Move nvtop to GPU-specific package lists; clean up git-bible

nvtop pulled nvidia-tesla-470-* via Recommends into the nogpu build.
Move it from bee.list.chroot into bee-nvidia and bee-amd lists so it
only appears in GPU variants.

Also remove the stray git-bible/ directory (was not gitignored) and
move grub-bitmap-error docs into bible-local/docs/.

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

.gitignore

@@ -1,6 +1,5 @@
 .env
 .DS_Store
 dist/
-iso/out/
 build-cache/
 audit/bee

audit/internal/collector/collector.go

@@ -40,6 +40,7 @@ func Run(_ runtimeenv.Mode) schema.HardwareIngestRequest {
 	snap.PCIeDevices = enrichPCIeWithAMD(snap.PCIeDevices)
 	snap.PCIeDevices = enrichPCIeWithPCISerials(snap.PCIeDevices)
 	snap.PCIeDevices = enrichPCIeWithNVIDIA(snap.PCIeDevices)
+	snap.PCIeDevices = enrichNVLinkBridgesWithGPUTopo(snap.PCIeDevices)
 	snap.PCIeDevices = enrichPCIeWithMellanox(snap.PCIeDevices)
 	snap.PCIeDevices = enrichPCIeWithNICTelemetry(snap.PCIeDevices)
 	snap.PCIeDevices = enrichPCIeWithRAIDTelemetry(snap.PCIeDevices)

audit/internal/collector/pcie.go

@@ -126,38 +126,39 @@ func parseLspciDevice(fields map[string]string) schema.HardwarePCIeDevice {
 	dev.Status = &status
 
 	// Slot is the BDF: "0000:00:02.0"
-	if bdf := fields["Slot"]; bdf != "" {
+	bdfStr := fields["Slot"]
-		dev.Slot = &bdf
+	if bdfStr != "" {
-		dev.BDF = &bdf
+		dev.Slot = &bdfStr
+		dev.BDF = &bdfStr
 		// parse vendor_id and device_id from sysfs
-		vendorID, deviceID := readPCIIDs(bdf)
+		vendorID, deviceID := readPCIIDs(bdfStr)
 		if vendorID != 0 {
 			dev.VendorID = &vendorID
 		}
 		if deviceID != 0 {
 			dev.DeviceID = &deviceID
 		}
-		if numaNode, ok := readPCINumaNode(bdf); ok {
+		if numaNode, ok := readPCINumaNode(bdfStr); ok {
 			dev.NUMANode = &numaNode
 		} else if numaNode, ok := parsePCINumaNode(fields["NUMANode"]); ok {
 			dev.NUMANode = &numaNode
 		}
-		if group, ok := readPCIIOMMUGroup(bdf); ok {
+		if group, ok := readPCIIOMMUGroup(bdfStr); ok {
 			dev.IOMMUGroup = &group
 		}
-		if width, ok := readPCIIntAttribute(bdf, "current_link_width"); ok {
+		if width, ok := readPCIIntAttribute(bdfStr, "current_link_width"); ok {
 			dev.LinkWidth = &width
 		}
-		if width, ok := readPCIIntAttribute(bdf, "max_link_width"); ok {
+		if width, ok := readPCIIntAttribute(bdfStr, "max_link_width"); ok {
 			dev.MaxLinkWidth = &width
 		}
-		if speed, ok := readPCIStringAttribute(bdf, "current_link_speed"); ok {
+		if speed, ok := readPCIStringAttribute(bdfStr, "current_link_speed"); ok {
 			linkSpeed := normalizePCILinkSpeed(speed)
 			if linkSpeed != "" {
 				dev.LinkSpeed = &linkSpeed
 			}
 		}
-		if speed, ok := readPCIStringAttribute(bdf, "max_link_speed"); ok {
+		if speed, ok := readPCIStringAttribute(bdfStr, "max_link_speed"); ok {
 			linkSpeed := normalizePCILinkSpeed(speed)
 			if linkSpeed != "" {
 				dev.MaxLinkSpeed = &linkSpeed
@@ -178,7 +179,15 @@ func parseLspciDevice(fields map[string]string) schema.HardwarePCIeDevice {
 
 	// SVendor/SDevice available but not in schema — skip
 
-	// Warn if PCIe link is running below its maximum negotiated speed.
+	// Detect NVLink bridge mezzanine cards (CPU→HGX internal link).
+	// These are Mellanox x2 devices with no host net interfaces and a DeviceName
+	// containing "NVLINK". The targeted lspci call is only executed for the small
+	// number of narrow-link Mellanox cards that pass the cheap pre-filter.
+	if bdfStr != "" && isNVLinkBridgeCandidate(bdfStr, dev) && confirmNVLinkBridgeDeviceName(bdfStr) {
+		markNVLinkBridge(&dev)
+	}
+
+	// Warn (or Critical for NVLink bridges) if PCIe link is running below max.
 	applyPCIeLinkSpeedWarning(&dev)
 
 	return dev
@@ -265,17 +274,27 @@ func readPCIStringAttribute(bdf, attribute string) (string, bool) {
 	return value, true
 }
 
-// applyPCIeLinkSpeedWarning sets the device status to Warning if the current PCIe link
+// applyPCIeLinkSpeedWarning sets device status when the current PCIe link speed is
-// speed is below the maximum negotiated speed supported by both ends.
+// below the device maximum. Regular PCIe slots get Warning; NVLink bridge cards
+// get Critical because they are fixed internal connectors that must always train
+// to max speed — any downgrade signals a hardware fault.
 func applyPCIeLinkSpeedWarning(dev *schema.HardwarePCIeDevice) {
 	if dev.LinkSpeed == nil || dev.MaxLinkSpeed == nil {
 		return
 	}
-	if pcieLinkSpeedRank(*dev.LinkSpeed) < pcieLinkSpeedRank(*dev.MaxLinkSpeed) {
+	if pcieLinkSpeedRank(*dev.LinkSpeed) >= pcieLinkSpeedRank(*dev.MaxLinkSpeed) {
+		return
+	}
+	desc := fmt.Sprintf("PCIe link speed degraded: running at %s, capable of %s", *dev.LinkSpeed, *dev.MaxLinkSpeed)
+	dev.ErrorDescription = &desc
+
+	isNVLinkBridge := dev.DeviceClass != nil && *dev.DeviceClass == "NVLinkBridge"
+	if isNVLinkBridge {
+		crit := statusCritical
+		dev.Status = &crit
+	} else {
 		warn := statusWarning
 		dev.Status = &warn
-		desc := fmt.Sprintf("PCIe link speed degraded: running at %s, capable of %s", *dev.LinkSpeed, *dev.MaxLinkSpeed)
-		dev.ErrorDescription = &desc
 	}
 }
 

audit/internal/collector/pcie_nvlink_bridge.go

@@ -0,0 +1,206 @@
+package collector
+
+import (
+	"bee/audit/internal/schema"
+	"log/slog"
+	"os/exec"
+	"regexp"
+	"strconv"
+	"strings"
+)
+
+var nv5re = regexp.MustCompile(`(?i)^NV(\d+)$`)
+
+// isNVLinkBridgeCandidate returns true for Mellanox PCIe devices that look like
+// NVLink bridge mezzanine cards: narrow link (x2), no host net interfaces.
+// These are the CPU-side PCIe control plane of the NVSwitch fabric on HGX/DGX systems.
+func isNVLinkBridgeCandidate(bdf string, dev schema.HardwarePCIeDevice) bool {
+	if !isMellanoxDevice(dev) {
+		return false
+	}
+	if dev.LinkWidth == nil || *dev.LinkWidth > 2 {
+		return false
+	}
+	if len(netIfacesByBDF(bdf)) > 0 {
+		return false
+	}
+	return true
+}
+
+// confirmNVLinkBridgeDeviceName checks if the lspci DeviceName for bdf contains
+// "NVLINK". This is a targeted single-device call, only executed for candidates
+// already pre-filtered by isNVLinkBridgeCandidate.
+func confirmNVLinkBridgeDeviceName(bdf string) bool {
+	out, err := exec.Command("lspci", "-s", bdf, "-v").Output()
+	if err != nil {
+		return false
+	}
+	for _, line := range strings.Split(string(out), "\n") {
+		if strings.Contains(strings.ToUpper(strings.TrimSpace(line)), "NVLINK") {
+			return true
+		}
+	}
+	return false
+}
+
+// markNVLinkBridge overwrites device_class and adds telemetry flags on a detected
+// NVLink bridge card. Must be called before applyPCIeLinkSpeedWarning so that the
+// correct severity (Critical) is applied.
+func markNVLinkBridge(dev *schema.HardwarePCIeDevice) {
+	class := "NVLinkBridge"
+	dev.DeviceClass = &class
+	if dev.Telemetry == nil {
+		dev.Telemetry = map[string]any{}
+	}
+	dev.Telemetry["nvlink_bridge"] = true
+}
+
+// enrichNVLinkBridgesWithGPUTopo cross-references NVLink bridge PCIe status with
+// the GPU-side NVLink topology reported by nvidia-smi. For each bridge device it
+// adds nvlink_topo_all_active and nvlink_topo_min_links to the telemetry, and
+// upgrades a degraded-link Warning to Critical when the fabric is also affected.
+func enrichNVLinkBridgesWithGPUTopo(devs []schema.HardwarePCIeDevice) []schema.HardwarePCIeDevice {
+	hasBridge := false
+	for _, d := range devs {
+		if d.DeviceClass != nil && *d.DeviceClass == "NVLinkBridge" {
+			hasBridge = true
+			break
+		}
+	}
+	if !hasBridge {
+		return devs
+	}
+
+	topo, err := queryNVIDIANVLinkTopo()
+	if err != nil {
+		slog.Info("nvlink-bridge: nvidia-smi topo unavailable, skipping cross-reference", "err", err)
+		return devs
+	}
+
+	for i := range devs {
+		if devs[i].DeviceClass == nil || *devs[i].DeviceClass != "NVLinkBridge" {
+			continue
+		}
+		if devs[i].Telemetry == nil {
+			devs[i].Telemetry = map[string]any{}
+		}
+		devs[i].Telemetry["nvlink_topo_all_active"] = topo.AllActive
+		devs[i].Telemetry["nvlink_topo_min_links"] = topo.MinNVLinks
+		devs[i].Telemetry["nvlink_topo_gpu_count"] = topo.GPUCount
+
+		// If the bridge PCIe is already degraded AND the fabric is also degraded
+		// (missing NVLink connections), escalate to Critical.
+		if devs[i].Status != nil && *devs[i].Status == statusCritical && !topo.AllActive {
+			devs[i].Telemetry["nvlink_fabric_affected"] = true
+		}
+	}
+
+	slog.Info("nvlink-bridge: topo cross-reference applied",
+		"gpu_count", topo.GPUCount,
+		"all_active", topo.AllActive,
+		"min_links", topo.MinNVLinks,
+	)
+	return devs
+}
+
+// nvlinkTopoResult summarises the GPU NVLink connectivity matrix.
+type nvlinkTopoResult struct {
+	GPUCount   int
+	AllActive  bool // true if every GPU pair has at least one NVLink bond
+	MinNVLinks int  // minimum NVLink bonds seen across any GPU pair (0 = some pair disconnected)
+}
+
+// queryNVIDIANVLinkTopo runs nvidia-smi topo -m and parses the NVLink matrix.
+func queryNVIDIANVLinkTopo() (nvlinkTopoResult, error) {
+	out, err := exec.Command("nvidia-smi", "topo", "-m").Output()
+	if err != nil {
+		return nvlinkTopoResult{}, err
+	}
+	return parseNVIDIATopologyMatrix(string(out)), nil
+}
+
+// parseNVIDIATopologyMatrix extracts the minimum NVLink bond count from the
+// nvidia-smi topo -m matrix.
+//
+// Format (abbreviated):
+//
+//	         GPU0  GPU1 ... NIC0 NIC1
+//	GPU0      X   NV18 ... NODE NODE
+//	GPU1     NV18   X  ... NODE NODE
+//	NIC0     NODE  NODE...   X   PIX
+//
+// The header row starts with "GPU0"; its columns may include non-GPU entries
+// (NIC, CPU) which are ignored. Only GPU×GPU cells containing NV# values are
+// counted. X is self; non-NV tokens (NODE, SYS, PHB, PIX) are skipped.
+func parseNVIDIATopologyMatrix(raw string) nvlinkTopoResult {
+	lines := strings.Split(raw, "\n")
+
+	// Locate the header line and record which column indices are GPU columns.
+	headerIdx := -1
+	var gpuColIndices []int // 0-based indices within fields (excluding the row label)
+	var gpuCount int
+	for i, line := range lines {
+		trimmed := strings.TrimSpace(line)
+		if strings.HasPrefix(trimmed, "GPU0") {
+			parts := strings.Fields(trimmed)
+			for j, col := range parts {
+				if strings.HasPrefix(col, "GPU") {
+					gpuColIndices = append(gpuColIndices, j)
+				}
+			}
+			gpuCount = len(gpuColIndices)
+			if gpuCount >= 2 {
+				headerIdx = i
+			}
+			break
+		}
+	}
+	if headerIdx < 0 || gpuCount == 0 {
+		return nvlinkTopoResult{}
+	}
+
+	minLinks := -1 // -1 = no NV pair seen yet
+	allActive := true
+
+	for _, line := range lines[headerIdx+1:] {
+		trimmed := strings.TrimSpace(line)
+		if !strings.HasPrefix(trimmed, "GPU") {
+			continue
+		}
+		cells := strings.Fields(trimmed)
+		// cells[0] is the row label (e.g. "GPU0"); cells[1..] are column values.
+		// gpuColIndices are 0-based within the header fields, so they map to
+		// cells[idx+1] in the data rows (shift by 1 for the row label).
+		for _, colIdx := range gpuColIndices {
+			dataIdx := colIdx + 1
+			if dataIdx >= len(cells) {
+				continue
+			}
+			cell := cells[dataIdx]
+			m := nv5re.FindStringSubmatch(cell)
+			if len(m) != 2 {
+				continue
+			}
+			n, err := strconv.Atoi(m[1])
+			if err != nil {
+				continue
+			}
+			if n == 0 {
+				allActive = false
+			}
+			if minLinks < 0 || n < minLinks {
+				minLinks = n
+			}
+		}
+	}
+
+	if minLinks < 0 {
+		minLinks = 0
+	}
+
+	return nvlinkTopoResult{
+		GPUCount:   gpuCount,
+		AllActive:  allActive && minLinks > 0,
+		MinNVLinks: minLinks,
+	}
+}

audit/internal/collector/pcie_nvlink_bridge_test.go

@@ -0,0 +1,124 @@
+package collector
+
+import (
+	"bee/audit/internal/schema"
+	"testing"
+)
+
+func TestParseNVIDIATopologyMatrix(t *testing.T) {
+	t.Parallel()
+
+	// Real-world B200 HGX output: 8 GPUs, all pairs connected via NV18.
+	input := `	GPU0	GPU1	GPU2	GPU3	GPU4	GPU5	GPU6	GPU7	NIC0	NIC1
+GPU0	 X 	NV18	NV18	NV18	NV18	NV18	NV18	NV18	NODE	NODE
+GPU1	NV18	 X 	NV18	NV18	NV18	NV18	NV18	NV18	NODE	NODE
+GPU2	NV18	NV18	 X 	NV18	NV18	NV18	NV18	NV18	NODE	NODE
+GPU3	NV18	NV18	NV18	 X 	NV18	NV18	NV18	NV18	NODE	NODE
+GPU4	NV18	NV18	NV18	NV18	 X 	NV18	NV18	NV18	SYS	SYS
+GPU5	NV18	NV18	NV18	NV18	NV18	 X 	NV18	NV18	SYS	SYS
+GPU6	NV18	NV18	NV18	NV18	NV18	NV18	 X 	NV18	SYS	SYS
+GPU7	NV18	NV18	NV18	NV18	NV18	NV18	NV18	 X 	SYS	SYS
+NIC0	NODE	NODE	NODE	NODE	SYS	SYS	SYS	SYS	 X 	PIX
+`
+	got := parseNVIDIATopologyMatrix(input)
+
+	if got.GPUCount != 8 {
+		t.Fatalf("GPUCount=%d want 8", got.GPUCount)
+	}
+	if !got.AllActive {
+		t.Fatalf("AllActive=false want true")
+	}
+	if got.MinNVLinks != 18 {
+		t.Fatalf("MinNVLinks=%d want 18", got.MinNVLinks)
+	}
+}
+
+func TestParseNVIDIATopologyMatrixPartialDegradation(t *testing.T) {
+	t.Parallel()
+
+	// GPU1-GPU3 pair shows NV12 (reduced) instead of NV18.
+	input := `	GPU0	GPU1	GPU2	GPU3
+GPU0	 X 	NV18	NV18	NV18
+GPU1	NV18	 X 	NV18	NV12
+GPU2	NV18	NV18	 X 	NV18
+GPU3	NV18	NV12	NV18	 X
+`
+	got := parseNVIDIATopologyMatrix(input)
+
+	if got.MinNVLinks != 12 {
+		t.Fatalf("MinNVLinks=%d want 12", got.MinNVLinks)
+	}
+	if !got.AllActive {
+		t.Fatalf("AllActive=false want true (12 links is still active)")
+	}
+}
+
+func TestParseNVIDIATopologyMatrixDisconnected(t *testing.T) {
+	t.Parallel()
+
+	// GPU0-GPU1 pair fully disconnected (NV0).
+	input := `	GPU0	GPU1
+GPU0	 X 	NV0
+GPU1	NV0	 X
+`
+	got := parseNVIDIATopologyMatrix(input)
+
+	if got.AllActive {
+		t.Fatalf("AllActive=true want false (NV0 means no links)")
+	}
+	if got.MinNVLinks != 0 {
+		t.Fatalf("MinNVLinks=%d want 0", got.MinNVLinks)
+	}
+}
+
+func TestParseNVIDIATopologyMatrixEmpty(t *testing.T) {
+	t.Parallel()
+
+	got := parseNVIDIATopologyMatrix("no gpus here")
+	if got.GPUCount != 0 {
+		t.Fatalf("GPUCount=%d want 0", got.GPUCount)
+	}
+}
+
+func TestApplyPCIeLinkSpeedWarningNVLinkBridgeEscalates(t *testing.T) {
+	t.Parallel()
+
+	bridgeClass := "NVLinkBridge"
+	linkSpeed := "Gen3"
+	maxLinkSpeed := "Gen4"
+	dev := schema.HardwarePCIeDevice{}
+	dev.DeviceClass = &bridgeClass
+	dev.LinkSpeed = &linkSpeed
+	dev.MaxLinkSpeed = &maxLinkSpeed
+	s := statusOK
+	dev.Status = &s
+
+	applyPCIeLinkSpeedWarning(&dev)
+
+	if dev.Status == nil || *dev.Status != statusCritical {
+		t.Fatalf("status=%v want Critical for NVLink bridge degradation", dev.Status)
+	}
+	if dev.ErrorDescription == nil {
+		t.Fatal("ErrorDescription nil, want degradation message")
+	}
+}
+
+func TestApplyPCIeLinkSpeedWarningRegularCardIsWarning(t *testing.T) {
+	t.Parallel()
+
+	regularClass := "NetworkController"
+	linkSpeed := "Gen3"
+	maxLinkSpeed := "Gen4"
+	dev := schema.HardwarePCIeDevice{}
+	dev.DeviceClass = &regularClass
+	dev.LinkSpeed = &linkSpeed
+	dev.MaxLinkSpeed = &maxLinkSpeed
+	s := statusOK
+	dev.Status = &s
+
+	applyPCIeLinkSpeedWarning(&dev)
+
+	if dev.Status == nil || *dev.Status != statusWarning {
+		t.Fatalf("status=%v want Warning for regular card degradation", dev.Status)
+	}
+}

audit/internal/collector/storage.go

@@ -36,6 +36,13 @@ func bestEffortRescanHotplugStorage() {
 		slog.Info("storage: scsi host scan skipped", "pattern", scsiHostScanGlob, "err", err)
 	} else {
 		for _, path := range hostPaths {
+			// SAS HBAs (e.g. smartpqi) block indefinitely in sas_user_scan when
+			// written to — SAS topology is discovered by the driver itself.
+			host := filepath.Base(filepath.Dir(path))
+			if _, err := os.Stat("/sys/class/sas_host/" + host); err == nil {
+				slog.Info("storage: scsi host scan skipped (SAS host)", "path", path)
+				continue
+			}
 			if err := hotplugWriteFile(path, []byte("- - -\n"), 0644); err != nil {
 				slog.Info("storage: scsi host scan write failed", "path", path, "err", err)
 				continue
@@ -66,17 +73,41 @@ func collectStorage() []schema.HardwareStorage {
 	return result
 }
 
+// jsonInt64 accepts both a bare JSON number and a JSON-quoted number string.
+// lsblk -J emits LOG-SEC / PHY-SEC as integers on util-linux ≥ 2.37 (Debian 12)
+// but older versions emit them as strings. This type handles both.
+type jsonInt64 int64
+
+func (j *jsonInt64) UnmarshalJSON(data []byte) error {
+	// bare number: 512
+	var n int64
+	if err := json.Unmarshal(data, &n); err == nil {
+		*j = jsonInt64(n)
+		return nil
+	}
+	// quoted string: "512"
+	var s string
+	if err := json.Unmarshal(data, &s); err == nil {
+		n, err := strconv.ParseInt(strings.TrimSpace(s), 10, 64)
+		if err == nil {
+			*j = jsonInt64(n)
+		}
+		return nil
+	}
+	return nil // null or unexpected type — leave zero
+}
+
 // lsblkDevice is a minimal lsblk JSON record.
 type lsblkDevice struct {
-	Name   string `json:"name"`
+	Name   string   `json:"name"`
-	Type   string `json:"type"`
+	Type   string   `json:"type"`
-	Size   string `json:"size"`
+	Size   string   `json:"size"`
-	Serial string `json:"serial"`
+	Serial string   `json:"serial"`
-	Model  string `json:"model"`
+	Model  string   `json:"model"`
-	Tran   string `json:"tran"`
+	Tran   string   `json:"tran"`
-	Hctl   string `json:"hctl"`
+	Hctl   string   `json:"hctl"`
-	LogSec string `json:"log-sec"`
+	LogSec jsonInt64 `json:"log-sec"`
-	PhySec string `json:"phy-sec"`
+	PhySec jsonInt64 `json:"phy-sec"`
 }
 
 type lsblkRoot struct {
@@ -620,8 +651,8 @@ func applyStorageBlockGeometry(s *schema.HardwareStorage, dev lsblkDevice) {
 	if s == nil {
 		return
 	}
-	logical := parseStorageBytes(dev.LogSec)
+	logical := int64(dev.LogSec)
-	physical := parseStorageBytes(dev.PhySec)
+	physical := int64(dev.PhySec)
 	if logical <= 0 && physical <= 0 {
 		return
 	}

audit/internal/collector/storage_discovery_test.go

@@ -1,6 +1,7 @@
 package collector
 
 import (
+	"encoding/json"
 	"os"
 	"os/exec"
 	"path/filepath"
@@ -38,6 +39,54 @@ func TestParseStorageBytes(t *testing.T) {
 	}
 }
 
+func TestJsonInt64UnmarshalBothFormats(t *testing.T) {
+	t.Parallel()
+
+	// util-linux ≥ 2.37 emits LOG-SEC / PHY-SEC as bare JSON numbers.
+	// Older versions emit quoted strings. Both must parse without error
+	// so that the entire lsblkDevices() call does not return nil on Debian 12.
+	cases := []struct {
+		json string
+		want int64
+	}{
+		{`512`, 512},
+		{`4096`, 4096},
+		{`"512"`, 512},
+		{`"4096"`, 4096},
+		{`null`, 0},
+	}
+	for _, tc := range cases {
+		var v jsonInt64
+		if err := v.UnmarshalJSON([]byte(tc.json)); err != nil {
+			t.Fatalf("UnmarshalJSON(%s): unexpected error %v", tc.json, err)
+		}
+		if int64(v) != tc.want {
+			t.Fatalf("UnmarshalJSON(%s)=%d want %d", tc.json, int64(v), tc.want)
+		}
+	}
+
+	// Simulate the exact JSON shape that triggered the bug on Debian 12.
+	input := []byte(`{
+		"blockdevices": [
+			{"name":"sda","type":"disk","size":"3.6T","serial":"S1234","model":"SEAGATE","tran":"sata","hctl":"0:0:0:0","log-sec":512,"phy-sec":4096},
+			{"name":"sdb","type":"disk","size":"3.6T","serial":"S5678","model":"SEAGATE","tran":"sata","hctl":"0:0:1:0","log-sec":512,"phy-sec":4096}
+		]
+	}`)
+	var root lsblkRoot
+	if err := json.Unmarshal(input, &root); err != nil {
+		t.Fatalf("lsblkRoot unmarshal with integer log-sec/phy-sec: %v", err)
+	}
+	if len(root.Blockdevices) != 2 {
+		t.Fatalf("got %d blockdevices want 2", len(root.Blockdevices))
+	}
+	if int64(root.Blockdevices[0].LogSec) != 512 {
+		t.Fatalf("LogSec=%d want 512", root.Blockdevices[0].LogSec)
+	}
+	if int64(root.Blockdevices[0].PhySec) != 4096 {
+		t.Fatalf("PhySec=%d want 4096", root.Blockdevices[0].PhySec)
+	}
+}
+
 func TestBestEffortRescanHotplugStorage(t *testing.T) {
 	t.Parallel()
 

audit/internal/platform/error_patterns.go

@@ -38,6 +38,15 @@ var HardwareErrorPatterns = []ErrorPattern{
 		Category: "gpu",
 		Severity: "warning",
 	},
+	// PCIe AER correctable from the NVIDIA driver — "bus correctable error" in SEL.
+	// Severity is warning (not critical): correctable errors are hardware-recovered.
+	{
+		Name:     "nvidia-aer-correctable",
+		Re:       mustPat(`(?i)nvidia\s+([\da-f]{4}:[\da-f]{2}:[\da-f]{2}\.\d).*AER.*[Cc]orrect`),
+		Category: "gpu",
+		Severity: "warning",
+		BDFGroup: 1,
+	},
 	{
 		Name:     "nvidia-aer",
 		Re:       mustPat(`(?i)nvidia\s+([\da-f]{4}:[\da-f]{2}:[\da-f]{2}\.\d).*AER`),
@@ -54,6 +63,15 @@ var HardwareErrorPatterns = []ErrorPattern{
 	},
 
 	// ── PCIe AER (generic) ──────────────────────────────────────────────────────
+	// PCIe AER correctable from the root port — captures the reported device BDF
+	// (second BDF in "pcieport X: AER: Correctable error received: Y").
+	{
+		Name:     "pcie-aer-correctable",
+		Re:       mustPat(`(?i)pcieport.*AER:.*[Cc]orrect.*:\s*([\da-f]{4}:[\da-f]{2}:[\da-f]{2}\.\d)`),
+		Category: "pcie",
+		Severity: "warning",
+		BDFGroup: 1,
+	},
 	{
 		Name:     "pcie-aer",
 		Re:       mustPat(`(?i)pcieport\s+([\da-f]{4}:[\da-f]{2}:[\da-f]{2}\.\d).*AER`),

audit/internal/platform/install.go

@@ -5,6 +5,7 @@ import (
 	"fmt"
 	"os"
 	"os/exec"
+	"path/filepath"
 	"strconv"
 	"strings"
 )
@@ -18,7 +19,7 @@ type InstallDisk struct {
 	MountedParts []string // partition mount points currently active
 }
 
-const squashfsPath = "/run/live/medium/live/filesystem.squashfs"
+const squashfsGlob = "/run/live/medium/live/*.squashfs"
 
 // ListInstallDisks returns block devices suitable for installation.
 // Excludes the current live boot medium but includes USB drives.
@@ -176,11 +177,22 @@ func inferLiveBootKind(fsType, source, deviceType, transport string) string {
 // squashfs size × 1.5 to allow for extracted filesystem and bootloader.
 // Returns 0 if the squashfs is not available (non-live environment).
 func MinInstallBytes() int64 {
-	fi, err := os.Stat(squashfsPath)
+	files, err := filepath.Glob(squashfsGlob)
-	if err != nil {
+	if err != nil || len(files) == 0 {
 		return 0
 	}
-	return fi.Size() * 3 / 2
+	var total int64
+	for _, path := range files {
+		fi, statErr := os.Stat(path)
+		if statErr != nil {
+			continue
+		}
+		total += fi.Size()
+	}
+	if total == 0 {
+		return 0
+	}
+	return total * 3 / 2
 }
 
 // toramActive returns true when the live system was booted with toram.
@@ -222,12 +234,10 @@ func DiskWarnings(d InstallDisk) []string {
 			humanBytes(min), humanBytes(d.SizeBytes)))
 	}
 	if toramActive() {
-		sqFi, err := os.Stat(squashfsPath)
+		free := freeMemBytes()
-		if err == nil {
+		min := MinInstallBytes()
-			free := freeMemBytes()
+		if free > 0 && min > 0 && free < (min*4/3) {
-			if free > 0 && free < sqFi.Size()*2 {
+			w = append(w, "toram mode — low RAM, extraction may be slow or fail")
-				w = append(w, "toram mode — low RAM, extraction may be slow or fail")
-			}
 		}
 	}
 	return w

audit/internal/platform/runtime.go

@@ -55,7 +55,6 @@ func (s *System) CollectRuntimeHealth(exportDir string) (schema.RuntimeHealth, e
 	if err == nil {
 		health.Interfaces = make([]schema.RuntimeInterface, 0, len(interfaces))
 		hasIPv4 := false
-		missingIPv4 := false
 		for _, iface := range interfaces {
 			outcome := "no_offer"
 			if len(iface.IPv4) > 0 {
@@ -63,8 +62,6 @@ func (s *System) CollectRuntimeHealth(exportDir string) (schema.RuntimeHealth, e
 				hasIPv4 = true
 			} else if strings.EqualFold(iface.State, "DOWN") {
 				outcome = "link_down"
-			} else {
-				missingIPv4 = true
 			}
 			health.Interfaces = append(health.Interfaces, schema.RuntimeInterface{
 				Name:    iface.Name,
@@ -73,17 +70,9 @@ func (s *System) CollectRuntimeHealth(exportDir string) (schema.RuntimeHealth, e
 				Outcome: outcome,
 			})
 		}
-		switch {
+		if hasIPv4 {
-		case hasIPv4 && !missingIPv4:
 			health.NetworkStatus = "OK"
-		case hasIPv4:
+		} else {
-			health.NetworkStatus = "PARTIAL"
-			health.Issues = append(health.Issues, schema.RuntimeIssue{
-				Code:        "dhcp_partial",
-				Severity:    "warning",
-				Description: "At least one interface did not obtain IPv4 connectivity.",
-			})
-		default:
 			health.NetworkStatus = "FAILED"
 			health.Issues = append(health.Issues, schema.RuntimeIssue{
 				Code:        "dhcp_failed",

audit/internal/webui/api.go

@@ -1679,6 +1679,56 @@ func (h *handler) handleAPIBenchmarkResults(w http.ResponseWriter, r *http.Reque
 	fmt.Fprint(w, renderBenchmarkResultsCard(h.opts.ExportDir))
 }
 
+// ── Hardware summary / component detail ──────────────────────────────────────
+
+// handleAPIHardwareSummary returns the hardware summary card HTML fragment for
+// htmx polling (hx-get="/api/hardware-summary" hx-swap="outerHTML").
+func (h *handler) handleAPIHardwareSummary(w http.ResponseWriter, _ *http.Request) {
+	w.Header().Set("Content-Type", "text/html; charset=utf-8")
+	w.Header().Set("Cache-Control", "no-store")
+	fmt.Fprint(w, renderHardwareSummaryCard(h.opts))
+}
+
+// handleAPIComponentDetail returns an HTML fragment describing the current and
+// historical status for one component type (cpu, memory, storage, gpu, psu).
+func (h *handler) handleAPIComponentDetail(w http.ResponseWriter, r *http.Request) {
+	compType := r.PathValue("type")
+	var exact, prefixes []string
+	var title string
+	switch compType {
+	case "cpu":
+		title = "CPU"
+		exact = []string{"cpu:all"}
+	case "memory":
+		title = "Memory"
+		exact = []string{"memory:all"}
+		prefixes = []string{"memory:"}
+	case "storage":
+		title = "Storage"
+		exact = []string{"storage:all"}
+		prefixes = []string{"storage:"}
+	case "gpu":
+		title = "GPU"
+		prefixes = []string{"pcie:gpu:"}
+	case "psu":
+		title = "PSU"
+		prefixes = []string{"psu:"}
+	default:
+		http.NotFound(w, r)
+		return
+	}
+
+	var records []app.ComponentStatusRecord
+	if h.opts.App != nil && h.opts.App.StatusDB != nil {
+		all := h.opts.App.StatusDB.All()
+		records = matchedRecords(all, exact, prefixes)
+	}
+
+	w.Header().Set("Content-Type", "text/html; charset=utf-8")
+	w.Header().Set("Cache-Control", "no-store")
+	fmt.Fprint(w, renderComponentDetail(title, records))
+}
+
 func (h *handler) rollbackPendingNetworkChange() error {
 	h.pendingNetMu.Lock()
 	pnc := h.pendingNet

audit/internal/webui/health_poller.go

@@ -0,0 +1,76 @@
+package webui
+
+import (
+	"bytes"
+	"context"
+	"log/slog"
+	"os/exec"
+	"time"
+
+	"bee/audit/internal/app"
+	"bee/audit/internal/collector"
+)
+
+const healthPollInterval = 60 * time.Second
+const psuIPMITimeout = 15 * time.Second
+
+// healthPoller runs periodic health checks for hardware components that do not
+// emit kernel log events (e.g. PSU). Results are written to ComponentStatusDB.
+type healthPoller struct {
+	statusDB *app.ComponentStatusDB
+}
+
+func newHealthPoller(statusDB *app.ComponentStatusDB) *healthPoller {
+	return &healthPoller{statusDB: statusDB}
+}
+
+func (p *healthPoller) start() {
+	goRecoverLoop("health poller", 5*time.Second, p.run)
+}
+
+func (p *healthPoller) run() {
+	ticker := time.NewTicker(healthPollInterval)
+	defer ticker.Stop()
+	for range ticker.C {
+		p.pollPSU()
+	}
+}
+
+func (p *healthPoller) pollPSU() {
+	if p.statusDB == nil {
+		return
+	}
+	ctx, cancel := context.WithTimeout(context.Background(), psuIPMITimeout)
+	defer cancel()
+
+	cmd := exec.CommandContext(ctx, "ipmitool", "sdr")
+	var out bytes.Buffer
+	cmd.Stdout = &out
+	if err := cmd.Run(); err != nil {
+		// IPMI not available or not a server — skip silently.
+		slog.Debug("health poller: ipmitool sdr unavailable", "err", err)
+		return
+	}
+
+	slots := collector.PSUSlotsFromSDR(out.String())
+	if len(slots) == 0 {
+		return
+	}
+
+	const source = "watchdog:psu"
+	for slot, psu := range slots {
+		key := "psu:" + slot
+		status := psu.Status
+		if status == "" {
+			status = "Unknown"
+		}
+		detail := ""
+		switch status {
+		case "Critical":
+			detail = "PSU sensor reported non-OK state"
+		case "Warning":
+			detail = "PSU sensor in warning state"
+		}
+		p.statusDB.Record(key, source, status, detail)
+	}
+}

audit/internal/webui/kmsg_watcher.go

@@ -73,6 +73,9 @@ func (w *kmsgWatcher) run() {
 			w.mu.Lock()
 			if w.window != nil {
 				w.recordEvent(evt)
+			} else {
+				evtCopy := evt
+				goRecoverOnce("kmsg flush immediate", func() { w.flushImmediate(evtCopy) })
 			}
 			w.mu.Unlock()
 		}
@@ -162,7 +165,9 @@ func (w *kmsgWatcher) flushWindow(window *kmsgWindow) {
 		for _, id := range evt.ids {
 			var key string
 			switch evt.category {
-			case "gpu", "pcie":
+			case "gpu":
+				key = "pcie:gpu:" + normalizeBDF(id)
+			case "pcie":
 				key = "pcie:" + normalizeBDF(id)
 			case "storage":
 				key = "storage:" + id
@@ -180,6 +185,54 @@ func (w *kmsgWatcher) flushWindow(window *kmsgWindow) {
 	}
 }
 
+// flushImmediate writes a single kmsg event directly to the status DB without a SAT window.
+// Called when an error is detected outside of any SAT task (always-on watching).
+func (w *kmsgWatcher) flushImmediate(evt kmsgEvent) {
+	if w.statusDB == nil {
+		return
+	}
+	const source = "watchdog:kmsg"
+	detail := "kernel: " + truncate(evt.raw, 120)
+
+	var severity string
+	for _, p := range platform.HardwareErrorPatterns {
+		if p.Re.MatchString(evt.raw) {
+			if p.Severity == "critical" {
+				severity = "Critical"
+			} else {
+				severity = "Warning"
+			}
+			break
+		}
+	}
+	if severity == "" {
+		severity = "Warning"
+	}
+
+	if len(evt.ids) == 0 {
+		key := "cpu:all"
+		if evt.category == "memory" {
+			key = "memory:all"
+		}
+		w.statusDB.Record(key, source, severity, detail)
+		return
+	}
+	for _, id := range evt.ids {
+		var key string
+		switch evt.category {
+		case "gpu":
+			key = "pcie:gpu:" + normalizeBDF(id)
+		case "pcie":
+			key = "pcie:" + normalizeBDF(id)
+		case "storage":
+			key = "storage:" + id
+		default:
+			key = "pcie:" + normalizeBDF(id)
+		}
+		w.statusDB.Record(key, source, severity, detail)
+	}
+}
+
 // parseKmsgLine parses a single /dev/kmsg line and returns an event if it matches
 // any pattern in platform.HardwareErrorPatterns.
 // kmsg format: "<priority>,<sequence>,<timestamp_usec>,-;message text"

audit/internal/webui/layout.go

@@ -17,6 +17,7 @@ func layoutHead(title string) string {
 <style>
 :root{--bg:#fff;--surface:#fff;--surface-2:#f9fafb;--border:rgba(34,36,38,.15);--border-lite:rgba(34,36,38,.1);--ink:rgba(0,0,0,.87);--muted:rgba(0,0,0,.6);--accent:#2185d0;--accent-dark:#1678c2;--crit-bg:#fff6f6;--crit-fg:#9f3a38;--crit-border:#e0b4b4;--ok-bg:#fcfff5;--ok-fg:#2c662d;--warn-bg:#fffaf3;--warn-fg:#573a08}
 *{box-sizing:border-box;margin:0;padding:0}
+dialog{margin:auto}
 body{font:14px/1.5 Lato,"Helvetica Neue",Arial,Helvetica,sans-serif;background:var(--bg);color:var(--ink);display:flex;min-height:100vh}
 a{color:var(--accent);text-decoration:none}
 /* Sidebar */

audit/internal/webui/pages.go

@@ -5,7 +5,9 @@ import (
 	"fmt"
 	"html"
 	"path/filepath"
+	"regexp"
 	"sort"
+	"strconv"
 	"strings"
 
 	"bee/audit/internal/app"
@@ -85,6 +87,7 @@ func renderPage(page string, opts HandlerOptions) string {
 		body +
 		`</div></div>` +
 		renderAuditModal() +
+		`<dialog id="component-detail-dialog" style="min-width:600px;max-width:900px;width:90vw;padding:0;border:1px solid var(--border);border-radius:8px;background:var(--surface)"><div id="component-detail-body" style="padding-bottom:20px"></div></dialog>` +
 		`<script>
 // Add copy button to every .terminal on the page
 document.querySelectorAll('.terminal').forEach(function(t){
@@ -94,6 +97,17 @@ document.querySelectorAll('.terminal').forEach(function(t){
   btn.onclick=function(){navigator.clipboard.writeText(t.textContent).then(function(){btn.textContent='Copied!';setTimeout(function(){btn.textContent='Copy';},1500);});};
   w.appendChild(btn);
 });
+function openComponentDetail(type) {
+  var dlg = document.getElementById('component-detail-dialog');
+  var body = document.getElementById('component-detail-body');
+  body.innerHTML = '<div style="padding:20px;color:var(--muted)">Loading…</div>';
+  dlg.showModal();
+  fetch('/api/components/' + type).then(function(r){ return r.text(); }).then(function(html){
+    body.innerHTML = html;
+  }).catch(function(){
+    body.innerHTML = '<div style="padding:20px;color:var(--crit-fg)">Error loading details.</div>';
+  });
+}
 </script>` +
 		`</body></html>`
 }
@@ -106,6 +120,14 @@ func renderDashboard(opts HandlerOptions) string {
 	b.WriteString(renderHardwareSummaryCard(opts))
 	b.WriteString(renderHealthCard(opts))
 	b.WriteString(renderMetrics())
+	b.WriteString(`<script>
+setInterval(function(){
+  fetch('/api/hardware-summary').then(function(r){return r.text();}).then(function(html){
+    var el=document.getElementById('hw-summary-card');
+    if(el){el.outerHTML=html;}
+  }).catch(function(){});
+},30000);
+</script>`)
 	return b.String()
 }
 
@@ -184,13 +206,14 @@ func renderAudit() string {
 }
 
 func renderHardwareSummaryCard(opts HandlerOptions) string {
+	const cardID = ` id="hw-summary-card"`
 	data, err := loadSnapshot(opts.AuditPath)
 	if err != nil {
-		return `<div class="card"><div class="card-head card-head-actions"><span>Hardware Summary</span><div class="card-head-buttons"><button class="btn btn-primary btn-sm" onclick="auditModalRun()">Run audit</button></div></div><div class="card-body"></div></div>`
+		return `<div class="card"` + cardID + `><div class="card-head card-head-actions"><span>Hardware Summary</span><div class="card-head-buttons"><button class="btn btn-primary btn-sm" onclick="auditModalRun()">Run audit</button></div></div><div class="card-body"></div></div>`
 	}
 	var ingest schema.HardwareIngestRequest
 	if err := json.Unmarshal(data, &ingest); err != nil {
-		return `<div class="card"><div class="card-head">Hardware Summary</div><div class="card-body"><span class="badge badge-err">Parse error</span></div></div>`
+		return `<div class="card"` + cardID + `><div class="card-head">Hardware Summary</div><div class="card-body"><span class="badge badge-err">Parse error</span></div></div>`
 	}
 	hw := ingest.Hardware
 
@@ -200,7 +223,7 @@ func renderHardwareSummaryCard(opts HandlerOptions) string {
 	}
 
 	var b strings.Builder
-	b.WriteString(`<div class="card"><div class="card-head">Hardware Summary</div><div class="card-body">`)
+	b.WriteString(`<div class="card"` + cardID + `><div class="card-head">Hardware Summary</div><div class="card-body">`)
 
 	// Server identity block above the component table.
 	{
@@ -229,22 +252,32 @@ func renderHardwareSummaryCard(opts HandlerOptions) string {
 	}
 
 	b.WriteString(`<table style="width:auto">`)
-	writeRow := func(label, value, badgeHTML string) {
+	// writeRow renders one component row. compType is the URL path segment for the detail
-		b.WriteString(fmt.Sprintf(`<tr><td style="padding:6px 14px 6px 0;font-weight:700;white-space:nowrap">%s</td><td style="padding:6px 0;color:var(--muted);font-size:13px">%s</td><td style="padding:6px 0 6px 12px">%s</td></tr>`,
+	// endpoint (e.g. "cpu"). Pass "" for rows that have no detail view.
-			html.EscapeString(label), html.EscapeString(value), badgeHTML))
+	writeRow := func(label, value, badgeHTML, compType string) {
+		var labelHTML string
+		if compType != "" {
+			labelHTML = fmt.Sprintf(
+				`<span style="cursor:pointer;text-decoration:underline dotted;text-underline-offset:3px" onclick="openComponentDetail('%s')">%s</span>`,
+				compType, html.EscapeString(label))
+		} else {
+			labelHTML = html.EscapeString(label)
+		}
+		fmt.Fprintf(&b, `<tr><td style="padding:6px 14px 6px 0;font-weight:700;white-space:nowrap">%s</td><td style="padding:6px 0;color:var(--muted);font-size:13px">%s</td><td style="padding:6px 0 6px 12px">%s</td></tr>`,
+			labelHTML, html.EscapeString(value), badgeHTML)
 	}
 
 	writeRow("CPU", hwDescribeCPU(hw),
-		renderComponentChips(matchedRecords(records, []string{"cpu:all"}, nil)))
+		renderComponentChips(matchedRecords(records, []string{"cpu:all"}, nil)), "cpu")
 
 	writeRow("Memory", hwDescribeMemory(hw),
-		renderComponentChips(matchedRecords(records, []string{"memory:all"}, []string{"memory:"})))
+		renderComponentChips(matchedRecords(records, []string{"memory:all"}, []string{"memory:"})), "memory")
 
 	writeRow("Storage", hwDescribeStorage(hw),
-		renderComponentChips(matchedRecords(records, []string{"storage:all"}, []string{"storage:"})))
+		renderComponentChips(matchedRecords(records, []string{"storage:all"}, []string{"storage:"})), "storage")
 
 	writeRow("GPU", hwDescribeGPU(hw),
-		renderComponentChips(matchedRecords(records, nil, []string{"pcie:gpu:"})))
+		renderComponentChips(matchedRecords(records, nil, []string{"pcie:gpu:"})), "gpu")
 
 	psuMatched := matchedRecords(records, nil, []string{"psu:"})
 	if len(psuMatched) == 0 && len(hw.PowerSupplies) > 0 {
@@ -252,10 +285,10 @@ func renderHardwareSummaryCard(opts HandlerOptions) string {
 		psuStatus := hwPSUStatus(hw.PowerSupplies)
 		psuMatched = []app.ComponentStatusRecord{{ComponentKey: "psu:ipmi", Status: psuStatus}}
 	}
-	writeRow("PSU", hwDescribePSU(hw), renderComponentChips(psuMatched))
+	writeRow("PSU", hwDescribePSU(hw), renderComponentChips(psuMatched), "psu")
 
 	if nicDesc := hwDescribeNIC(hw); nicDesc != "" {
-		writeRow("Network", nicDesc, "")
+		writeRow("Network", nicDesc, "", "")
 	}
 
 	b.WriteString(`</table>`)
@@ -614,7 +647,7 @@ func buildRuntimeNetworkRow(health schema.RuntimeHealth) runtimeHealthRow {
 	if status == "" {
 		status = "UNKNOWN"
 	}
-	issue := runtimeIssueDescriptions(health.Issues, "dhcp_partial", "dhcp_failed")
+	issue := runtimeIssueDescriptions(health.Issues, "dhcp_failed")
 	return runtimeHealthRow{Title: "Network", Status: status, Source: "ListInterfaces / DHCP", Issue: issue}
 }
 
@@ -672,12 +705,12 @@ func buildRuntimeServicesRow(health schema.RuntimeHealth) runtimeHealthRow {
 	nonActive := make([]string, 0)
 	for _, svc := range health.Services {
 		state := strings.TrimSpace(strings.ToLower(svc.Status))
-		// "activating" and "deactivating" are transient states for oneshot services
+		// "inactive" is OK for oneshot services that have completed successfully
-		// (RemainAfterExit=yes) — the service is running normally, not failed.
+		// (bee-sshsetup, bee-preflight, bee-audit, bee-network, etc.).
-		// Only "failed" and "inactive" (after services should be running) are problems.
+		// Only "failed" is a genuine problem.
 		switch state {
-		case "active", "activating", "deactivating", "reloading":
+		case "active", "activating", "deactivating", "reloading", "inactive":
-			// OK — service is running or transitioning normally
+			// OK — service is running, transitioning normally, or completed successfully
 		default:
 			nonActive = append(nonActive, svc.Name+"="+svc.Status)
 		}
@@ -999,3 +1032,200 @@ func rowIssueHTML(issue string) string {
 	}
 	return html.EscapeString(issue)
 }
+
+var aerStatusRe = regexp.MustCompile(`aer_status:\s*0x([0-9a-fA-F]{1,8})`)
+
+// decodeAERStatus parses an AER status hex value from a kernel error detail string
+// and returns a human-readable list of set bit names with correctable/uncorrectable label,
+// or "" if no AER status is found.
+func decodeAERStatus(detail string) string {
+	m := aerStatusRe.FindStringSubmatch(detail)
+	if m == nil {
+		return ""
+	}
+	v64, err := strconv.ParseUint(m[1], 16, 32)
+	if err != nil {
+		return ""
+	}
+	val := uint32(v64)
+
+	type bitDef struct {
+		bit  uint32
+		name string
+	}
+	corrBits := []bitDef{
+		{0, "Receiver Error"}, {6, "Replay Timer Timeout"}, {7, "Advisory Non-Fatal"},
+		{8, "Corrected Internal Error"}, {9, "Header Log Overflow"},
+		{13, "Replay Num Rollover"}, {14, "Bad DLLP"}, {15, "Bad TLP"},
+	}
+	uncorrBits := []bitDef{
+		{4, "Data Link Protocol Error"}, {5, "Surprise Down Error"},
+		{12, "Poisoned TLP Received"}, {13, "Flow Control Protocol Error"},
+		{14, "Completion Timeout"}, {15, "Completer Abort"}, {16, "Unexpected Completion"},
+		{17, "Receiver Overflow"}, {18, "Malformed TLP"}, {19, "ECRC Error"},
+		{20, "Unsupported Request Error"}, {21, "ACS Violation"}, {22, "Uncorrectable Internal Error"},
+	}
+	var corrNames, uncorrNames []string
+	for _, b := range corrBits {
+		if val&(1<<b.bit) != 0 {
+			corrNames = append(corrNames, b.name)
+		}
+	}
+	for _, b := range uncorrBits {
+		if val&(1<<b.bit) != 0 {
+			uncorrNames = append(uncorrNames, b.name)
+		}
+	}
+	if len(corrNames) >= len(uncorrNames) && len(corrNames) > 0 {
+		return strings.Join(corrNames, ", ") + " (correctable)"
+	}
+	if len(uncorrNames) > 0 {
+		return strings.Join(uncorrNames, ", ") + " (uncorrectable)"
+	}
+	return fmt.Sprintf("unknown bits: 0x%08x", val)
+}
+
+// renderSparkline returns a small inline SVG showing non-OK events over time.
+// Events are positioned proportionally along the time axis; if all share the same
+// timestamp they are spaced evenly. Width is always 100px.
+func renderSparkline(history []app.ComponentStatusEntry) string {
+	const (
+		svgW = 100
+		svgH = 20
+		barW = 3
+		barH = 14
+	)
+	var events []app.ComponentStatusEntry
+	for _, e := range history {
+		if e.Status != "OK" {
+			events = append(events, e)
+		}
+	}
+	if len(events) == 0 {
+		return ""
+	}
+	n := len(events)
+	barColor := func(status string) string {
+		if status == "Critical" {
+			return "#c0392b"
+		}
+		return "#d97706"
+	}
+	yTop := (svgH - barH) / 2
+
+	var bars strings.Builder
+	if n == 1 {
+		x := (svgW - barW) / 2
+		fmt.Fprintf(&bars, `<rect x="%d" y="%d" width="%d" height="%d" fill="%s" rx="1"/>`,
+			x, yTop, barW, barH, barColor(events[0].Status))
+	} else {
+		minT := events[0].At
+		maxT := events[n-1].At
+		dur := maxT.Sub(minT).Seconds()
+		for i, e := range events {
+			var x int
+			if dur <= 0 {
+				step := svgW / n
+				x = i*step + (step-barW)/2
+			} else {
+				frac := e.At.Sub(minT).Seconds() / dur
+				x = int(frac * float64(svgW-barW))
+			}
+			fmt.Fprintf(&bars, `<rect x="%d" y="%d" width="%d" height="%d" fill="%s" rx="1"/>`,
+				x, yTop, barW, barH, barColor(e.Status))
+		}
+	}
+	return fmt.Sprintf(
+		`<svg width="%d" height="%d" style="display:inline-block;vertical-align:middle;margin-left:6px;flex-shrink:0" xmlns="http://www.w3.org/2000/svg">`+
+			`<rect x="0" y="0" width="%d" height="%d" fill="var(--surface-alt,#ebebeb)" rx="3"/>%s</svg>`,
+		svgW, svgH, svgW, svgH, bars.String())
+}
+
+// renderComponentDetail renders a modal content fragment for one component type.
+// Called by handleAPIComponentDetail and displayed inside #component-detail-dialog.
+func renderComponentDetail(title string, records []app.ComponentStatusRecord) string {
+	var b strings.Builder
+	fmt.Fprintf(&b, `<div style="padding:20px 24px 0">`)
+	fmt.Fprintf(&b, `<div style="display:flex;align-items:center;justify-content:space-between;margin-bottom:16px">`)
+	fmt.Fprintf(&b, `<span style="font-size:16px;font-weight:700">%s — Status Detail</span>`, html.EscapeString(title))
+	b.WriteString(`<button class="btn btn-sm btn-secondary" onclick="document.getElementById('component-detail-dialog').close()">Close</button>`)
+	b.WriteString(`</div>`)
+
+	if len(records) == 0 {
+		b.WriteString(`<p style="color:var(--muted)">No status data recorded yet for this component type.</p>`)
+		b.WriteString(`</div>`)
+		return b.String()
+	}
+
+	sort.Slice(records, func(i, j int) bool {
+		return records[i].ComponentKey < records[j].ComponentKey
+	})
+
+	for _, rec := range records {
+		letter, cls := chipLetterClass(rec.Status)
+
+		// Count non-OK events across the full history for the badge + sparkline.
+		warnCount := 0
+		for _, e := range rec.History {
+			if e.Status != "OK" {
+				warnCount++
+			}
+		}
+
+		fmt.Fprintf(&b, `<div style="margin-bottom:20px">`)
+		fmt.Fprintf(&b, `<div style="display:flex;align-items:center;gap:8px;margin-bottom:8px;flex-wrap:wrap">`)
+		fmt.Fprintf(&b, `<span class="chip %s">%s</span>`, cls, letter)
+		fmt.Fprintf(&b, `<span style="font-weight:700;font-size:13px">%s</span>`, html.EscapeString(rec.ComponentKey))
+		if !rec.LastCheckedAt.IsZero() {
+			fmt.Fprintf(&b, `<span style="color:var(--muted);font-size:12px">checked %s</span>`, rec.LastCheckedAt.Format("2006-01-02 15:04:05"))
+		}
+		if warnCount > 0 {
+			noun := "events"
+			if warnCount == 1 {
+				noun = "event"
+			}
+			fmt.Fprintf(&b,
+				`<span style="font-size:11px;background:var(--warn-bg,#fffbeb);color:var(--warn-fg,#92400e);border:1px solid var(--warn-border,#fde68a);border-radius:10px;padding:1px 7px;white-space:nowrap">%d %s</span>`,
+				warnCount, noun)
+			b.WriteString(renderSparkline(rec.History))
+		}
+		b.WriteString(`</div>`)
+
+		if rec.ErrorSummary != "" {
+			fmt.Fprintf(&b, `<div style="font-size:12px;margin-bottom:4px;color:var(--muted)">%s</div>`, html.EscapeString(rec.ErrorSummary))
+			if decoded := decodeAERStatus(rec.ErrorSummary); decoded != "" {
+				fmt.Fprintf(&b,
+					`<div style="font-size:12px;margin-bottom:8px;color:var(--muted)"><span style="background:var(--surface-alt,#f5f5f5);border-radius:4px;padding:1px 6px;font-family:monospace">AER: %s</span></div>`,
+					html.EscapeString(decoded))
+			}
+		}
+
+		// History table — newest first, cap at 20 entries.
+		history := rec.History
+		if len(history) > 20 {
+			history = history[len(history)-20:]
+		}
+		b.WriteString(`<table style="width:100%;font-size:12px;border-collapse:collapse">`)
+		b.WriteString(`<tr style="color:var(--muted)"><th style="text-align:left;padding:2px 10px 2px 0;white-space:nowrap">Time</th><th style="text-align:left;padding:2px 10px 2px 0">Status</th><th style="text-align:left;padding:2px 10px 2px 0">Source</th><th style="text-align:left;padding:2px 0">Detail</th></tr>`)
+		for i := len(history) - 1; i >= 0; i-- {
+			e := history[i]
+			eLetter, eCls := chipLetterClass(e.Status)
+			detail := e.Detail
+			if detail == "" {
+				detail = "—"
+			}
+			fmt.Fprintf(&b,
+				`<tr><td style="padding:3px 10px 3px 0;white-space:nowrap;color:var(--muted)">%s</td><td style="padding:3px 10px 3px 0"><span class="chip %s" style="font-size:10px;width:16px;height:16px">%s</span></td><td style="padding:3px 10px 3px 0;white-space:nowrap">%s</td><td style="padding:3px 0;color:var(--muted)">%s</td></tr>`,
+				html.EscapeString(e.At.Format("2006-01-02 15:04:05")),
+				eCls, eLetter,
+				html.EscapeString(e.Source),
+				html.EscapeString(detail),
+			)
+		}
+		b.WriteString(`</table>`)
+		b.WriteString(`</div>`)
+	}
+
+	b.WriteString(`</div>`)
+	return b.String()
+}

audit/internal/webui/server.go

@@ -221,6 +221,11 @@ func NewHandler(opts HandlerOptions) http.Handler {
 		h.kmsg = newKmsgWatcher(opts.App.StatusDB)
 		h.kmsg.start()
 		globalQueue.kmsgWatcher = h.kmsg
+
+		// Start periodic health poller for components that don't emit kernel log events (e.g. PSU).
+		if opts.App.StatusDB != nil {
+			newHealthPoller(opts.App.StatusDB).start()
+		}
 	}
 
 	globalQueue.startWorker(&opts)
@@ -328,6 +333,10 @@ func NewHandler(opts HandlerOptions) http.Handler {
 	mux.HandleFunc("GET /api/install/disks", h.handleAPIInstallDisks)
 	mux.HandleFunc("POST /api/install/run", h.handleAPIInstallRun)
 
+	// Hardware component detail (fragment for modal in Hardware Summary card)
+	mux.HandleFunc("GET /api/hardware-summary", h.handleAPIHardwareSummary)
+	mux.HandleFunc("GET /api/components/{type}", h.handleAPIComponentDetail)
+
 	// Metrics — SSE stream of live sensor data + server-side SVG charts + CSV export
 	mux.HandleFunc("GET /api/metrics/stream", h.handleAPIMetricsStream)
 	mux.HandleFunc("GET /api/metrics/latest", h.handleAPIMetricsLatest)

bible-local/docs/grub-bitmap-error-history.md

@@ -0,0 +1,312 @@
+# GRUB Bitmap Error History
+
+## Symptom
+
+On some servers GRUB prints:
+
+```text
+error: null src bitmap in grub_video_bitmap_create_scaled.
+Press any key to continue...
+```
+
+The important new observation as of `v10.7` is:
+
+- the error still appears even when the logo image block is removed from
+  `iso/builder/config/bootloaders/grub-efi/live-theme/theme.txt`
+- therefore the current error can no longer be explained only by
+  `bee-logo.png` / `bee-logo.tga`
+
+That does not prove the theme system is healthy. It proves only that the
+currently remaining failure is deeper than "bad logo file".
+
+## Current State
+
+Current source files:
+
+- [iso/builder/config/bootloaders/grub-efi/live-theme/theme.txt](/Users/mchusavitin/Documents/git/bee/iso/builder/config/bootloaders/grub-efi/live-theme/theme.txt:1)
+  has no `image` block anymore
+- [iso/builder/config/bootloaders/grub-efi/config.cfg](/Users/mchusavitin/Documents/git/bee/iso/builder/config/bootloaders/grub-efi/config.cfg:1)
+  still does `insmod tga` and then `source /boot/grub/theme.cfg`
+
+Implication:
+
+- if the error still fires, the trigger is likely elsewhere in GRUB theme
+  rendering or in the assets/config GRUB resolves while sourcing `theme.cfg`
+- the old "PNG parser fragility" story is no longer a sufficient explanation
+  for the current failure mode
+
+Current artifact facts:
+
+- the provided `easy-bee-nvidia-v10.7-amd64.logs` build logs reference
+  `linux-image-6.1.0-45`
+- the provided `easy-bee-nvidia-v10.7-amd64.iso` contains
+  `live/initrd.img-6.1.0-45-amd64` and `live/vmlinuz-6.1.0-45-amd64`
+- a later `BOOT FAILED!` screenshot showed `live/initrd.img-6.1.0-44-amd64`
+  and `live/vmlinuz-6.1.0-44-amd64`
+
+Implication:
+
+- the `BOOT FAILED!` screenshot is not from the same artifact as the provided
+  `v10.7` ISO/log set
+- until the exact ISO filename and checksum are tied to that failure, the
+  GRUB bitmap issue and the live-boot failure must be treated as separate
+  problems
+
+## Chronology
+
+### 1. Initial bee GRUB theme introduction
+
+Relevant commit:
+
+- `d52ec67` `Stability hardening, build script fixes, GRUB bee logo`
+
+What changed:
+
+- bee-branded GRUB theme introduced
+- image block with explicit `width` / `height`
+
+Observed result:
+
+- bitmap error appeared
+
+### 2. Remove explicit scaling dimensions
+
+Relevant commit:
+
+- `aa284ae` `fix(iso): avoid grub logo scaling error`
+
+What changed:
+
+- removed `width = 400`
+- removed `height = 400`
+
+Reason stated by the change:
+
+- try to avoid the scaling path
+
+Observed result:
+
+- error persisted
+
+Conclusion:
+
+- explicit width/height were not the sole trigger
+
+### 3. Rework PNG handling and menu rendering
+
+Relevant commit:
+
+- `6112094` `fix(grub): fix bitmap error and menu rendering`
+
+Commit message says the change was intended to:
+
+- convert `bee-logo.png` to RGBA and strip metadata
+- move `terminal_output gfxterm` before `insmod png` / theme load
+- remove ASCII banner from GRUB menu area
+- fix theme typography/layout fields
+
+Observed result:
+
+- error persisted
+
+Notes:
+
+- this was still operating under the assumption that the issue was the PNG
+  payload or the order of gfxterm/theme init
+
+### 4. Convert logo PNG back to RGB
+
+Relevant commit:
+
+- `333c44f` `Fix GRUB splash: convert bee-logo.png from RGBA to RGB`
+
+Intended reason:
+
+- GRUB might dislike RGBA PNG and want RGB PNG
+
+Observed result:
+
+- error still persisted according to later project notes
+
+### 5. Add post-build canonical GRUB/isolinux sync
+
+Relevant commit:
+
+- `0cdfbc5` `fix(iso): restore boot UX and boot logs`
+
+What this introduced:
+
+- post-`lb build` rewriting of `binary/boot/grub/grub.cfg`
+- post-`lb build` rewriting of `binary/isolinux/live.cfg`
+- forced rebuild of `binary_checksums`, `binary_iso`, `binary_zsync`
+
+Why it was added:
+
+- restore canonical EASY-BEE boot UX after live-build wrote its own bootloader
+  outputs
+- restore expected boot menu and logs
+
+Important note:
+
+- this commit did not directly solve the bitmap issue
+- it added a second layer of bootloader mutation after live-build
+
+### 6. Switch from PNG to TGA
+
+Relevant commit:
+
+- `626763e` `Fix GRUB bitmap error: switch from PNG to TGA for splash logo`
+
+Commit message says:
+
+- GRUB PNG reader was considered fragile
+- switch to uncompressed 24-bit TGA
+- `config.cfg`: `insmod png` -> `insmod tga`
+- `theme.txt`: `bee-logo.png` -> `bee-logo.tga`
+
+Observed result:
+
+- this did not eliminate the problem in the current lineage
+- today the system still errors even after the entire image block was removed
+
+Conclusion:
+
+- switching PNG -> TGA was not a durable root-cause fix
+
+### 7. Patch EFI image after build
+
+Relevant commit:
+
+- `4f20c92` `Make UEFI boot safe and remove GRUB logo`
+
+What this introduced:
+
+- `sync_efi_grub_theme_assets`
+- direct `mtools` patching of `efi.img`
+- copying `config.cfg`, `theme.cfg`, and `live-theme/*` into the EFI FAT image
+- removal of the logo image block from `theme.txt`
+
+Why it was added:
+
+- make UEFI path "safe"
+- keep EFI GRUB image aligned with canonical bootloader assets
+
+Observed result:
+
+- later this became the direct cause of `Disk full` during build once
+  `bee-logo.tga` was large enough
+- and even with the logo removed from `theme.txt`, the bitmap error still
+  remained
+
+Conclusion:
+
+- EFI post-build patching increased build complexity
+- removing the logo alone did not remove the runtime GRUB error
+
+### 8. Remove ASCII logo banners
+
+Relevant commit:
+
+- `14505ef` `Remove easy bee ASCII logo banners`
+
+What changed:
+
+- web loading page ASCII cleanup only
+
+Relevance here:
+
+- none for GRUB bitmap error
+- included here only to avoid confusion with other "logo removal" work
+
+### 9. Remove EFI post-build patching
+
+Relevant commit:
+
+- `5dc022d` `Drop post-build EFI bootloader patching`
+
+Why it was done:
+
+- stop mutating `efi.img` post-build
+- remove dependence on `mtools` for EFI patching
+- remove the `Disk full` failure mode
+
+Impact:
+
+- this did not target the GRUB bitmap error directly
+- it targeted build-system complexity and EFI image overflow
+
+### 10. Restore only GRUB/isolinux post-build sync
+
+Relevant commit:
+
+- `42774d4` `Restore post-build GRUB and isolinux sync`
+
+Why it was needed:
+
+- removing all post-build sync caused final ISO validation to fail with
+  missing canonical EASY-BEE boot entries
+- memtest was still fine, but final GRUB menu was no longer canonical
+
+What it restored:
+
+- only `binary/boot/grub/grub.cfg`
+- only `binary/isolinux/live.cfg`
+
+What it did not restore:
+
+- no EFI FAT image patching
+- no `mtools` path
+
+## What Is Proven False
+
+The current evidence rules out several simplistic explanations:
+
+- "the error is only caused by explicit image scaling"
+- "the error is only caused by PNG vs TGA"
+- "the error is only caused by the logo file itself"
+
+Why:
+
+- scaling dimensions were removed and error persisted
+- PNG was replaced with TGA and error still survived in the lineage
+- the image block itself is now absent, and the error still occurs
+
+## Working Hypotheses Left
+
+The remaining plausible layers are:
+
+- GRUB theme engine still tries to render some bitmap-related element even
+  without the logo image block
+- GRUB is resolving stale theme assets from the built EFI/ISO path rather than
+  what we think the source tree says
+- `theme.cfg` / `theme.txt` / GRUB module loading order still triggers a bitmap
+  code path elsewhere
+- live-build may still package a stale `theme.txt` or stale `live-theme`
+  directory into the final image
+- the GRUB environment on the failing hardware may behave differently from the
+  assumptions in our source tree
+
+## Decision Boundary
+
+Before making another change, the next step should be evidence gathering from
+the real built artifact, not another speculative edit.
+
+That means checking on the actual built ISO or EFI image:
+
+- exact `boot/grub/theme.cfg`
+- exact `boot/grub/live-theme/theme.txt`
+- exact contents of `boot/grub/live-theme/`
+- whether the final image still contains a stale logo reference
+- whether the EFI path and non-EFI path differ
+
+## Relevant Commits
+
+- `d52ec67` `Stability hardening, build script fixes, GRUB bee logo`
+- `aa284ae` `fix(iso): avoid grub logo scaling error`
+- `6112094` `fix(grub): fix bitmap error and menu rendering`
+- `333c44f` `Fix GRUB splash: convert bee-logo.png from RGBA to RGB`
+- `0cdfbc5` `fix(iso): restore boot UX and boot logs`
+- `626763e` `Fix GRUB bitmap error: switch from PNG to TGA for splash logo`
+- `4f20c92` `Make UEFI boot safe and remove GRUB logo`
+- `5dc022d` `Drop post-build EFI bootloader patching`
+- `42774d4` `Restore post-build GRUB and isolinux sync`

iso/builder/auto/config

@@ -38,7 +38,7 @@ lb config noauto \
     --memtest memtest86+ \
     --iso-volume "${LB_ISO_VOLUME}" \
     --iso-application "EASY-BEE-${BEE_GPU_VENDOR_UPPER:-NVIDIA}" \
-    --bootappend-live "boot=live live-media-label=${LB_ISO_VOLUME} components video=1920x1080 console=ttyS0,115200n8 console=tty0 loglevel=3 systemd.show_status=1 username=bee user-fullname=Bee modprobe.blacklist=nouveau,snd_hda_intel,snd_hda_codec_realtek,snd_hda_codec_generic,soundcore" \
+    --bootappend-live "boot=live live-media=/dev/disk/by-label/${LB_ISO_VOLUME} live-media-label=${LB_ISO_VOLUME} components video=1920x1080 console=ttyS0,115200n8 console=tty0 loglevel=3 systemd.show_status=1 username=bee user-fullname=Bee modprobe.blacklist=nouveau,snd_hda_intel,snd_hda_codec_realtek,snd_hda_codec_generic,soundcore" \
     --debootstrap-options "--include=ca-certificates" \
     --apt-recommends false \
     --chroot-squashfs-compression-type zstd \

iso/builder/build-in-container.sh

@@ -8,7 +8,7 @@ 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}"
+CACHE_DIR="${BEE_BUILDER_CACHE_DIR:-${REPO_ROOT}/dist/cache}"
 AUTH_KEYS=""
 CLEAN_CACHE=0
 VARIANT="all"
@@ -54,14 +54,14 @@ if [ "$CLEAN_CACHE" = "1" ]; then
            "${CACHE_DIR:?}/bee" \
            "${CACHE_DIR:?}/lb-packages"
     echo "=== cleaning live-build work dirs ==="
-    rm -rf "${REPO_ROOT}/dist/live-build-work-nvidia"
+    rm -rf "${REPO_ROOT}/dist/cache/live-build-work-nvidia"
-    rm -rf "${REPO_ROOT}/dist/live-build-work-nvidia-legacy"
+    rm -rf "${REPO_ROOT}/dist/cache/live-build-work-nvidia-legacy"
-    rm -rf "${REPO_ROOT}/dist/live-build-work-amd"
+    rm -rf "${REPO_ROOT}/dist/cache/live-build-work-amd"
-    rm -rf "${REPO_ROOT}/dist/live-build-work-nogpu"
+    rm -rf "${REPO_ROOT}/dist/cache/live-build-work-nogpu"
-    rm -rf "${REPO_ROOT}/dist/overlay-stage-nvidia"
+    rm -rf "${REPO_ROOT}/dist/cache/overlay-stage-nvidia"
-    rm -rf "${REPO_ROOT}/dist/overlay-stage-nvidia-legacy"
+    rm -rf "${REPO_ROOT}/dist/cache/overlay-stage-nvidia-legacy"
-    rm -rf "${REPO_ROOT}/dist/overlay-stage-amd"
+    rm -rf "${REPO_ROOT}/dist/cache/overlay-stage-amd"
-    rm -rf "${REPO_ROOT}/dist/overlay-stage-nogpu"
+    rm -rf "${REPO_ROOT}/dist/cache/overlay-stage-nogpu"
     echo "=== caches cleared, proceeding with build ==="
 fi
 

iso/builder/build.sh

@@ -51,8 +51,8 @@ case "$BUILD_VARIANT" in
         ;;
 esac
 
-BUILD_WORK_DIR="${DIST_DIR}/live-build-work-${BUILD_VARIANT}"
+BUILD_WORK_DIR="${DIST_DIR}/cache/live-build-work-${BUILD_VARIANT}"
-OVERLAY_STAGE_DIR="${DIST_DIR}/overlay-stage-${BUILD_VARIANT}"
+OVERLAY_STAGE_DIR="${DIST_DIR}/cache/overlay-stage-${BUILD_VARIANT}"
 
 export BEE_GPU_VENDOR BEE_NVIDIA_MODULE_FLAVOR BUILD_VARIANT
 
@@ -63,7 +63,7 @@ export PATH="$PATH:/usr/local/go/bin"
 
 # Allow git to read the bind-mounted repo (different UID inside container).
 git config --global safe.directory "${REPO_ROOT}"
-mkdir -p "${DIST_DIR}"
+mkdir -p "${DIST_DIR}/cache" "${DIST_DIR}/release"
 mkdir -p "${CACHE_ROOT}"
 : "${GOCACHE:=${CACHE_ROOT}/go-build}"
 : "${GOMODCACHE:=${CACHE_ROOT}/go-mod}"
@@ -516,12 +516,12 @@ validate_iso_live_boot_entries() {
         exit 1
     fi
 
-    grep -q 'menuentry "EASY-BEE"' "$grub_cfg" || {
+    grep -q 'menuentry "EASY-BEE v' "$grub_cfg" || {
         echo "ERROR: GRUB default EASY-BEE entry is missing" >&2
         rm -f "$grub_cfg" "$isolinux_cfg"
         exit 1
     }
-    grep -q 'menuentry "EASY-BEE -- load to RAM (toram)"' "$grub_cfg" || {
+    grep -q 'menuentry "EASY-BEE v.* -- load to RAM (toram)"' "$grub_cfg" || {
         echo "ERROR: GRUB toram entry is missing" >&2
         rm -f "$grub_cfg" "$isolinux_cfg"
         exit 1
@@ -562,40 +562,38 @@ validate_iso_live_boot_entries() {
     echo "=== live boot validation OK ==="
 }
 
-validate_iso_grub_theme_assets() {
+validate_iso_grub_assets() {
     iso_path="$1"
-    echo "=== validating GRUB theme assets in ISO ==="
+    echo "=== validating GRUB assets in ISO ==="
 
     [ -f "$iso_path" ] || {
-        echo "ERROR: ISO not found for GRUB theme validation: $iso_path" >&2
+        echo "ERROR: ISO not found for GRUB asset validation: $iso_path" >&2
         exit 1
     }
     require_iso_reader "$iso_path" >/dev/null 2>&1 || {
-        echo "ERROR: ISO reader unavailable for GRUB theme validation" >&2
+        echo "ERROR: ISO reader unavailable for GRUB asset validation" >&2
         exit 1
     }
 
     iso_files="$(mktemp)"
     iso_list_files "$iso_path" > "$iso_files" || {
-        echo "ERROR: failed to list ISO files for GRUB theme validation" >&2
+        echo "ERROR: failed to list ISO files for GRUB asset validation" >&2
         rm -f "$iso_files"
         exit 1
     }
 
     for required in \
         boot/grub/config.cfg \
-        boot/grub/theme.cfg \
+        boot/grub/grub.cfg; do
-        boot/grub/live-theme/theme.txt \
-        boot/grub/live-theme/bee-logo.tga; do
         grep -q "^${required}$" "$iso_files" || {
-            echo "ERROR: missing GRUB theme asset in ISO: ${required}" >&2
+            echo "ERROR: missing GRUB asset in ISO: ${required}" >&2
             rm -f "$iso_files"
             exit 1
         }
     done
 
     rm -f "$iso_files"
-    echo "=== GRUB theme validation OK ==="
+    echo "=== GRUB asset validation OK ==="
 }
 
 validate_iso_nvidia_runtime() {
@@ -610,29 +608,37 @@ validate_iso_nvidia_runtime() {
 
     squashfs_tmp="$(mktemp)"
     squashfs_list="$(mktemp)"
-    iso_read_member "$iso_path" live/filesystem.squashfs "$squashfs_tmp" || {
+    iso_files="$(mktemp)"
-        rm -f "$squashfs_tmp" "$squashfs_list"
+    iso_list_files "$iso_path" > "$iso_files" || {
-        nvidia_runtime_fail "failed to extract live/filesystem.squashfs from ISO"
+        rm -f "$squashfs_tmp" "$squashfs_list" "$iso_files"
-    }
+        nvidia_runtime_fail "failed to list ISO files for NVIDIA runtime validation"
-    unsquashfs -ll "$squashfs_tmp" > "$squashfs_list" 2>/dev/null || {
-        rm -f "$squashfs_tmp" "$squashfs_list"
-        nvidia_runtime_fail "failed to inspect filesystem.squashfs from ISO"
     }
+    grep '^live/.*\.squashfs$' "$iso_files" | while IFS= read -r squashfs_member; do
+        iso_read_member "$iso_path" "$squashfs_member" "$squashfs_tmp" || {
+            rm -f "$squashfs_tmp" "$squashfs_list" "$iso_files"
+            nvidia_runtime_fail "failed to extract $squashfs_member from ISO"
+        }
+        unsquashfs -ll "$squashfs_tmp" >> "$squashfs_list" 2>/dev/null || {
+            rm -f "$squashfs_tmp" "$squashfs_list" "$iso_files"
+            nvidia_runtime_fail "failed to inspect $squashfs_member from ISO"
+        }
+        : > "$squashfs_tmp"
+    done
 
     grep -Eq 'usr/bin/dcgmi$' "$squashfs_list" || {
-        rm -f "$squashfs_tmp" "$squashfs_list"
+        rm -f "$squashfs_tmp" "$squashfs_list" "$iso_files"
         nvidia_runtime_fail "dcgmi missing from final NVIDIA ISO"
     }
     grep -Eq 'usr/bin/nv-hostengine$' "$squashfs_list" || {
-        rm -f "$squashfs_tmp" "$squashfs_list"
+        rm -f "$squashfs_tmp" "$squashfs_list" "$iso_files"
         nvidia_runtime_fail "nv-hostengine missing from final NVIDIA ISO"
     }
     grep -Eq 'usr/bin/dcgmproftester([0-9]+)?$' "$squashfs_list" || {
-        rm -f "$squashfs_tmp" "$squashfs_list"
+        rm -f "$squashfs_tmp" "$squashfs_list" "$iso_files"
         nvidia_runtime_fail "dcgmproftester missing from final NVIDIA ISO"
     }
 
-    rm -f "$squashfs_tmp" "$squashfs_list"
+    rm -f "$squashfs_tmp" "$squashfs_list" "$iso_files"
     echo "=== NVIDIA runtime validation OK ==="
 }
 
@@ -726,26 +732,26 @@ write_canonical_grub_cfg() {
     kernel="$2"
     append_live="$3"
     initrd="$4"
+    version_label="${PROJECT_VERSION_EFFECTIVE}"
 
     cat > "$cfg" <<EOF
 source /boot/grub/config.cfg
 
-menuentry "EASY-BEE" {
+menuentry "EASY-BEE v${version_label}" {
     linux   ${kernel} ${append_live} nomodeset bee.nvidia.mode=normal pci=realloc net.ifnames=0 biosdevname=0 mitigations=off transparent_hugepage=always numa_balancing=disable pcie_aspm=off intel_idle.max_cstate=1 processor.max_cstate=1 nowatchdog nosoftlockup
     initrd  ${initrd}
 }
 
-menuentry "EASY-BEE -- load to RAM (toram)" {
+menuentry "EASY-BEE v${version_label} -- load to RAM (toram)" {
     linux   ${kernel} ${append_live} toram nomodeset bee.nvidia.mode=normal pci=realloc net.ifnames=0 biosdevname=0 mitigations=off transparent_hugepage=always numa_balancing=disable pcie_aspm=off intel_idle.max_cstate=1 processor.max_cstate=1 nowatchdog nosoftlockup
     initrd  ${initrd}
 }
 
-menuentry "EASY-BEE -- no GUI / no X11" {
+menuentry "EASY-BEE v${version_label} -- no GUI / no X11" {
     linux   ${kernel} ${append_live} nomodeset bee.gui=off bee.nvidia.mode=gsp-off pci=realloc net.ifnames=0 biosdevname=0 mitigations=off transparent_hugepage=always numa_balancing=disable pcie_aspm=off intel_idle.max_cstate=1 processor.max_cstate=1 nowatchdog nosoftlockup
     initrd  ${initrd}
 }
 
-
 if [ "\${grub_platform}" = "efi" ]; then
     menuentry "Memory Test (memtest86+)" {
         chainloader /boot/memtest86+x64.efi
@@ -769,23 +775,24 @@ write_canonical_isolinux_cfg() {
     kernel="$2"
     initrd="$3"
     append_live="$4"
+    version_label="${PROJECT_VERSION_EFFECTIVE}"
 
     cat > "$cfg" <<EOF
 label live-@FLAVOUR@-normal
-    menu label ^EASY-BEE
+    menu label ^EASY-BEE v${version_label}
     linux ${kernel}
     initrd ${initrd}
     append ${append_live} nomodeset bee.nvidia.mode=normal net.ifnames=0 biosdevname=0 mitigations=off transparent_hugepage=always numa_balancing=disable pcie_aspm=off intel_idle.max_cstate=1 processor.max_cstate=1 nowatchdog nosoftlockup
 
 label live-@FLAVOUR@-toram
-    menu label EASY-BEE (^load to RAM)
+    menu label EASY-BEE v${version_label} (^load to RAM)
     menu default
     linux ${kernel}
     initrd ${initrd}
     append ${append_live} toram nomodeset bee.nvidia.mode=normal net.ifnames=0 biosdevname=0 mitigations=off transparent_hugepage=always numa_balancing=disable pcie_aspm=off intel_idle.max_cstate=1 processor.max_cstate=1 nowatchdog nosoftlockup
 
 label live-@FLAVOUR@-console
-    menu label EASY-BEE (^no GUI / no X11)
+    menu label EASY-BEE v${version_label} (^no GUI / no X11)
     linux ${kernel}
     initrd ${initrd}
     append ${append_live} nomodeset bee.gui=off bee.nvidia.mode=gsp-off net.ifnames=0 biosdevname=0 mitigations=off transparent_hugepage=always numa_balancing=disable pcie_aspm=off intel_idle.max_cstate=1 processor.max_cstate=1 nowatchdog nosoftlockup
@@ -833,10 +840,7 @@ enforce_live_build_bootloader_assets() {
 
     if [ -f "$grub_cfg" ]; then
         if extract_live_grub_entry "$grub_cfg"; then
-            mkdir -p "$grub_dir/live-theme"
             cp "${BUILDER_DIR}/config/bootloaders/grub-efi/config.cfg" "$grub_dir/config.cfg"
-            cp "${BUILDER_DIR}/config/bootloaders/grub-efi/theme.cfg" "$grub_dir/theme.cfg"
-            cp -R "${BUILDER_DIR}/config/bootloaders/grub-efi/live-theme/." "$grub_dir/live-theme/"
             write_canonical_grub_cfg "$grub_cfg" "$grub_kernel" "${live_build_append:-$grub_append}" "$grub_initrd"
             echo "bootloader sync: rewrote binary/boot/grub/grub.cfg with canonical EASY-BEE menu"
         else
@@ -852,60 +856,6 @@ enforce_live_build_bootloader_assets() {
             echo "bootloader sync: WARNING: could not extract live entry from $isolinux_cfg" >&2
         fi
     fi
-
-    sync_efi_grub_theme_assets "$lb_dir"
-}
-
-fat_image_has_file() {
-    img="$1"
-    path="$2"
-    mtype -i "$img" "$path" >/dev/null 2>&1
-}
-
-is_efi_grub_fat_image() {
-    img="$1"
-    fat_image_has_file "$img" "::/EFI/boot/grubx64.efi" || return 1
-    fat_image_has_file "$img" "::/boot/grub/grub.cfg" || return 1
-    return 0
-}
-
-copy_file_to_fat_image() {
-    img="$1"
-    src="$2"
-    dst="$3"
-    mcopy -o -i "$img" "$src" "$dst" >/dev/null
-}
-
-sync_efi_grub_theme_assets() {
-    lb_dir="$1"
-    found=0
-
-    for img in $(find "$lb_dir" -type f \( -name '*.img' -o -name '*.ima' -o -name '*.fat' \) 2>/dev/null); do
-        if ! is_efi_grub_fat_image "$img"; then
-            continue
-        fi
-        found=1
-        echo "bootloader sync: patching EFI GRUB image $img"
-        mmd -i "$img" "::/boot" >/dev/null 2>&1 || true
-        mmd -i "$img" "::/boot/grub" >/dev/null 2>&1 || true
-
-        copy_file_to_fat_image "$img" "${BUILDER_DIR}/config/bootloaders/grub-efi/config.cfg" "::/boot/grub/config.cfg"
-        copy_file_to_fat_image "$img" "${BUILDER_DIR}/config/bootloaders/grub-efi/theme.cfg" "::/boot/grub/theme.cfg"
-
-        fat_image_has_file "$img" "::/boot/grub/config.cfg" || {
-            echo "ERROR: EFI GRUB image missing /boot/grub/config.cfg after sync: $img" >&2
-            exit 1
-        }
-        fat_image_has_file "$img" "::/boot/grub/theme.cfg" || {
-            echo "ERROR: EFI GRUB image missing /boot/grub/theme.cfg after sync: $img" >&2
-            exit 1
-        }
-    done
-
-    if [ "$found" != "1" ]; then
-        echo "ERROR: no EFI GRUB FAT image found in live-build workdir; cannot sync theme assets" >&2
-        exit 1
-    fi
 }
 
 copy_memtest_from_deb() {
@@ -944,8 +894,11 @@ FULL_BUILD_MARKER="${BUILD_WORK_DIR}/.bee-full-build-marker"
 # hooks, archives, Dockerfile, auto/config) require a full lb build.
 needs_full_build() {
     [ -f "${FULL_BUILD_MARKER}" ]                                        || return 0
-    [ -f "${BUILD_WORK_DIR}/binary/live/filesystem.squashfs" ]           || return 0
     [ -f "${BUILD_WORK_DIR}/live-image-amd64.hybrid.iso" ]               || return 0
+    # Accept any versioned squashfs (filesystem-v*.squashfs or legacy filesystem.squashfs)
+    _any_sq=$(find "${BUILD_WORK_DIR}/binary/live" -maxdepth 1 \
+        -name 'filesystem*.squashfs' 2>/dev/null | head -1)
+    [ -n "$_any_sq" ]                                                    || return 0
 
     _heavy=$(find \
         "${BUILDER_DIR}/VERSIONS" \
@@ -968,40 +921,109 @@ needs_full_build() {
 # Fast-path: unsquash existing filesystem, rsync overlay on top, repack.
 # Requires ~10 GB free in BEE_CACHE_DIR for the unpacked squashfs.
 fast_path_repack_squashfs() {
-    _sq="${BUILD_WORK_DIR}/binary/live/filesystem.squashfs"
+    _old_sq=$(find "${BUILD_WORK_DIR}/binary/live" -maxdepth 1 \
+        -name 'filesystem*.squashfs' | sort | head -1)
+    _sq="${BUILD_WORK_DIR}/binary/live/${SQUASHFS_FILENAME}"
     _tmp="${BEE_CACHE_DIR}/fast-unsquash-${BUILD_VARIANT}"
-    echo "=== fast-path: unsquash ($(du -sh "$_sq" | cut -f1) compressed) ==="
+    echo "=== fast-path: unsquash $(basename "$_old_sq") ($(du -sh "$_old_sq" | cut -f1) compressed) ==="
     rm -rf "$_tmp"
-    unsquashfs -d "$_tmp" "$_sq"
+    unsquashfs -d "$_tmp" "$_old_sq"
     echo "=== fast-path: syncing overlay stage ==="
     rsync -a --checksum "${OVERLAY_STAGE_DIR}/" "$_tmp/"
-    echo "=== fast-path: repacking squashfs ==="
+    echo "=== fast-path: repacking as ${SQUASHFS_FILENAME} ==="
     _sq_new="${_sq}.new"
     rm -f "$_sq_new"
-    mksquashfs "$_tmp" "$_sq_new" -comp zstd -b 1048576 -noappend -no-progress
+    mksquashfs "$_tmp" "$_sq_new" -comp zstd -b 1048576 -noappend -no-progress -no-xattrs
     mv "$_sq_new" "$_sq"
     rm -rf "$_tmp"
+    [ "$_old_sq" != "$_sq" ] && rm -f "$_old_sq"
     echo "=== fast-path: squashfs repacked ($(du -sh "$_sq" | cut -f1)) ==="
 }
 
-# Fast-path: rebuild ISO by replacing only live/filesystem.squashfs via xorriso.
+# Fast-path: rebuild ISO replacing the squashfs via xorriso.
 # Boot structure (El Torito, EFI, MBR hybrid) is replayed from the prior ISO.
 fast_path_rebuild_iso() {
-    _sq="${BUILD_WORK_DIR}/binary/live/filesystem.squashfs"
+    _sq="${BUILD_WORK_DIR}/binary/live/${SQUASHFS_FILENAME}"
     _prior="${BUILD_WORK_DIR}/live-image-amd64.hybrid.iso"
     _new="${BUILD_WORK_DIR}/live-image-amd64.hybrid.iso.new"
     echo "=== fast-path: rebuilding ISO with xorriso ==="
     rm -f "$_new"
+    # Remove any old squashfs entries from the prior ISO before adding the new one
+    _old_entries=$(xorriso -indev "$_prior" -find /live -name 'filesystem*.squashfs' -- 2>/dev/null \
+        | grep -E '^/live/filesystem.*\.squashfs$' || true)
+    _rm_args=""
+    for _e in $_old_entries; do
+        _rm_args="$_rm_args -rm $_e --"
+    done
+    # shellcheck disable=SC2086
     xorriso \
         -indev  "$_prior" \
         -outdev "$_new" \
-        -map    "$_sq" /live/filesystem.squashfs \
+        ${_rm_args} \
+        -map    "$_sq" /live/${SQUASHFS_FILENAME} \
         -boot_image any replay \
         -commit
     mv "$_new" "$_prior"
     echo "=== fast-path: ISO rebuilt ==="
 }
 
+dir_has_entries() {
+    _dir="$1"
+    [ -d "$_dir" ] || return 1
+    find "$_dir" -mindepth 1 -print -quit 2>/dev/null | grep -q .
+}
+
+move_tree_to_layer() {
+    _src_root="$1"
+    _rel="$2"
+    _dst_root="$3"
+    [ -e "${_src_root}/${_rel}" ] || return 0
+    mkdir -p "${_dst_root}/$(dirname "$_rel")"
+    mv "${_src_root}/${_rel}" "${_dst_root}/${_rel}"
+}
+
+split_live_squashfs_layers() {
+    lb_dir="$1"
+    live_dir="${lb_dir}/binary/live"
+    base_sq="${live_dir}/filesystem.squashfs"
+    usr_sq="${live_dir}/10-usr.squashfs"
+    fw_sq="${live_dir}/20-firmware.squashfs"
+
+    [ -f "$base_sq" ] || return 0
+    command -v unsquashfs >/dev/null 2>&1 || return 0
+    command -v mksquashfs >/dev/null 2>&1 || return 0
+
+    tmp_root="$(mktemp -d)"
+    tmp_usr="$(mktemp -d)"
+    tmp_fw="$(mktemp -d)"
+
+    echo "=== splitting live squashfs into smaller layers ==="
+    unsquashfs -d "$tmp_root/root" "$base_sq" >/dev/null
+    mkdir -p "$tmp_usr/root" "$tmp_fw/root"
+
+    move_tree_to_layer "$tmp_root/root" "usr" "$tmp_usr/root"
+    move_tree_to_layer "$tmp_root/root" "lib/firmware" "$tmp_fw/root"
+    move_tree_to_layer "$tmp_root/root" "usr/lib/firmware" "$tmp_fw/root"
+    move_tree_to_layer "$tmp_root/root" "boot/firmware" "$tmp_fw/root"
+
+    rm -f "$usr_sq" "$fw_sq"
+    mksquashfs "$tmp_root/root" "${base_sq}.new" -comp zstd -b 1048576 -noappend -no-progress -no-xattrs >/dev/null
+    mv "${base_sq}.new" "$base_sq"
+
+    if dir_has_entries "$tmp_usr/root"; then
+        mksquashfs "$tmp_usr/root" "${usr_sq}.new" -comp zstd -b 1048576 -noappend -no-progress -no-xattrs >/dev/null
+        mv "${usr_sq}.new" "$usr_sq"
+    fi
+    if dir_has_entries "$tmp_fw/root"; then
+        mksquashfs "$tmp_fw/root" "${fw_sq}.new" -comp zstd -b 1048576 -noappend -no-progress -no-xattrs >/dev/null
+        mv "${fw_sq}.new" "$fw_sq"
+    fi
+
+    echo "=== live squashfs layers ==="
+    find "$live_dir" -maxdepth 1 -type f -name '*.squashfs' -exec du -sh {} \; | sort
+    rm -rf "$tmp_root" "$tmp_usr" "$tmp_fw"
+}
+
 recover_iso_memtest() {
     lb_dir="$1"
     iso_path="$2"
@@ -1080,9 +1102,10 @@ recover_iso_memtest() {
 }
 
 PROJECT_VERSION_EFFECTIVE="$(resolve_project_version)"
+SQUASHFS_FILENAME="filesystem-v${PROJECT_VERSION_EFFECTIVE}.squashfs"
 ISO_BASENAME="easy-bee-${BUILD_VARIANT}-v${PROJECT_VERSION_EFFECTIVE}-amd64"
 # Versioned output directory: dist/easy-bee-v4.1/ — all final artefacts live here.
-OUT_DIR="${DIST_DIR}/easy-bee-v${PROJECT_VERSION_EFFECTIVE}"
+OUT_DIR="${DIST_DIR}/release/easy-bee-v${PROJECT_VERSION_EFFECTIVE}"
 ISO_VERSION_LABEL_TOKEN="$(printf '%s' "${PROJECT_VERSION_EFFECTIVE}" | tr '[:lower:].-' '[:upper:]__')"
 mkdir -p "${OUT_DIR}"
 LOG_DIR="${OUT_DIR}/${ISO_BASENAME}.logs"
@@ -1267,7 +1290,7 @@ run_step "sync git submodules" "05-git-submodules" \
 
 # --- compile bee binary (static, Linux amd64) ---
 # Shared between variants — built once, reused on second pass.
-BEE_BIN="${DIST_DIR}/bee-linux-amd64"
+BEE_BIN="${DIST_DIR}/cache/bee-linux-amd64"
 NEED_BUILD=1
 if [ -f "$BEE_BIN" ]; then
     NEWEST_SRC=$(find "${REPO_ROOT}/audit" -name '*.go' -newer "$BEE_BIN" | head -1)
@@ -1298,16 +1321,16 @@ else
 fi
 
 # --- NVIDIA-only build steps ---
-GPU_BURN_WORKER_BIN="${DIST_DIR}/bee-gpu-burn-worker-linux-amd64"
+GPU_BURN_WORKER_BIN="${DIST_DIR}/cache/bee-gpu-burn-worker-linux-amd64"
 if [ "$BEE_GPU_VENDOR" = "nvidia" ]; then
     run_step "download cuBLAS/cuBLASLt/cudart ${NCCL_CUDA_VERSION} userspace" "20-cublas" \
         sh "${BUILDER_DIR}/build-cublas.sh" \
         "${CUBLAS_VERSION}" \
         "${CUDA_USERSPACE_VERSION}" \
         "${NCCL_CUDA_VERSION}" \
-        "${DIST_DIR}"
+        "${DIST_DIR}/cache"
 
-    CUBLAS_CACHE="${DIST_DIR}/cublas-${CUBLAS_VERSION}+cuda${NCCL_CUDA_VERSION}"
+    CUBLAS_CACHE="${DIST_DIR}/cache/cublas-${CUBLAS_VERSION}+cuda${NCCL_CUDA_VERSION}"
 
     echo "=== bee-gpu-burn FP4 header probe ==="
     fp4_type_match="$(grep -Rsnm 1 'CUDA_R_4F_E2M1' "${CUBLAS_CACHE}/include" 2>/dev/null || true)"
@@ -1433,7 +1456,7 @@ fi
 
 # --- copy bee binary into overlay ---
 mkdir -p "${OVERLAY_STAGE_DIR}/usr/local/bin"
-cp "${DIST_DIR}/bee-linux-amd64" "${OVERLAY_STAGE_DIR}/usr/local/bin/bee"
+cp "$BEE_BIN" "${OVERLAY_STAGE_DIR}/usr/local/bin/bee"
 chmod +x "${OVERLAY_STAGE_DIR}/usr/local/bin/bee"
 
 if [ "$BEE_GPU_VENDOR" = "nvidia" ] && [ -f "$GPU_BURN_WORKER_BIN" ]; then
@@ -1463,10 +1486,10 @@ done
 # --- NVIDIA kernel modules and userspace libs ---
 if [ "$BEE_GPU_VENDOR" = "nvidia" ]; then
     run_step "build NVIDIA ${NVIDIA_DRIVER_VERSION} modules" "40-nvidia-module" \
-        sh "${BUILDER_DIR}/build-nvidia-module.sh" "${NVIDIA_DRIVER_VERSION}" "${DIST_DIR}" "${DEBIAN_KERNEL_ABI}" "${BEE_NVIDIA_MODULE_FLAVOR}"
+        sh "${BUILDER_DIR}/build-nvidia-module.sh" "${NVIDIA_DRIVER_VERSION}" "${DIST_DIR}/cache" "${DEBIAN_KERNEL_ABI}" "${BEE_NVIDIA_MODULE_FLAVOR}"
 
     KVER="${DEBIAN_KERNEL_ABI}-amd64"
-    NVIDIA_CACHE="${DIST_DIR}/nvidia-${BEE_NVIDIA_MODULE_FLAVOR}-${NVIDIA_DRIVER_VERSION}-${KVER}"
+    NVIDIA_CACHE="${DIST_DIR}/cache/nvidia-${BEE_NVIDIA_MODULE_FLAVOR}-${NVIDIA_DRIVER_VERSION}-${KVER}"
 
     # Inject .ko files into overlay at /usr/local/lib/nvidia/
     OVERLAY_KMOD_DIR="${OVERLAY_STAGE_DIR}/usr/local/lib/nvidia"
@@ -1492,9 +1515,9 @@ if [ "$BEE_GPU_VENDOR" = "nvidia" ]; then
 
     # --- build / download NCCL ---
     run_step "download NCCL ${NCCL_VERSION}+cuda${NCCL_CUDA_VERSION}" "50-nccl" \
-        sh "${BUILDER_DIR}/build-nccl.sh" "${NCCL_VERSION}" "${NCCL_CUDA_VERSION}" "${DIST_DIR}" "${NCCL_SHA256:-}"
+        sh "${BUILDER_DIR}/build-nccl.sh" "${NCCL_VERSION}" "${NCCL_CUDA_VERSION}" "${DIST_DIR}/cache" "${NCCL_SHA256:-}"
 
-    NCCL_CACHE="${DIST_DIR}/nccl-${NCCL_VERSION}+cuda${NCCL_CUDA_VERSION}"
+    NCCL_CACHE="${DIST_DIR}/cache/nccl-${NCCL_VERSION}+cuda${NCCL_CUDA_VERSION}"
 
     # Inject libnccl.so.* into overlay alongside other NVIDIA userspace libs
     cp "${NCCL_CACHE}/lib/"* "${OVERLAY_STAGE_DIR}/usr/lib/"
@@ -1510,19 +1533,19 @@ if [ "$BEE_GPU_VENDOR" = "nvidia" ]; then
         "${NCCL_TESTS_VERSION}" \
         "${NCCL_VERSION}" \
         "${NCCL_CUDA_VERSION}" \
-        "${DIST_DIR}" \
+        "${DIST_DIR}/cache" \
         "${NVCC_VERSION}" \
         "${DEBIAN_VERSION}"
 
-    NCCL_TESTS_CACHE="${DIST_DIR}/nccl-tests-${NCCL_TESTS_VERSION}"
+    NCCL_TESTS_CACHE="${DIST_DIR}/cache/nccl-tests-${NCCL_TESTS_VERSION}"
     cp "${NCCL_TESTS_CACHE}/bin/all_reduce_perf" "${OVERLAY_STAGE_DIR}/usr/local/bin/all_reduce_perf"
     chmod +x "${OVERLAY_STAGE_DIR}/usr/local/bin/all_reduce_perf"
     cp "${NCCL_TESTS_CACHE}/lib/"* "${OVERLAY_STAGE_DIR}/usr/lib/" 2>/dev/null || true
     echo "=== all_reduce_perf injected ==="
 
     run_step "build john jumbo ${JOHN_JUMBO_COMMIT}" "70-john" \
-        sh "${BUILDER_DIR}/build-john.sh" "${JOHN_JUMBO_COMMIT}" "${DIST_DIR}"
+        sh "${BUILDER_DIR}/build-john.sh" "${JOHN_JUMBO_COMMIT}" "${DIST_DIR}/cache"
-    JOHN_CACHE="${DIST_DIR}/john-${JOHN_JUMBO_COMMIT}"
+    JOHN_CACHE="${DIST_DIR}/cache/john-${JOHN_JUMBO_COMMIT}"
     mkdir -p "${OVERLAY_STAGE_DIR}/usr/local/lib/bee/john"
     rsync -a --delete "${JOHN_CACHE}/run/" "${OVERLAY_STAGE_DIR}/usr/local/lib/bee/john/run/"
     ln -sfn ../lib/bee/john/run/john "${OVERLAY_STAGE_DIR}/usr/local/bin/john"
@@ -1650,7 +1673,7 @@ if ! needs_full_build; then
     fast_path_rebuild_iso
     ISO_RAW="${LB_DIR}/live-image-amd64.hybrid.iso"
     validate_iso_live_boot_entries "$ISO_RAW"
-    validate_iso_grub_theme_assets "$ISO_RAW"
+    validate_iso_grub_assets "$ISO_RAW"
     validate_iso_nvidia_runtime "$ISO_RAW"
     cp "$ISO_RAW" "$ISO_OUT"
     echo ""
@@ -1665,16 +1688,30 @@ echo "=== building ISO (variant: ${BUILD_VARIANT}) ==="
 
 # Export for auto/config
 BEE_GPU_VENDOR_UPPER="$(echo "${BUILD_VARIANT}" | tr 'a-z-' 'A-Z_')"
-BEE_ISO_VOLUME="EASY_BEE_${BEE_GPU_VENDOR_UPPER}_V${ISO_VERSION_LABEL_TOKEN}"
+# ISO 9660 volume ID is limited to 32 characters; truncate the version token to fit.
+_vol_prefix="EASY_BEE_${BEE_GPU_VENDOR_UPPER}_V"
+_max_token=$(( 32 - ${#_vol_prefix} ))
+_vol_token="$(printf '%s' "${ISO_VERSION_LABEL_TOKEN}" | cut -c1-${_max_token})"
+BEE_ISO_VOLUME="${_vol_prefix}${_vol_token}"
+unset _vol_prefix _max_token _vol_token
 export BEE_GPU_VENDOR_UPPER BEE_ISO_VOLUME
 
 cd "${LB_DIR}"
 run_step_sh "live-build clean" "80-lb-clean" "lb clean --all 2>&1 | tail -3"
 run_step_sh "live-build config" "81-lb-config" "lb config 2>&1 | tail -5"
 dump_memtest_debug "pre-build" "${LB_DIR}"
+export MKSQUASHFS_OPTIONS="-no-xattrs"
 run_step_sh "live-build build" "90-lb-build" "lb build 2>&1"
 echo "=== enforcing canonical bootloader assets ==="
 enforce_live_build_bootloader_assets "${LB_DIR}"
+# Rename lb's default filesystem.squashfs to the versioned filename so the
+# ISO contains a version-stamped squashfs (e.g. filesystem-v10.15.squashfs).
+_std_sq="${LB_DIR}/binary/live/filesystem.squashfs"
+_ver_sq="${LB_DIR}/binary/live/${SQUASHFS_FILENAME}"
+if [ -f "${_std_sq}" ] && [ "${_std_sq}" != "${_ver_sq}" ]; then
+    mv "${_std_sq}" "${_ver_sq}"
+    echo "=== squashfs renamed: filesystem.squashfs → ${SQUASHFS_FILENAME} ==="
+fi
 reset_live_build_stage "${LB_DIR}" "binary_checksums"
 reset_live_build_stage "${LB_DIR}" "binary_iso"
 reset_live_build_stage "${LB_DIR}" "binary_zsync"
@@ -1706,7 +1743,7 @@ if [ -f "$ISO_RAW" ]; then
     fi
     validate_iso_memtest "$ISO_RAW"
     validate_iso_live_boot_entries "$ISO_RAW"
-    validate_iso_grub_theme_assets "$ISO_RAW"
+    validate_iso_grub_assets "$ISO_RAW"
     validate_iso_nvidia_runtime "$ISO_RAW"
     cp "$ISO_RAW" "$ISO_OUT"
     touch "${FULL_BUILD_MARKER}"

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

@@ -1,5 +1,7 @@
 set default=1
 set timeout=10
+set color_normal=yellow/black
+set color_highlight=white/brown
 
 if [ x$feature_default_font_path = xy ] ; then
     font=unicode
@@ -26,6 +28,3 @@ insmod gfxterm
 
 terminal_input console serial
 terminal_output gfxterm serial
-
-insmod tga
-source /boot/grub/theme.cfg

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

@@ -1,21 +1,26 @@
 source /boot/grub/config.cfg
 
-menuentry "EASY-BEE" {
+menuentry "EASY-BEE v@VERSION@" {
     linux   @KERNEL_LIVE@ @APPEND_LIVE@ nomodeset bee.nvidia.mode=normal pci=realloc net.ifnames=0 biosdevname=0 mitigations=off transparent_hugepage=always numa_balancing=disable pcie_aspm=off intel_idle.max_cstate=1 processor.max_cstate=1 nowatchdog nosoftlockup
     initrd  @INITRD_LIVE@
 }
 
-menuentry "EASY-BEE -- load to RAM (toram)" {
+menuentry "EASY-BEE v@VERSION@ -- load to RAM (toram)" {
     linux   @KERNEL_LIVE@ @APPEND_LIVE@ toram nomodeset bee.nvidia.mode=normal pci=realloc net.ifnames=0 biosdevname=0 mitigations=off transparent_hugepage=always numa_balancing=disable pcie_aspm=off intel_idle.max_cstate=1 processor.max_cstate=1 nowatchdog nosoftlockup
     initrd  @INITRD_LIVE@
 }
 
-menuentry "EASY-BEE -- no GUI / no X11" {
+menuentry "EASY-BEE v@VERSION@ -- no GUI / no X11" {
     linux   @KERNEL_LIVE@ @APPEND_LIVE@ nomodeset bee.gui=off bee.nvidia.mode=gsp-off pci=realloc net.ifnames=0 biosdevname=0 mitigations=off transparent_hugepage=always numa_balancing=disable pcie_aspm=off intel_idle.max_cstate=1 processor.max_cstate=1 nowatchdog nosoftlockup
     initrd  @INITRD_LIVE@
 }
 
 
+menuentry "*** WIPE ALL DISKS (irreversible!) ***" {
+    linux   @KERNEL_LIVE@ @APPEND_LIVE@ toram nomodeset bee.gui=off bee.wipe=all net.ifnames=0 biosdevname=0
+    initrd  @INITRD_LIVE@
+}
+
 if [ "${grub_platform}" = "efi" ]; then
     menuentry "Memory Test (memtest86+)" {
         chainloader /boot/memtest86+x64.efi

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

@@ -1,18 +1,18 @@
 label live-@FLAVOUR@-normal
-    menu label ^EASY-BEE
+    menu label ^EASY-BEE v@VERSION@
     linux @LINUX@
     initrd @INITRD@
     append @APPEND_LIVE@ nomodeset bee.nvidia.mode=normal net.ifnames=0 biosdevname=0 mitigations=off transparent_hugepage=always numa_balancing=disable pcie_aspm=off intel_idle.max_cstate=1 processor.max_cstate=1 nowatchdog nosoftlockup
 
 label live-@FLAVOUR@-toram
-    menu label EASY-BEE (^load to RAM)
+    menu label EASY-BEE v@VERSION@ (^load to RAM)
     menu default
     linux @LINUX@
     initrd @INITRD@
     append @APPEND_LIVE@ toram nomodeset bee.nvidia.mode=normal net.ifnames=0 biosdevname=0 mitigations=off transparent_hugepage=always numa_balancing=disable pcie_aspm=off intel_idle.max_cstate=1 processor.max_cstate=1 nowatchdog nosoftlockup
 
 label live-@FLAVOUR@-console
-    menu label EASY-BEE (^no GUI / no X11)
+    menu label EASY-BEE v@VERSION@ (^no GUI / no X11)
     linux @LINUX@
     initrd @INITRD@
     append @APPEND_LIVE@ nomodeset bee.gui=off bee.nvidia.mode=gsp-off net.ifnames=0 biosdevname=0 mitigations=off transparent_hugepage=always numa_balancing=disable pcie_aspm=off intel_idle.max_cstate=1 processor.max_cstate=1 nowatchdog nosoftlockup
@@ -41,6 +41,12 @@ label live-@FLAVOUR@-failsafe
     initrd @INITRD@
     append @APPEND_LIVE@ nomodeset bee.nvidia.mode=gsp-off noapic noapm nodma nomce nolapic nosmp vga=normal net.ifnames=0 biosdevname=0
 
+label wipe-disks
+    menu label *** WIPE ALL DISKS (irreversible!) ***
+    linux @LINUX@
+    initrd @INITRD@
+    append @APPEND_LIVE@ toram nomodeset bee.gui=off bee.wipe=all net.ifnames=0 biosdevname=0
+
 label memtest
     menu label ^Memory Test (memtest86+)
     linux /boot/memtest86+x64.bin

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

@@ -69,6 +69,7 @@ chmod +x /usr/local/bin/bee-boot-status    2>/dev/null || true
 chmod +x /usr/local/bin/bee-install        2>/dev/null || true
 chmod +x /usr/local/bin/bee-gui-gate       2>/dev/null || true
 chmod +x /usr/local/bin/bee-remount-medium 2>/dev/null || true
+chmod +x /usr/local/bin/bee-check-nvswitch 2>/dev/null || true
 if [ "$GPU_VENDOR" = "nvidia" ]; then
     chmod +x /usr/local/bin/bee-nvidia-load 2>/dev/null || true
     chmod +x /usr/local/bin/bee-gpu-burn 2>/dev/null || true

iso/builder/config/hooks/normal/9012-wipe.hook.chroot

@@ -0,0 +1,57 @@
+#!/bin/sh
+# 9012-wipe.hook.chroot
+#
+# Adds bee-initramfs-wipe to the initramfs so that selecting the
+# "WIPE ALL DISKS" boot menu entry runs the wipe tool before squashfs
+# is mounted — i.e. it works even when live boot fails.
+#
+# Two files are installed inside the chroot:
+#   /etc/initramfs-tools/hooks/bee-wipe         — copies binaries into initrd
+#   /etc/initramfs-tools/scripts/local-premount/bee-wipe — runs at boot
+
+set -e
+
+HOOK_DIR="/etc/initramfs-tools/hooks"
+SCRIPT_DIR="/etc/initramfs-tools/scripts/local-premount"
+
+mkdir -p "${HOOK_DIR}" "${SCRIPT_DIR}"
+
+# ── initramfs hook: copy binaries ────────────────────────────────────────────
+cat > "${HOOK_DIR}/bee-wipe" << 'EOF'
+#!/bin/sh
+PREREQ=""
+prereqs() { echo "$PREREQ"; }
+case "$1" in prereqs) prereqs; exit 0 ;; esac
+
+. /usr/share/initramfs-tools/hook-functions
+
+for bin in lsblk blkid blkdiscard blockdev; do
+    b=$(command -v "$bin" 2>/dev/null) && copy_exec "$b" /bin
+done
+
+[ -x /usr/sbin/nvme ] && copy_exec /usr/sbin/nvme /sbin
+
+copy_exec /usr/local/bin/bee-initramfs-wipe /bin/bee-wipe
+EOF
+
+chmod +x "${HOOK_DIR}/bee-wipe"
+
+# ── initramfs premount script: trigger on bee.wipe=all ───────────────────────
+cat > "${SCRIPT_DIR}/bee-wipe" << 'EOF'
+#!/bin/sh
+PREREQ=""
+prereqs() { echo "$PREREQ"; }
+case "$1" in prereqs) prereqs; exit 0 ;; esac
+
+grep -qw 'bee.wipe=all' /proc/cmdline 2>/dev/null || exit 0
+exec /bin/bee-wipe
+EOF
+
+chmod +x "${SCRIPT_DIR}/bee-wipe"
+
+echo "9012-wipe: installed initramfs hook and premount script"
+
+KVER=$(ls /lib/modules | sort -V | tail -1)
+echo "9012-wipe: rebuilding initramfs for kernel ${KVER}"
+update-initramfs -u -k "${KVER}"
+echo "9012-wipe: done"

iso/builder/config/hooks/normal/9998-strip-xattrs.hook.chroot

@@ -0,0 +1,37 @@
+#!/usr/bin/env python3
+# 9998-strip-xattrs.hook.chroot
+#
+# mksquashfs 4.5.1 (Debian bookworm) writes a non-INVALID xattr_id_table_start
+# even with -no-xattrs when the source tree contains POSIX ACL xattrs set by
+# dpkg/install-time.  Linux 6.1 squashfs driver then fails with
+# "unable to read xattr id index table" and aborts the mount.
+#
+# Strip all xattrs from the live chroot before mksquashfs sees the tree so the
+# resulting squashfs has SQUASHFS_INVALID_BLK in xattr_id_table_start.
+
+import os
+
+def strip(path):
+    try:
+        for attr in os.listxattr(path, follow_symlinks=False):
+            try:
+                os.removexattr(path, attr, follow_symlinks=False)
+            except OSError:
+                pass
+    except OSError:
+        pass
+
+removed = 0
+for root, dirs, files in os.walk('/', topdown=True, followlinks=False):
+    for name in dirs + files:
+        p = os.path.join(root, name)
+        try:
+            attrs = os.listxattr(p, follow_symlinks=False)
+            if attrs:
+                strip(p)
+                removed += len(attrs)
+        except OSError:
+            pass
+    strip(root)
+
+print(f"9998-strip-xattrs: removed xattrs from {removed} entries")

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

@@ -1,5 +1,6 @@
 # AMD GPU firmware
 firmware-amd-graphics
+nvtop
 
 # AMD ROCm — GPU monitoring, bandwidth test, and compute stress (RVS GST)
 rocm-smi-lib=%%ROCM_SMI_VERSION%%

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

@@ -5,6 +5,7 @@
 # DCGM 4 is packaged per CUDA major. The image ships NVIDIA driver 590 with
 # CUDA 13 userspace, so install the CUDA 13 build plus proprietary components
 # explicitly.
+nvtop
 nvidia-fabricmanager=%%NVIDIA_FABRICMANAGER_VERSION%%
 datacenter-gpu-manager-4-cuda13=1:%%DCGM_VERSION%%
 datacenter-gpu-manager-4-proprietary=1:%%DCGM_VERSION%%

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

@@ -38,6 +38,7 @@ exfat-fuse
 ntfs-3g
 
 # Utilities
+infiniband-diags
 bash
 procps
 lsof
@@ -46,7 +47,6 @@ less
 vim-tiny
 mc
 htop
-nvtop
 sudo
 zstd
 mstflint

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

@@ -2,6 +2,8 @@
 Description=Bee: load NVIDIA kernel modules and create device nodes
 After=local-fs.target udev.service bee-blackbox.service
 Before=bee-audit.service
+# Skip silently if bee-nvidia-load is absent (non-nvidia builds).
+ConditionPathExists=/usr/local/bin/bee-nvidia-load
 
 [Service]
 Type=oneshot

iso/overlay/etc/systemd/system/nvidia-fabricmanager.service.d/bee-nvswitch-check.conf

@@ -0,0 +1,4 @@
+[Service]
+# Skip fabricmanager on systems without NVSwitch hardware.
+# ExecCondition exits 1-254 → unit is silently skipped (inactive, not failed).
+ExecCondition=/usr/local/bin/bee-check-nvswitch

iso/overlay/usr/local/bin/bee-boot-status

@@ -3,8 +3,14 @@
 # Shows live service status until all bee services are done or failed,
 # then exits so getty can show the login prompt.
 
-CRITICAL="bee-preflight bee-nvidia bee-audit"
+GPU_VENDOR="$(cat /etc/bee-gpu-vendor 2>/dev/null || echo nvidia)"
-ALL="bee-sshsetup ssh bee-network bee-nvidia bee-preflight bee-audit bee-web"
+if [ "$GPU_VENDOR" = "nvidia" ]; then
+    CRITICAL="bee-preflight bee-nvidia bee-audit"
+    ALL="bee-sshsetup ssh bee-network bee-nvidia bee-preflight bee-audit bee-web"
+else
+    CRITICAL="bee-preflight bee-audit"
+    ALL="bee-sshsetup ssh bee-network bee-preflight bee-audit bee-web"
+fi
 
 svc_state() { systemctl is-active "$1.service" 2>/dev/null || echo "inactive"; }
 

iso/overlay/usr/local/bin/bee-check-nvswitch

@@ -0,0 +1,4 @@
+#!/bin/sh
+# Exit 0 if NVSwitch hardware is detected; exit 1 to skip fabricmanager on non-NVSwitch systems.
+# NVSwitch appears in lspci as vendor 10de, class 0680 (Bridge, Other).
+lspci -Dn 2>/dev/null | awk '$2 == "0680:" && $3 ~ /^10de:/ { found=1; exit } END { exit(found ? 0 : 1) }'

iso/overlay/usr/local/bin/bee-initramfs-wipe

@@ -0,0 +1,166 @@
+#!/bin/sh
+# bee-initramfs-wipe — interactive disk wipe running entirely in the initramfs.
+# Triggered by bee.wipe=all on the kernel cmdline (via local-premount hook).
+# Works before squashfs is mounted, so it runs even when live boot fails.
+
+RED='\033[1;31m'
+YEL='\033[1;33m'
+GRN='\033[1;32m'
+CYN='\033[1;36m'
+NC='\033[0m'
+
+p()  { printf '%b\n' "$*"; }
+pp() { printf '%b'   "$*"; }
+
+banner() {
+    p ""
+    p "${RED}╔══════════════════════════════════════════════════════════╗${NC}"
+    p "${RED}║            BEE DRIVE WIPE — initramfs stage              ║${NC}"
+    p "${RED}╚══════════════════════════════════════════════════════════╝${NC}"
+    p ""
+}
+
+# ── find boot device ─────────────────────────────────────────────────────────
+boot_dev() {
+    local label token
+    for token in $(cat /proc/cmdline 2>/dev/null); do
+        case "$token" in
+            live-media-label=*) label="${token#*=}" ;;
+        esac
+    done
+    [ -z "$label" ] && return
+
+    local dev
+    dev=$(blkid -L "$label" 2>/dev/null) || return
+    # strip partition suffix: /dev/sdb1 → /dev/sdb, /dev/nvme0n1p1 → /dev/nvme0n1
+    echo "$dev" | sed 's/p\?[0-9]\+$//'
+}
+
+# ── enumerate candidate disks ─────────────────────────────────────────────────
+list_disks() {
+    local boot
+    boot=$(boot_dev)
+
+    lsblk -d -n -o NAME,TYPE,SIZE,MODEL 2>/dev/null | while read -r name type size model; do
+        [ "$type" = "disk" ] || continue
+        [ "$size" = "0B"  ] && continue
+        local dev="/dev/$name"
+        [ "$dev" = "$boot" ] && continue
+        printf '%s\t%s\t%s\n' "$dev" "$size" "${model:-}"
+    done
+}
+
+# ── wipe one disk ─────────────────────────────────────────────────────────────
+wipe_one() {
+    local dev="$1"
+    p ""
+    p "=== ${YEL}${dev}${NC} ==="
+
+    if echo "$dev" | grep -q '^/dev/nvme'; then
+        if nvme format --ses=1 "$dev" 2>&1; then
+            p "  ${GRN}nvme format OK${NC}"
+            blockdev --flushbufs "$dev" 2>/dev/null || true
+            return
+        fi
+        p "  nvme format failed — falling back to blkdiscard"
+    fi
+
+    if blkdiscard -f "$dev" 2>&1; then
+        p "  ${GRN}blkdiscard OK${NC}"
+        blockdev --flushbufs "$dev" 2>/dev/null || true
+        return
+    fi
+
+    p "  blkdiscard not supported — zeroing partition tables (HDD fallback)"
+    local size_bytes mb32 skip
+    size_bytes=$(blockdev --getsize64 "$dev" 2>/dev/null || echo 0)
+    mb32=$(( 32 * 1024 * 1024 ))
+
+    dd if=/dev/zero of="$dev" bs=4M count=8 conv=fsync status=progress 2>&1 || true
+
+    if [ "$size_bytes" -gt $(( mb32 * 2 )) ]; then
+        skip=$(( (size_bytes - mb32) / (4 * 1024 * 1024) ))
+        dd if=/dev/zero of="$dev" bs=4M count=8 seek="$skip" conv=fsync status=progress 2>&1 || true
+    fi
+
+    blockdev --flushbufs "$dev" 2>/dev/null || true
+    p "  ${GRN}done (partition tables zeroed)${NC}"
+}
+
+# ── main ──────────────────────────────────────────────────────────────────────
+banner
+
+BOOT=$(boot_dev)
+[ -n "$BOOT" ] && p "Boot device (excluded): ${CYN}${BOOT}${NC}\n"
+
+# build indexed list
+i=0
+DEVS=""
+IFS='
+'
+for line in $(list_disks); do
+    i=$(( i + 1 ))
+    dev=$(echo "$line" | cut -f1)
+    size=$(echo "$line" | cut -f2)
+    model=$(echo "$line" | cut -f3)
+    DEVS="${DEVS}${i}:${dev}:${size}:${model}
+"
+    printf "  ${CYN}[%d]${NC}  %-16s %8s  %s\n" "$i" "$dev" "$size" "$model"
+done
+IFS='
+'
+
+if [ "$i" -eq 0 ]; then
+    p "\nNo physical disks found (boot device excluded)."
+    p "Dropping to shell — type 'exit' to continue boot."
+    exec /bin/sh
+fi
+
+p ""
+pp "Enter numbers to wipe (space-separated), ${YEL}all${NC} for all, ${YEL}q${NC} to abort: "
+read -r SELECTION
+
+case "$SELECTION" in
+    q|Q|'') p "\nAborted."; exec /bin/sh ;;
+esac
+
+# resolve selection → list of devs
+SELECTED=""
+if [ "$SELECTION" = "all" ] || [ "$SELECTION" = "ALL" ]; then
+    SELECTED=$(echo "$DEVS" | grep -v '^$' | cut -d: -f2 | tr '\n' ' ')
+else
+    for num in $SELECTION; do
+        match=$(echo "$DEVS" | grep "^${num}:" | cut -d: -f2)
+        if [ -z "$match" ]; then
+            p "${RED}Unknown index: ${num}${NC}"; exec /bin/sh
+        fi
+        SELECTED="${SELECTED}${match} "
+    done
+fi
+
+SELECTED=$(echo "$SELECTED" | tr -s ' ' | sed 's/ $//')
+
+p ""
+p "Selected for wipe: ${YEL}${SELECTED}${NC}"
+p "${RED}WARNING: This is IRREVERSIBLE. All data on the selected disks will be lost.${NC}"
+p ""
+pp "Type  YES  to confirm, anything else to abort: "
+read -r CONFIRM
+
+if [ "$CONFIRM" != "YES" ]; then
+    p "\nAborted — no disks were touched."
+    exec /bin/sh
+fi
+
+p "\nStarting wipe..."
+for dev in $SELECTED; do
+    wipe_one "$dev"
+done
+
+sync
+p ""
+p "${GRN}=== All selected disks wiped and flushed. ===${NC}"
+p ""
+pp "Press Enter to reboot..."
+read -r _
+reboot

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

@@ -8,7 +8,7 @@
 # Layout (UEFI):  GPT, /dev/sdX1=EFI 512MB vfat, /dev/sdX2=root ext4
 # Layout (BIOS):  MBR, /dev/sdX1=root ext4
 #
-# Squashfs source: /run/live/medium/live/filesystem.squashfs
+# Squashfs sources: /run/live/medium/live/*.squashfs
 
 set -euo pipefail
 
@@ -62,9 +62,9 @@ for tool in parted mkfs.vfat mkfs.ext4 unsquashfs grub-install update-grub; do
     fi
 done
 
-SQUASHFS="/run/live/medium/live/filesystem.squashfs"
+mapfile -t SQUASHFS_FILES < <(find /run/live/medium/live -maxdepth 1 -type f -name '*.squashfs' | sort)
-if [ ! -f "$SQUASHFS" ]; then
+if [ "${#SQUASHFS_FILES[@]}" -eq 0 ]; then
-    echo "ERROR: squashfs not found at $SQUASHFS" >&2
+    echo "ERROR: no squashfs files found under /run/live/medium/live" >&2
     echo "  The live medium may have been disconnected." >&2
     echo "  Reconnect the disc and run:  bee-remount-medium --wait" >&2
     echo "  Then re-run bee-install." >&2
@@ -106,7 +106,10 @@ log "=== BEE DISK INSTALLER ==="
 log "Target device : $DEVICE"
 log "Root partition: $PART_ROOT"
 [ "$UEFI" = "1" ] && log "EFI partition : $PART_EFI"
-log "Squashfs      : $SQUASHFS ($(du -sh "$SQUASHFS" | cut -f1))"
+log "Squashfs      : ${#SQUASHFS_FILES[@]} layer(s)"
+for sf in "${SQUASHFS_FILES[@]}"; do
+    log "  - $sf ($(du -sh "$sf" | cut -f1))"
+done
 log "Log           : $LOGFILE"
 log ""
 
@@ -163,7 +166,9 @@ log "  Mounted."
 
 # ------------------------------------------------------------------
 log "--- Step 5/7: Unpacking filesystem (this takes 10-20 minutes) ---"
-log "  Source: $SQUASHFS"
+for sf in "${SQUASHFS_FILES[@]}"; do
+    log "  Source: $sf"
+done
 log "  Target: $MOUNT_ROOT"
 
 # unsquashfs does not support resume, so retry the entire unpack step if the
@@ -177,9 +182,9 @@ while true; do
     fi
     [ "$UNPACK_ATTEMPTS" -gt 1 ] && log "  Retry attempt $UNPACK_ATTEMPTS / $UNPACK_MAX ..."
 
-    # Re-check squashfs is reachable before each attempt
+    mapfile -t SQUASHFS_FILES < <(find /run/live/medium/live -maxdepth 1 -type f -name '*.squashfs' | sort)
-    if [ ! -f "$SQUASHFS" ]; then
+    if [ "${#SQUASHFS_FILES[@]}" -eq 0 ]; then
-        log "  SOURCE LOST: $SQUASHFS not found."
+        log "  SOURCE LOST: no squashfs files found under /run/live/medium/live."
         log "  Reconnect the disc and run 'bee-remount-medium --wait' in another terminal,"
         log "  then press Enter here to retry."
         read -r _
@@ -194,12 +199,17 @@ while true; do
     fi
 
     UNPACK_OK=0
-    unsquashfs -f -d "$MOUNT_ROOT" "$SQUASHFS" 2>&1 | \
+    for sf in "${SQUASHFS_FILES[@]}"; do
-        grep -E '^\[|^inod|^created|^extract|^ERROR|failed' | \
+        log "  Unpacking $(basename "$sf") ..."
-        while IFS= read -r line; do log "  $line"; done || UNPACK_OK=$?
+        unsquashfs -f -d "$MOUNT_ROOT" "$sf" 2>&1 | \
+            grep -E '^\[|^inod|^created|^extract|^ERROR|failed' | \
+            while IFS= read -r line; do log "  $line"; done || UNPACK_OK=$?
+        [ "$UNPACK_OK" -eq 0 ] || break
+    done
 
     # Check squashfs is still reachable (gone = disc pulled during copy)
-    if [ ! -f "$SQUASHFS" ]; then
+    mapfile -t SQUASHFS_FILES < <(find /run/live/medium/live -maxdepth 1 -type f -name '*.squashfs' | sort)
+    if [ "${#SQUASHFS_FILES[@]}" -eq 0 ]; then
         log "  WARNING: source medium lost during unpack — will retry after remount."
         log "  Run 'bee-remount-medium --wait' in another terminal, then press Enter."
         read -r _

iso/overlay/usr/local/bin/bee-remount-medium

@@ -2,7 +2,7 @@
 # bee-remount-medium — find and remount the live ISO medium to /run/live/medium
 #
 # Run this after reconnecting the ISO source disc (USB/CD) if the live medium
-# was lost and /run/live/medium/live/filesystem.squashfs is missing.
+# was lost and /run/live/medium/live/*.squashfs are missing.
 #
 # Usage: bee-remount-medium [--wait]
 #   --wait  keep retrying every 5 seconds until the medium is found (useful
@@ -11,7 +11,7 @@
 set -euo pipefail
 
 MEDIUM_DIR="/run/live/medium"
-SQUASHFS_REL="live/filesystem.squashfs"
+SQUASHFS_GLOB="live/*.squashfs"
 WAIT_MODE=0
 
 for arg in "$@"; do
@@ -56,7 +56,7 @@ try_mount() {
     local tmpdir
     tmpdir=$(mktemp -d /tmp/bee-probe-XXXXXX)
     if mount -o ro "$dev" "$tmpdir" 2>/dev/null; then
-        if [ -f "${tmpdir}/${SQUASHFS_REL}" ]; then
+        if find "${tmpdir}/live" -maxdepth 1 -type f -name '*.squashfs' 2>/dev/null | grep -q .; then
             # Unmount probe mount and mount properly onto live path
             umount "$tmpdir" 2>/dev/null || true
             rmdir "$tmpdir"  2>/dev/null || true
@@ -82,8 +82,9 @@ attempt() {
     for dev in $(find_candidates); do
         log "  Trying $dev ..."
         if try_mount "$dev"; then
-            local sq="${MEDIUM_DIR}/${SQUASHFS_REL}"
+            local count
-            log "SUCCESS: squashfs available at $sq ($(du -sh "$sq" | cut -f1))"
+            count=$(find "${MEDIUM_DIR}/live" -maxdepth 1 -type f -name '*.squashfs' 2>/dev/null | wc -l | tr -d ' ')
+            log "SUCCESS: ${count} squashfs layer(s) available under ${MEDIUM_DIR}/live"
             return 0
         fi
     done
@@ -100,5 +101,5 @@ if [ "$WAIT_MODE" = "1" ]; then
         sleep 5
     done
 else
-    attempt || die "No ISO medium with ${SQUASHFS_REL} found. Reconnect the disc and re-run, or use --wait."
+    attempt || die "No ISO medium with ${SQUASHFS_GLOB} found. Reconnect the disc and re-run, or use --wait."
 fi

iso/overlay/usr/local/bin/bee-wipe-disks

@@ -0,0 +1,132 @@
+#!/bin/bash
+# bee-wipe-disks — erase all physical disks (interactive, confirmation required)
+#
+# Triggered automatically when the kernel cmdline contains bee.wipe=all.
+# Can also be run manually from a root shell.
+#
+# Wipe strategy:
+#   NVMe  — nvme format (ATA-style secure erase, fast)
+#   Other — blkdiscard -f (TRIM/UNMAP, fast on SSDs)
+#           dd if=/dev/zero  (fallback for HDDs, zeros first+last 32 MB)
+
+set -euo pipefail
+
+RED=$'\033[1;31m'
+YEL=$'\033[1;33m'
+GRN=$'\033[1;32m'
+NC=$'\033[0m'
+
+banner() {
+    echo ""
+    echo "${RED}╔══════════════════════════════════════════════════════════╗${NC}"
+    echo "${RED}║      BEE DISK WIPE — ALL DATA WILL BE DESTROYED          ║${NC}"
+    echo "${RED}╚══════════════════════════════════════════════════════════╝${NC}"
+    echo ""
+}
+
+# ── find boot device to skip ──────────────────────────────────────────────────
+live_dev() {
+    local src
+    src=$(findmnt -n -o SOURCE /run/live/medium 2>/dev/null || true)
+    [ -z "$src" ] && return
+    # Strip partition suffix: /dev/sdb1 → /dev/sdb, /dev/nvme0n1p1 → /dev/nvme0n1
+    echo "$src" | sed 's/p\?[0-9]\+$//'
+}
+
+# ── enumerate target disks ────────────────────────────────────────────────────
+find_disks() {
+    local boot_dev
+    boot_dev=$(live_dev)
+
+    lsblk -d -n -o NAME,TYPE,SIZE,MODEL | while read -r name type size model; do
+        [ "$type" = "disk" ] || continue
+        [ "$size" = "0B"  ] && continue           # empty virtual media
+
+        local dev="/dev/$name"
+        [ "$dev" = "$boot_dev" ] && continue      # skip boot device
+
+        printf '%s\t%s\t%s\n' "$dev" "$size" "$model"
+    done
+}
+
+# ── wipe one disk ─────────────────────────────────────────────────────────────
+wipe_disk() {
+    local dev="$1"
+    echo ""
+    echo "=== ${YEL}${dev}${NC} ==="
+
+    if echo "$dev" | grep -q '^/dev/nvme'; then
+        # NVMe format (ses=1 = user data erase)
+        if nvme format --ses=1 "$dev" 2>&1; then
+            echo "  ${GRN}nvme format OK${NC}"
+            return
+        fi
+        echo "  nvme format failed, falling back to blkdiscard"
+    fi
+
+    if blkdiscard -f "$dev" 2>&1; then
+        echo "  ${GRN}blkdiscard OK${NC}"
+        return
+    fi
+
+    echo "  blkdiscard not supported — zeroing partition tables (HDD fallback)"
+    local size_bytes
+    size_bytes=$(blockdev --getsize64 "$dev")
+    local mb32=$(( 32 * 1024 * 1024 ))
+
+    # Zero first 32 MB (MBR, GPT, filesystem superblocks)
+    dd if=/dev/zero of="$dev" bs=4M count=8 conv=fsync status=progress 2>&1 || true
+
+    # Zero last 32 MB (backup GPT)
+    if [ "$size_bytes" -gt $(( mb32 * 2 )) ]; then
+        local skip=$(( (size_bytes - mb32) / (4 * 1024 * 1024) ))
+        dd if=/dev/zero of="$dev" bs=4M count=8 seek="$skip" conv=fsync status=progress 2>&1 || true
+    fi
+
+    echo "  ${GRN}done (partition tables zeroed)${NC}"
+}
+
+# ── main ──────────────────────────────────────────────────────────────────────
+banner
+
+mapfile -t DISKS < <(find_disks | awk '{print $1}')
+
+if [ ${#DISKS[@]} -eq 0 ]; then
+    echo "No physical disks found (boot device excluded)."
+    echo "Nothing to wipe."
+    exit 0
+fi
+
+echo "Disks to be ${RED}COMPLETELY ERASED${NC}:"
+echo ""
+find_disks | while IFS=$'\t' read -r dev size model; do
+    printf "  ${YEL}%-16s${NC} %8s  %s\n" "$dev" "$size" "$model"
+done
+echo ""
+echo "${RED}WARNING: This is IRREVERSIBLE. All data on the listed disks will be lost.${NC}"
+echo ""
+printf "Type  YES  to confirm wipe, anything else to abort: "
+read -r CONFIRM
+
+if [ "$CONFIRM" != "YES" ]; then
+    echo ""
+    echo "Aborted — no disks were touched."
+    exit 0
+fi
+
+echo ""
+echo "Starting wipe..."
+
+for dev in "${DISKS[@]}"; do
+    wipe_disk "$dev"
+done
+
+echo ""
+echo "${GRN}=== All disks wiped. ===${NC}"
+echo ""
+printf "Reboot now to return to the boot menu? [Y/n] "
+read -r REBOOT
+case "${REBOOT:-Y}" in
+    [Nn]*) echo "You can reboot manually when ready." ;;
+    *)     echo "Rebooting..."; sleep 2; reboot ;;
+esac

scripts/build.sh

@@ -0,0 +1,125 @@
+#!/bin/sh
+# build.sh -- single entry point for ISO builds.
+#
+# Local build (default):
+#   sh scripts/build.sh
+#   sh scripts/build.sh --variant nvidia
+#   sh scripts/build.sh --clean-build
+#
+# Remote build (set BUILDER_HOST + BUILDER_USER in .env):
+#   sh scripts/build.sh
+#   sh scripts/build.sh --authorized-keys ~/.ssh/authorized_keys
+#
+# All flags are forwarded to build-in-container.sh.
+
+set -e
+
+SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
+REPO_ROOT="$(cd "${SCRIPT_DIR}/.." && pwd)"
+
+ENV_FILE="${REPO_ROOT}/.env"
+if [ -f "$ENV_FILE" ]; then
+    # shellcheck disable=SC1090
+    . "$ENV_FILE"
+fi
+
+BUILDER_HOST="${BUILDER_HOST:-}"
+BUILDER_USER="${BUILDER_USER:-}"
+
+# Cache lives inside the repo under dist/ (gitignored).
+CACHE_DIR="${REPO_ROOT}/dist/cache"
+
+# Forward all arguments as-is to the underlying build script.
+EXTRA_ARGS="$*"
+
+# ── Remote build ────────────────────────────────────────────────────────────
+if [ -n "$BUILDER_HOST" ]; then
+    if [ -z "$BUILDER_USER" ]; then
+        echo "ERROR: BUILDER_USER not set. Set it in .env."
+        exit 1
+    fi
+
+    echo "=== bee builder (remote: ${BUILDER_USER}@${BUILDER_HOST}) ==="
+    echo ""
+
+    cd "${REPO_ROOT}"
+    git fetch --quiet origin main
+    LOCAL=$(git rev-parse HEAD)
+    REMOTE=$(git rev-parse origin/main)
+    if [ "$LOCAL" != "$REMOTE" ]; then
+        echo "ERROR: local repo is not in sync with remote."
+        echo "  local:  $LOCAL"
+        echo "  remote: $REMOTE"
+        echo ""
+        echo "Push or pull before building:"
+        echo "  git push   -- if you have unpushed commits"
+        echo "  git pull   -- if remote is ahead"
+        exit 1
+    fi
+    echo "repo: in sync with remote ($LOCAL)"
+    echo ""
+
+    ssh -o StrictHostKeyChecking=no "${BUILDER_USER}@${BUILDER_HOST}" /bin/sh <<ENDSSH
+set -e
+REPO="/home/${BUILDER_USER}/bee"
+LOG=/tmp/bee-build.log
+
+if [ ! -d "\$REPO/.git" ]; then
+    echo "--- cloning bee repo ---"
+    git clone https://git.mchus.pro/reanimator/bee.git "\$REPO"
+fi
+
+cd "\$REPO"
+echo "--- pulling latest ---"
+sudo git checkout -- .
+git pull --ff-only
+chmod +x iso/overlay/usr/local/bin/* 2>/dev/null || true
+
+screen -S bee-build -X quit 2>/dev/null || true
+
+echo "--- starting build in screen session (survives SSH disconnect) ---"
+echo "--- log: \$LOG ---"
+screen -dmS bee-build sh -c "sh iso/builder/build-in-container.sh --cache-dir \$REPO/dist/cache ${EXTRA_ARGS} > \$LOG 2>&1; echo \$? > /tmp/bee-build-exit"
+
+echo "--- streaming build log (Ctrl+C safe -- build continues on VM) ---"
+tail -n +1 -f "\$LOG" 2>/dev/null &
+TAIL_PID=\$!
+while screen -list 2>/dev/null | grep -q bee-build; do
+    sleep 2
+done
+sleep 1
+kill \$TAIL_PID 2>/dev/null || true
+
+tail -n 20 "\$LOG" 2>/dev/null || true
+
+EXIT_CODE=\$(cat /tmp/bee-build-exit 2>/dev/null || echo 1)
+exit \$EXIT_CODE
+ENDSSH
+
+    echo ""
+    echo "=== downloading ISO ==="
+    LOCAL_ISO_DIR="${REPO_ROOT}/dist/release"
+    mkdir -p "${LOCAL_ISO_DIR}"
+    if command -v rsync >/dev/null 2>&1 && ssh -o StrictHostKeyChecking=no "${BUILDER_USER}@${BUILDER_HOST}" command -v rsync >/dev/null 2>&1; then
+        rsync -az --progress \
+            -e "ssh -o StrictHostKeyChecking=no" \
+            "${BUILDER_USER}@${BUILDER_HOST}:/home/${BUILDER_USER}/bee/dist/*.iso" \
+            "${LOCAL_ISO_DIR}/"
+    else
+        scp -o StrictHostKeyChecking=no \
+            "${BUILDER_USER}@${BUILDER_HOST}:/home/${BUILDER_USER}/bee/dist/*.iso" \
+            "${LOCAL_ISO_DIR}/"
+    fi
+    echo ""
+    echo "=== build complete ==="
+    echo "ISO saved to: ${LOCAL_ISO_DIR}/"
+    ls -lh "${LOCAL_ISO_DIR}/"*.iso 2>/dev/null || true
+    exit 0
+fi
+
+# ── Local build ─────────────────────────────────────────────────────────────
+echo "=== bee builder (local) ==="
+echo "cache: ${CACHE_DIR}"
+echo ""
+# shellcheck disable=SC2086
+exec sh "${REPO_ROOT}/iso/builder/build-in-container.sh" --cache-dir "${CACHE_DIR}" $EXTRA_ARGS

scripts/deploy.sh

@@ -48,7 +48,7 @@ echo "==> Сборка бинарника..."
 echo "    OK: $(ls -lh "${LOCAL_BIN}" | awk '{print $5, $9}')"
 
 LOCAL_SHA="$(shasum -a 256 "${LOCAL_BIN}" | awk '{print $1}')"
-REMOTE_SHA="$("${SSH_CMD[@]}" "$REMOTE" "if [ -f '${REMOTE_BIN}' ] && command -v sha256sum >/dev/null 2>&1; then sha256sum '${REMOTE_BIN}' | awk '{print \\$1}'; fi" 2>/dev/null || true)"
+REMOTE_SHA="$("${SSH_CMD[@]}" "$REMOTE" "if [ -f '${REMOTE_BIN}' ] && command -v sha256sum >/dev/null 2>&1; then sha256sum '${REMOTE_BIN}' | awk '{print \$1}'; fi" 2>/dev/null || true)"
 if [[ -n "${REMOTE_SHA}" && "${LOCAL_SHA}" == "${REMOTE_SHA}" ]]; then
     echo "==> Бинарник не изменился (${LOCAL_SHA}); копирование и перезапуск сервисов пропущены."
     exit 0