feat(nccl): add nccl-tests all_reduce_perf for GPU bandwidth testing

- Dockerfile: install cuda-nvcc-13-0 from NVIDIA repo for compilation
- build-nccl-tests.sh: downloads libnccl-dev for nccl.h, builds all_reduce_perf
- build.sh: runs nccl-tests build, injects binary into /usr/local/bin/
- platform: RunNCCLTests() auto-detects GPU count, runs all_reduce_perf
- TUI: NCCL bandwidth test entry in Burn-in Tests screen [N] hotkey

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

audit/internal/tui/forms.go

@@ -1,6 +1,7 @@
 package tui
 
 import (
+	"context"
 	"time"
 
 	"bee/audit/internal/platform"
@@ -140,6 +141,15 @@ func (m model) updateConfirm(msg tea.KeyMsg) (tea.Model, tea.Cmd) {
 			)
 		case actionRunFanStress:
 			return m.startGPUStressTest()
+		case actionRunNCCLTests:
+			m.busy = true
+			m.busyTitle = "NCCL bandwidth test"
+			ctx, cancel := context.WithCancel(context.Background())
+			m.ncclCancel = cancel
+			return m, func() tea.Msg {
+				result, err := m.app.RunNCCLTestsResult(ctx)
+				return resultMsg{title: result.Title, body: result.Body, err: err, back: screenBurnInTests}
+			}
 		}
 	case "ctrl+c":
 		return m, tea.Quit
@@ -153,7 +163,7 @@ func (m model) confirmCancelTarget() screen {
 		return screenExportTargets
 	case actionRunAll, actionRunMemorySAT, actionRunStorageSAT, actionRunCPUSAT, actionRunAMDGPUSAT:
 		return screenHealthCheck
-	case actionRunFanStress:
+	case actionRunFanStress, actionRunNCCLTests:
 		return screenBurnInTests
 	default:
 		return screenMain

audit/internal/tui/screen_burn_in.go

@@ -8,12 +8,13 @@ import (
 )
 
 const (
-	burnCurGPUStress = 0
-	burnCurModeQuick = 1
-	burnCurModeStd   = 2
-	burnCurModeExpr  = 3
-	burnCurRun       = 4
-	burnCurTotal     = 5
+	burnCurGPUStress  = 0
+	burnCurModeQuick  = 1
+	burnCurModeStd    = 2
+	burnCurModeExpr   = 3
+	burnCurRun        = 4
+	burnCurNCCLTests  = 5
+	burnCurTotal      = 6
 )
 
 func (m model) enterBurnInTests() (tea.Model, tea.Cmd) {
@@ -48,9 +49,13 @@ func (m model) updateBurnInTests(msg tea.KeyMsg) (tea.Model, tea.Cmd) {
 			return m.burnRunSelected()
 		case burnCurModeQuick, burnCurModeStd, burnCurModeExpr:
 			m.burnMode = m.burnCursor - burnCurModeQuick
+		case burnCurNCCLTests:
+			return m.burnRunNCCL()
 		}
 	case "f", "F", "r", "R":
 		return m.burnRunSelected()
+	case "n", "N":
+		return m.burnRunNCCL()
 	case "1":
 		m.burnMode = 0
 	case "2":
@@ -70,6 +75,13 @@ func (m model) burnRunSelected() (tea.Model, tea.Cmd) {
 	return m.hcRunFanStress()
 }
 
+func (m model) burnRunNCCL() (tea.Model, tea.Cmd) {
+	m.pendingAction = actionRunNCCLTests
+	m.screen = screenConfirm
+	m.cursor = 0
+	return m, nil
+}
+
 func renderBurnInTests(m model) string {
 	var b strings.Builder
 
@@ -110,8 +122,15 @@ func renderBurnInTests(m model) string {
 	}
 	fmt.Fprintf(&b, "%s[ RUN SELECTED [R] ]\n", pfx)
 
+	fmt.Fprintln(&b)
+	pfx = "  "
+	if m.burnCursor == burnCurNCCLTests {
+		pfx = "> "
+	}
+	fmt.Fprintf(&b, "%s[ NCCL BANDWIDTH TEST [N] ]         (all_reduce_perf, NVLink/PCIe bandwidth)\n", pfx)
+
 	fmt.Fprintln(&b)
 	fmt.Fprintln(&b, "─────────────────────────────────────────────────────────────────")
-	fmt.Fprint(&b, "[↑↓] move  [space/enter] select  [1/2/3] mode  [R/F] run  [Esc] back")
+	fmt.Fprint(&b, "[↑↓] move  [space/enter] select  [1/2/3] mode  [R/F] run  [N] nccl  [Esc] back")
 	return b.String()
 }

audit/internal/tui/tui_test.go

@@ -268,12 +268,9 @@ func TestHealthCheckGPUOpensNvidiaSATSetup(t *testing.T) {
 	m.hcInitialized = true
 	m.hcSel = [4]bool{true, true, true, true}
 
-	next, cmd := m.hcRunSingle(hcGPU)
+	next, _ := m.hcRunSingle(hcGPU)
 	got := next.(model)
 
-	if cmd == nil {
-		t.Fatal("expected non-nil cmd (GPU list loader)")
-	}
 	if got.screen != screenNvidiaSATSetup {
 		t.Fatalf("screen=%q want %q", got.screen, screenNvidiaSATSetup)
 	}

audit/internal/tui/types.go

@@ -44,6 +44,7 @@ const (
 	actionRunCPUSAT     actionKind = "run_cpu_sat"
 	actionRunAMDGPUSAT  actionKind = "run_amd_gpu_sat"
 	actionRunFanStress  actionKind = "run_fan_stress"
+	actionRunNCCLTests  actionKind = "run_nccl_tests"
 )
 
 type model struct {
@@ -98,6 +99,9 @@ type model struct {
 	nvidiaSATCancel  func()
 	nvidiaSATAborted bool
 
+	// NCCL tests running
+	ncclCancel func()
+
 	// GPU Platform Stress Test running
 	gpuStressCancel  func()
 	gpuStressAborted bool
@@ -202,6 +206,8 @@ func (m model) confirmBody() (string, string) {
 		return "CPU test", "Run stress-ng? Mode: " + modes[m.hcMode]
 	case actionRunAMDGPUSAT:
 		return "AMD GPU test", "Run AMD GPU diagnostic pack (rocm-smi)?"
+	case actionRunNCCLTests:
+		return "NCCL bandwidth test", "Run all_reduce_perf across all GPUs?\n\nMeasures collective bandwidth over NVLink/PCIe.\nRequires 2+ GPUs for meaningful results."
 	case actionRunFanStress:
 		modes := []string{"Quick (2×2min)", "Standard (2×5min)", "Express (2×10min)"}
 		return "GPU Platform Stress Test", "Two-phase GPU thermal cycling test.\n" +