$2"
+
+    meta="$(lookup_pkg "$pkg" "$ver")"
+    [ -n "$meta" ] || { echo "ERROR: package metadata not found for ${pkg} ${ver}"; exit 1; }
+
+    repo_file="$(printf '%s\n' "$meta" | awk '{print $1}')"
+    repo_sha="$(printf '%s\n' "$meta" | awk '{print $2}')"
+    [ -n "$repo_file" ] || { echo "ERROR: package filename missing for ${pkg}"; exit 1; }
+    [ -n "$repo_sha" ]  || { echo "ERROR: package sha missing for ${pkg}"; exit 1; }
+
+    out="${DOWNLOAD_CACHE_DIR}/$(basename "$repo_file")"
+    if [ -f "$out" ]; then
+        actual_sha="$(sha256sum "$out" | awk '{print $1}')"
+        if [ "$actual_sha" = "$repo_sha" ]; then
+            echo "=== using cached $(basename "$repo_file") ==="
+            printf '%s\n' "$out"
+            return 0
+        fi
+        echo "=== removing stale $(basename "$repo_file") (sha256 mismatch) ==="
+        rm -f "$out"
+    fi
+
+    echo "=== downloading $(basename "$repo_file") ==="
+    wget --show-progress -O "$out" "${REPO_BASE}/$(basename "$repo_file")"
+
+    actual_sha="$(sha256sum "$out" | awk '{print $1}')"
+    if [ "$actual_sha" != "$repo_sha" ]; then
+        echo "ERROR: sha256 mismatch for $(basename "$repo_file")"
+        echo "  expected: $repo_sha"
+        echo "  actual:   $actual_sha"
+        rm -f "$out"
+        exit 1
+    fi
+    echo "sha256 OK: $(basename "$repo_file")"
+    printf '%s\n' "$out"
+}
+
+extract_deb() {
+    deb="$1"
+    dst="$2"
+    mkdir -p "$dst"
+    (
+        cd "$dst"
+        ar x "$deb"
+        data_tar=$(ls data.tar.* 2>/dev/null | head -1)
+        [ -n "$data_tar" ] || { echo "ERROR: data.tar.* not found in $deb"; exit 1; }
+        tar xf "$data_tar"
+    )
+}
+
+copy_headers() {
+    from="$1"
+    if [ -d "${from}/usr/include" ]; then
+        cp -a "${from}/usr/include/." "${CACHE_DIR}/include/"
+    fi
+}
+
+copy_libs() {
+    from="$1"
+    find "$from" \( -name 'libcublas.so*' -o -name 'libcublasLt.so*' -o -name 'libcudart.so*' \) \
+        \( -type f -o -type l \) -exec cp -a {} "${CACHE_DIR}/lib/" \;
+}
+
+make_links() {
+    base="$1"
+    versioned=$(find "${CACHE_DIR}/lib" -maxdepth 1 -name "${base}.so.[0-9]*" -type f | sort | head -1)
+    [ -n "$versioned" ] || return 0
+    soname=$(printf '%s\n' "$versioned" | sed -E "s#.*/(${base}\.so\.[0-9]+).*#\\1#")
+    target=$(basename "$versioned")
+    ln -sf "$target" "${CACHE_DIR}/lib/${soname}" 2>/dev/null || true
+    ln -sf "${soname}" "${CACHE_DIR}/lib/${base}.so" 2>/dev/null || true
+}
+
+TMP_DIR=$(mktemp -d)
+trap 'rm -rf "$TMP_DIR"' EXIT INT TERM
+
+CUBLAS_RT_DEB=$(download_verified_pkg "libcublas-${CUDA_SERIES_DASH}" "${CUBLAS_VERSION}")
+CUBLAS_DEV_DEB=$(download_verified_pkg "libcublas-dev-${CUDA_SERIES_DASH}" "${CUBLAS_VERSION}")
+CUDART_RT_DEB=$(download_verified_pkg "cuda-cudart-${CUDA_SERIES_DASH}" "${CUDA_USERSPACE_VERSION}")
+CUDART_DEV_DEB=$(download_verified_pkg "cuda-cudart-dev-${CUDA_SERIES_DASH}" "${CUDA_USERSPACE_VERSION}")
+
+extract_deb "$CUBLAS_RT_DEB" "${TMP_DIR}/cublas-rt"
+extract_deb "$CUBLAS_DEV_DEB" "${TMP_DIR}/cublas-dev"
+extract_deb "$CUDART_RT_DEB" "${TMP_DIR}/cudart-rt"
+extract_deb "$CUDART_DEV_DEB" "${TMP_DIR}/cudart-dev"
+
+copy_headers "${TMP_DIR}/cublas-dev"
+copy_headers "${TMP_DIR}/cudart-dev"
+copy_libs "${TMP_DIR}/cublas-rt"
+copy_libs "${TMP_DIR}/cudart-rt"
+
+make_links "libcublas"
+make_links "libcublasLt"
+make_links "libcudart"
+
+[ -f "${CACHE_DIR}/include/cublasLt.h" ]      || { echo "ERROR: cublasLt.h not extracted"; exit 1; }
+[ -f "${CACHE_DIR}/include/cuda_runtime_api.h" ] || { echo "ERROR: cuda_runtime_api.h not extracted"; exit 1; }
+[ "$(find "${CACHE_DIR}/lib" -maxdepth 1 -name 'libcublasLt.so*' | wc -l)" -gt 0 ] || { echo "ERROR: libcublasLt not extracted"; exit 1; }
+[ "$(find "${CACHE_DIR}/lib" -maxdepth 1 -name 'libcublas.so*' | wc -l)" -gt 0 ] || { echo "ERROR: libcublas not extracted"; exit 1; }
+[ "$(find "${CACHE_DIR}/lib" -maxdepth 1 -name 'libcudart.so*' | wc -l)" -gt 0 ] || { echo "ERROR: libcudart not extracted"; exit 1; }
+
+echo "=== cuBLAS extraction complete ==="
+echo "cache: $CACHE_DIR"
+echo "headers: $(find "${CACHE_DIR}/include" -type f | wc -l)"
+echo "libs:    $(find "${CACHE_DIR}/lib" -maxdepth 1 \( -name 'libcublas*.so*' -o -name 'libcudart.so*' \) | wc -l)"

iso/builder/build-in-container.sh

@@ -7,6 +7,7 @@ REPO_ROOT="$(cd "$(dirname "$0")/../.." && pwd)"
 BUILDER_DIR="${REPO_ROOT}/iso/builder"
 CONTAINER_TOOL="${CONTAINER_TOOL:-docker}"
 IMAGE_TAG="${BEE_BUILDER_IMAGE:-bee-iso-builder}"
+BUILDER_PLATFORM="${BEE_BUILDER_PLATFORM:-linux/amd64}"
 CACHE_DIR="${BEE_BUILDER_CACHE_DIR:-${REPO_ROOT}/dist/container-cache}"
 AUTH_KEYS=""
 REBUILD_IMAGE=0
@@ -40,6 +41,13 @@ if ! command -v "$CONTAINER_TOOL" >/dev/null 2>&1; then
     exit 1
 fi
 
+PLATFORM_OS="${BUILDER_PLATFORM%/*}"
+PLATFORM_ARCH="${BUILDER_PLATFORM#*/}"
+if [ -z "$PLATFORM_OS" ] || [ -z "$PLATFORM_ARCH" ] || [ "$PLATFORM_OS" = "$BUILDER_PLATFORM" ]; then
+    echo "invalid BEE_BUILDER_PLATFORM: ${BUILDER_PLATFORM} (expected os/arch, e.g. linux/amd64)" >&2
+    exit 1
+fi
+
 if [ -n "$AUTH_KEYS" ]; then
     [ -f "$AUTH_KEYS" ] || { echo "authorized_keys not found: $AUTH_KEYS" >&2; exit 1; }
     AUTH_KEYS_ABS="$(cd "$(dirname "$AUTH_KEYS")" && pwd)/$(basename "$AUTH_KEYS")"
@@ -56,17 +64,35 @@ mkdir -p \
 
 IMAGE_REF="${IMAGE_TAG}:debian${DEBIAN_VERSION}"
 
-if [ "$REBUILD_IMAGE" = "1" ] || ! "$CONTAINER_TOOL" image inspect "${IMAGE_REF}" >/dev/null 2>&1; then
+image_matches_platform() {
+    actual_platform="$("$CONTAINER_TOOL" image inspect --format '{{.Os}}/{{.Architecture}}' "${IMAGE_REF}" 2>/dev/null || true)"
+    [ "$actual_platform" = "${BUILDER_PLATFORM}" ]
+}
+
+NEED_BUILD_IMAGE=0
+if [ "$REBUILD_IMAGE" = "1" ]; then
+    NEED_BUILD_IMAGE=1
+elif ! "$CONTAINER_TOOL" image inspect "${IMAGE_REF}" >/dev/null 2>&1; then
+    NEED_BUILD_IMAGE=1
+elif ! image_matches_platform; then
+    actual_platform="$("$CONTAINER_TOOL" image inspect --format '{{.Os}}/{{.Architecture}}' "${IMAGE_REF}" 2>/dev/null || echo unknown)"
+    echo "=== rebuilding builder image ${IMAGE_REF}: platform mismatch (${actual_platform} != ${BUILDER_PLATFORM}) ==="
+    NEED_BUILD_IMAGE=1
+fi
+
+if [ "$NEED_BUILD_IMAGE" = "1" ]; then
     "$CONTAINER_TOOL" build \
+        --platform "${BUILDER_PLATFORM}" \
         --build-arg GO_VERSION="${GO_VERSION}" \
         -t "${IMAGE_REF}" \
         "${BUILDER_DIR}"
 else
-    echo "=== using existing builder image ${IMAGE_REF} ==="
+    echo "=== using existing builder image ${IMAGE_REF} (${BUILDER_PLATFORM}) ==="
 fi
 
 set -- \
     run --rm --privileged \
+    --platform "${BUILDER_PLATFORM}" \
     -v "${REPO_ROOT}:/work" \
     -v "${CACHE_DIR}:/cache" \
     -e BEE_CONTAINER_BUILD=1 \
@@ -80,6 +106,7 @@ set -- \
 
 if [ -n "$AUTH_KEYS" ]; then
     set -- run --rm --privileged \
+        --platform "${BUILDER_PLATFORM}" \
         -v "${REPO_ROOT}:/work" \
         -v "${CACHE_DIR}:/cache" \
         -v "${AUTH_KEYS_DIR}:/tmp/bee-authkeys:ro" \

iso/builder/build.sh

@@ -159,6 +159,16 @@ else
     echo "=== bee binary up to date, skipping build ==="
 fi
 
+echo ""
+echo "=== downloading cuBLAS/cuBLASLt/cudart ${NCCL_CUDA_VERSION} userspace ==="
+sh "${BUILDER_DIR}/build-cublas.sh" \
+    "${CUBLAS_VERSION}" \
+    "${CUDA_USERSPACE_VERSION}" \
+    "${NCCL_CUDA_VERSION}" \
+    "${DIST_DIR}"
+
+CUBLAS_CACHE="${DIST_DIR}/cublas-${CUBLAS_VERSION}+cuda${NCCL_CUDA_VERSION}"
+
 GPU_STRESS_NEED_BUILD=1
 if [ -f "$GPU_STRESS_BIN" ] && [ "${BUILDER_DIR}/bee-gpu-stress.c" -ot "$GPU_STRESS_BIN" ]; then
     GPU_STRESS_NEED_BUILD=0
@@ -167,6 +177,7 @@ fi
 if [ "$GPU_STRESS_NEED_BUILD" = "1" ]; then
     echo "=== building bee-gpu-stress ==="
     gcc -O2 -s -Wall -Wextra \
+        -I"${CUBLAS_CACHE}/include" \
         -o "$GPU_STRESS_BIN" \
         "${BUILDER_DIR}/bee-gpu-stress.c" \
         -ldl
@@ -283,6 +294,10 @@ NCCL_CACHE="${DIST_DIR}/nccl-${NCCL_VERSION}+cuda${NCCL_CUDA_VERSION}"
 cp "${NCCL_CACHE}/lib/"* "${OVERLAY_STAGE_DIR}/usr/lib/"
 echo "=== NCCL: $(ls "${NCCL_CACHE}/lib/" | wc -l) files injected into /usr/lib/ ==="
 
+# Inject cuBLAS/cuBLASLt/cudart runtime libs used by bee-gpu-stress tensor-core GEMM path
+cp "${CUBLAS_CACHE}/lib/"* "${OVERLAY_STAGE_DIR}/usr/lib/"
+echo "=== cuBLAS: $(ls "${CUBLAS_CACHE}/lib/" | wc -l) files injected into /usr/lib/ ==="
+
 # --- embed build metadata ---
 mkdir -p "${OVERLAY_STAGE_DIR}/etc"
 BUILD_DATE="$(date +%Y-%m-%d)"
@@ -297,6 +312,8 @@ DEBIAN_KERNEL_ABI=${DEBIAN_KERNEL_ABI}
 NVIDIA_DRIVER_VERSION=${NVIDIA_DRIVER_VERSION}
 NCCL_VERSION=${NCCL_VERSION}
 NCCL_CUDA_VERSION=${NCCL_CUDA_VERSION}
+CUBLAS_VERSION=${CUBLAS_VERSION}
+CUDA_USERSPACE_VERSION=${CUDA_USERSPACE_VERSION}
 EOF
 
 # Patch motd with build info

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

@@ -20,6 +20,7 @@ openssh-server
 
 # Filesystem support for USB export targets
 exfatprogs
+exfat-fuse
 ntfs-3g
 
 # Utilities

iso/overlay/etc/profile.d/bee.sh

@@ -17,4 +17,5 @@ if [ -z "${SSH_CONNECTION:-}" ] \
     && [ "$(tty 2>/dev/null)" = "/dev/tty1" ]; then
     echo "Bee live environment ready."
     echo "Run 'menu' to open the TUI."
+    echo "Kernel logs: Alt+F2  |  Extra shell: Alt+F3"
 fi

iso/overlay/etc/systemd/journald.conf.d/console-to-sol.conf

@@ -1,4 +1,4 @@
 [Journal]
 ForwardToConsole=yes
 TTYPath=/dev/ttyS0
-MaxLevelConsole=debug
+MaxLevelConsole=info