avds2 · Oct 1, 2026 · Oct 1, 2026 · Oct 1, 2026 · Oct 1, 2026 · Oct 1, 2026
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -1,5 +1,28 @@
 # Changelog
 
+
+## Unreleased
+
+- Accept IEC capacity units (GiB/MiB and related sizes) from newer dmidecode
+  output, restoring module/installed capacity and memory-array limits.
+- Bound Bluetooth controller discovery to five seconds with probe telemetry;
+  failed/truncated discovery leaves the controller count unknown instead of zero.
+- Stop and reap collector process trees before private staging cleanup on INT/TERM
+  or early failure, including producers in GNU timeout's separate process groups.
+- Add hanging/failed/empty/successful Bluetooth and INT/TERM cleanup regressions.
+- Add standard-profile typed SMBIOS RAM arrays/modules with capacity, DDR type,
+  manufacturer/part number, slot/bank, rank, ECC, rated/configured speeds and
+  explicit missing-tool/access/firmware/truncation coverage. Keep serial/asset
+  identifiers out of acquisition output.
+- Add typed CPU topology/cache/selected ISA features, firmware models, GPU PCI
+  name/driver/available VRAM, disk models/transport/capacity, system
+  battery full/design capacity health, available RAM and swap counters.
+- Reduce duplicate hardware text in automatic snapshots; retain focused evidence
+  under named hardware targets and `all`, without changing the schema or budgets.
+- Add a lossless dictionary-coded `context.ai.json` AI view with round-trip
+  validation, residual scanning and manifest/archive coverage.
+- Add hardware parser/sysfs/privacy fixtures and memory-collector integration tests.
+
 ## 1.0.0
 
 - Fixed the Docker local-socket regression fixture so it validates the security invariant (explicit local Unix-socket pinning) without assuming a rootless socket wins over a real `/run/docker.sock` on developer workstations.

diff --git a/README.md b/README.md
@@ -253,6 +253,7 @@ Typical output:
 ```text
 linux-context-HOST-YYYYMMDDTHHMMSSZ/
 ├── context.json
+├── context.ai.json            # lossless dictionary-coded AI view
 ├── manifest.sha256
 ├── REDACTION-REPORT.md
 ├── meta/
@@ -287,6 +288,25 @@ This is the file an AI should receive/read first. It contains:
 
 It is intentionally minified and optimized for machine consumption, not human aesthetics.
 
+### `context.ai.json` — lossless compact AI view
+
+For richer hardware inventories, this alternate entrypoint reduces repeated
+attribute names, device types and relation predicates using small dictionaries.
+It keeps every canonical fact, value, relationship, provenance row, coverage
+state and evidence route. Its embedded `encoding` explains how to read it.
+Send **one** entrypoint to an AI; there is no benefit to sending both.
+
+To reconstruct the original document exactly:
+
+```bash
+python3 -B -S lib/ai_view.py decode BUNDLE/context.ai.json reconstructed.json
+```
+
+Size savings depend on the inventory. If the dictionary header outweighs the
+savings, this file uses the original canonical encoding, so it is never larger
+in bytes than `context.json`. Actual token savings also depend on the model tokenizer.
+`context.json` remains the canonical v5 format for existing integrations.
+
 ### `meta/graph.json` — overflow graph
 
 Only present when the complete typed graph would exceed the profile's hard `context.json` ceiling. The full graph is preserved here rather than deleted; `context.json` becomes a bounded routing/summary layer.
@@ -348,7 +368,7 @@ Current domains include:
 
 - host identity, distribution and capabilities;
 - kernel/modules/taint/high-value sysctls and source configuration;
-- CPU/memory/PCI/USB/block hardware;
+- CPU topology/cache/features, RAM module/slot/type/speed/manufacturer, GPU model/driver/VRAM, and PCI/USB/block hardware;
 - filesystem/mount/swap/LVM/RAID/ZFS/Btrfs topology;
 - SMART/NVMe/filesystem health where safely available;
 - interfaces, addresses, routes, DNS, sockets and firewall topology;
@@ -371,6 +391,29 @@ Current domains include:
 
 The core does not assume all of these technologies exist. Feature detection is explicit; missing tools/subsystems are represented as unavailable rather than guessed.
 
+### Identifying installed RAM and hardware
+
+From `standard` onward, hardware identity is included as typed facts/entities in
+`context.json`; it does not require `--target all` or reading raw DMI evidence.
+
+```bash
+sudo ./bin/linux-context --profile max
+# Keep extra hardware evidence for a focused follow-up:
+sudo ./bin/linux-context --profile max --target hardware
+```
+
+With optional `dmidecode` installed and firmware access available, the model
+reports memory arrays and modules: capacity, DDR type, manufacturer/part number,
+slot/bank, form factor, rank, rated/configured speed and reported ECC policy.
+Firmware-installed capacity is distinct from RAM visible to Linux. Unknown
+identity, missing tooling and inaccessible firmware are explicit; the exporter
+never guesses DDR type from “16 GB” or installs dependencies automatically.
+
+CPU topology/cache/features, motherboard/BIOS models, GPU PCI name/driver/VRAM,
+disk model/transport and battery full/design capacity are also modeled
+when exposed by the host. Automatic runs omit redundant hardware command dumps;
+focused targets preserve extra evidence. See [hardware model details](docs/COLLECTORS.md#hardware-model-standard-and-above).
+
 ---
 
 ## Security and privacy model

diff --git a/bin/linux-context b/bin/linux-context
@@ -59,6 +59,7 @@ Recommended:
       Exhaustive target-specific diagnostic acquisition, still bounded/read-only/redacted.
 
 AI contract:
+  context.ai.json                     Lossless dictionary-coded AI view.
   context.json                        Canonical compact typed host graph/index.
   meta/evidence.json                  Evidence routing catalog; open only on demand.
   meta/collection.json                Collector/probe telemetry; not host context.
@@ -145,6 +146,18 @@ init_bundle
 on_exit() {
     local rc=$?
     trap - EXIT
+    trap '' INT TERM
+    # Background shells can ignore terminal SIGINT and continue writing into
+    # staging. Stop and reap their full trees before removing private material.
+    local -a workers=()
+    mapfile -t workers < <(jobs -pr)
+    if (( ${#workers[@]} )); then
+        if ! lctx_python "$LCTX_PROJECT_ROOT/lib/stop_workers.py" "${workers[@]}"; then
+            log ERROR "Unable to stop collectors; private staging retained at $LCTX_TMP_DIR"
+            exit "$rc"
+        fi
+        wait || true
+    fi
     cleanup_tmp || true
     exit "$rc"
 }
@@ -263,6 +276,11 @@ log INFO 'compiling compact canonical AI context'
 compile_context || fatal "canonical context compilation/validation failed: $LCTX_OUTPUT_DIR/meta/validation.json"
 # Defense in depth: sanitize compiler-generated JSON/metadata too, then rescan.
 sanitize_bundle_text_in_place || fatal 'final bundle redaction pass failed'
+# Derive a lossless AI view from sanitized canonical data, before the final
+# residual scan and manifest. It receives the same privacy/publication contract.
+lctx_python "$LCTX_PROJECT_ROOT/lib/ai_view.py" encode \
+    "$LCTX_OUTPUT_DIR/context.json" "$LCTX_OUTPUT_DIR/context.ai.json" \
+    || fatal 'lossless AI view generation failed'
 final_scan || fatal "residual secret-risk scan failed: $LCTX_OUTPUT_DIR/meta/redaction.json"
 # Redaction rewrites first-party JSON structurally, so validate once more afterward.
 compile_validation_only() {
@@ -282,5 +300,6 @@ cleanup_tmp
 if (( LCTX_ARCHIVE && ! LCTX_KEEP_DIR )); then remove_published_bundle; fi
 
 printf 'Bundle: %s\n' "$LCTX_OUTPUT_DIR"
+if (( LCTX_KEEP_DIR )); then printf 'AI context: %s/context.ai.json\n' "$LCTX_OUTPUT_DIR"; fi
 if [[ -n "$archive_path" ]]; then printf 'Archive: %s\n' "$archive_path"; fi
 exit 0
diff --git a/collectors/hardware/hardware.sh b/collectors/hardware/hardware.sh
@@ -2,7 +2,7 @@
 set -euo pipefail
 COLLECTOR_ID='hardware.inventory'
 COLLECTOR_MIN_PROFILE='standard'
-COLLECTOR_TARGETS='system hardware'
+COLLECTOR_TARGETS='system hardware cpu gpu memory'
 COLLECTOR_PRIVILEGE='optional-root'
 COLLECTOR_BASELINE=1
 COLLECTOR_DESCRIPTION='CPU, memory, PCI/USB devices, and safe block-device hardware inventory.'
@@ -21,14 +21,14 @@ collector_collect() {
         [[ "$value" =~ ^[0-9]+$ ]] && emit_fact hardware.memory.total_bytes "$value" /proc/meminfo observed 1.0 number
     fi
 
-    command_exists lscpu && run_capture cpu "$LCTX_COMMAND_TIMEOUT" "$LCTX_COMMAND_MAX_BYTES" -- lscpu || true
-    command_exists free && run_capture memory "$LCTX_COMMAND_TIMEOUT" "$LCTX_COMMAND_MAX_BYTES" -- free -h || true
-    command_exists lsblk && run_capture block_hardware "$LCTX_COMMAND_TIMEOUT" "$LCTX_COMMAND_MAX_BYTES" -- lsblk -e 7 -o NAME,KNAME,TYPE,SIZE,FSTYPE,FSVER,MOUNTPOINTS,RO,RM,ROTA,TRAN,MODEL || true
+    { target_requested hardware || target_requested cpu; } && command_exists lscpu && run_capture cpu "$LCTX_COMMAND_TIMEOUT" "$LCTX_COMMAND_MAX_BYTES" -- lscpu || true
+    { target_requested hardware || target_requested memory; } && command_exists free && run_capture memory "$LCTX_COMMAND_TIMEOUT" "$LCTX_COMMAND_MAX_BYTES" -- free -h || true
+    { target_requested hardware || target_requested storage; } && command_exists lsblk && run_capture block_hardware "$LCTX_COMMAND_TIMEOUT" "$LCTX_COMMAND_MAX_BYTES" -- lsblk -e 7 -o NAME,KNAME,TYPE,SIZE,FSTYPE,FSVER,MOUNTPOINTS,RO,RM,ROTA,TRAN,MODEL || true
     command_exists lspci && run_capture pci "$LCTX_COMMAND_TIMEOUT" "$LCTX_COMMAND_MAX_BYTES" -- lspci -nnk || true
     command_exists lsusb && run_capture usb "$LCTX_COMMAND_TIMEOUT" "$LCTX_COMMAND_MAX_BYTES" -- lsusb || true
 
     # Sysfs DMI is available on many systems even when dmidecode is not installed.
-    run_shell_capture firmware_identity 5 262144 '
+    target_requested hardware && run_shell_capture firmware_identity 5 262144 '
         for f in sys_vendor product_name product_version board_vendor board_name board_version bios_vendor bios_version bios_date; do
             p="/sys/class/dmi/id/$f"; [ -r "$p" ] || continue; printf "%s=" "$f"; cat "$p"; done
         [ -r /sys/devices/system/cpu/cpu0/microcode/version ] && { printf "microcode="; cat /sys/devices/system/cpu/cpu0/microcode/version; } || true' || true
@@ -37,12 +37,12 @@ collector_collect() {
         command_exists nvidia-smi && run_capture nvidia_gpu 10 524288 -- nvidia-smi --query-gpu=name,driver_version,pstate,memory.total,memory.used,temperature.gpu,utilization.gpu --format=csv,noheader,nounits || true
         command_exists sensors && run_capture sensors 10 524288 -- sensors || true
     fi
-    if (( EUID == 0 )) && profile_at_least deep && command_exists dmidecode; then
+    if target_requested hardware && (( EUID == 0 )) && profile_at_least deep && command_exists dmidecode; then
         # dmidecode includes unique Serial Number/UUID/Asset Tag fields by default.
         # Those add negligible diagnostic value and are intentionally excluded at
         # collection time. Keep model/topology/memory characteristics only.
         run_shell_capture dmi 15 "$LCTX_COMMAND_MAX_BYTES" '
-            dmidecode -t system -t baseboard -t memory 2>/dev/null |
+            dmidecode -t system -t baseboard 2>/dev/null |
             awk "
               /^(System Information|Base Board Information|Memory Device)$/ {print; next}
               /^[[:space:]]+(Manufacturer|Product Name|Version|Family|Type|Size|Form Factor|Locator|Bank Locator|Type Detail|Speed|Configured Memory Speed|Part Number|Rank|Configured Voltage|Minimum Voltage|Maximum Voltage):/ {print}

diff --git a/collectors/hardware/memory.sh b/collectors/hardware/memory.sh
@@ -0,0 +1,34 @@
+#!/usr/bin/env bash
+set -euo pipefail
+COLLECTOR_ID='hardware.memory'
+COLLECTOR_MIN_PROFILE='standard'
+COLLECTOR_TARGETS='system hardware memory'
+COLLECTOR_PRIVILEGE='optional-root'
+COLLECTOR_BASELINE=1
+COLLECTOR_DESCRIPTION='Typed RAM modules, firmware-reported slots/capacity/ECC, and explicit discovery gaps.'
+source "$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")/../../lib" && pwd)/collector_api.sh"
+
+collector_collect() {
+    local status='tool_missing' input=''
+    if command_exists dmidecode; then
+        status='access_or_firmware_unavailable'
+        # Filter identifiers BEFORE persistence. Headers retain local DMI handles
+        # for array/module links; serial numbers, UUIDs and asset tags are omitted.
+        if run_shell_capture memory_dmi 15 "$LCTX_COMMAND_MAX_BYTES" '
+            dmidecode --type 16,17 |
+            awk '\''
+              /^Handle 0x[0-9A-Fa-f]+, DMI type (16|17),/ { print; next }
+              /^(Physical Memory Array|Memory Device)$/ { print; next }
+              /^[[:space:]]+(Location|Use|Error Correction Type|Maximum Capacity|Number Of Devices|Array Handle|Size|Form Factor|Locator|Bank Locator|Type|Type Detail|Speed|Configured Memory Speed|Configured Clock Speed|Manufacturer|Part Number|Rank|Total Width|Data Width|Configured Voltage|Minimum Voltage|Maximum Voltage|Memory Technology):/ { print }
+            '\'''; then
+            if (( LCTX_CAPTURE_TRUNCATED == 0 )); then
+                input="$LCTX_SECTION_DIR/memory_dmi.txt"
+                status='no_records'
+            else
+                status='truncated'
+            fi
+        fi
+    fi
+    lctx_python "$COLLECTOR_LIB_DIR/hardware_model.py" memory "$input" "$status"
+}
+collector_main "$@"
diff --git a/collectors/hardware/platform.sh b/collectors/hardware/platform.sh
@@ -0,0 +1,46 @@
+#!/usr/bin/env bash
+set -euo pipefail
+COLLECTOR_ID='hardware.platform'
+COLLECTOR_MIN_PROFILE='standard'
+COLLECTOR_TARGETS='system hardware cpu gpu display power'
+COLLECTOR_PRIVILEGE='user'
+COLLECTOR_BASELINE=1
+COLLECTOR_DESCRIPTION='Typed CPU topology, firmware models, GPU drivers/VRAM, and battery capacity health.'
+source "$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")/../../lib" && pwd)/collector_api.sh"
+
+collector_collect() {
+    local cpu_file='' pci_file='' block_file=''
+    if command_exists lscpu; then
+        if probe_capture cpu_json 5 262144 -- lscpu --json; then
+            (( LCTX_CAPTURE_TRUNCATED == 0 )) && cpu_file=$LCTX_PROBE_FILE
+        fi
+    fi
+    if command_exists lspci; then
+        if probe_capture pci_models 5 262144 -- lspci -Dmm -nn; then
+            (( LCTX_CAPTURE_TRUNCATED == 0 )) && pci_file=$LCTX_PROBE_FILE
+        fi
+    fi
+    if command_exists lsblk; then
+        if probe_capture block_models 5 262144 -- lsblk --json --bytes -e 1,7 -o NAME,KNAME,TYPE,SIZE,RO,RM,ROTA,TRAN,MODEL; then
+            (( LCTX_CAPTURE_TRUNCATED == 0 )) && block_file=$LCTX_PROBE_FILE
+        fi
+    fi
+    # The helper reads only bounded allowlisted sysfs fields; no EDID, serials,
+    # arbitrary device properties, or writes to power-management controls.
+    local snapshot_file
+    if probe_capture platform_snapshot 10 "$LCTX_COMMAND_MAX_BYTES" -- \
+        python3 -B -S "$COLLECTOR_LIB_DIR/hardware_model.py" snapshot "$cpu_file" "$pci_file" "$block_file"; then
+        snapshot_file=$LCTX_PROBE_FILE
+        if (( LCTX_CAPTURE_TRUNCATED == 0 )); then
+            lctx_python "$COLLECTOR_LIB_DIR/hardware_model.py" platform "$snapshot_file"
+        else
+            record_collector_note 'platform-snapshot-truncated'
+        fi
+    fi
+    release_probe
+    local file
+    for file in "$cpu_file" "$pci_file" "$block_file"; do
+        [[ -z "$file" ]] || rm -f -- "$file"
+    done
+}
+collector_main "$@"
diff --git a/collectors/workstation/devices.sh b/collectors/workstation/devices.sh
@@ -50,13 +50,25 @@ collector_collect() {
     fi
 
     if command_exists bluetoothctl; then
-        local bt_list bt_count=0
+        local bt_list line bt_count=0
         emit_fact workstation.bluetooth.client_present true 'command -v bluetoothctl' observed 1.0 boolean
-        bt_list=$(LC_ALL=C bluetoothctl list 2>/dev/null || true)
-        if [[ -n "$bt_list" ]]; then
-            while IFS= read -r _; do bt_count=$((bt_count+1)); done <<< "$bt_list"
+        # BlueZ can wait indefinitely when the daemon/bus is unavailable. A
+        # failed probe means unknown, not zero controllers; never parse partial
+        # output or persist personal controller names from this discovery step.
+        if probe_capture bluetooth_list 5 262144 -- bluetoothctl list; then
+            if (( ! LCTX_CAPTURE_TRUNCATED )); then
+                bt_list=$(<"$LCTX_PROBE_FILE")
+                while IFS= read -r line; do
+                    [[ "$line" == Controller\ * ]] && bt_count=$((bt_count+1))
+                done <<< "$bt_list"
+                emit_fact workstation.bluetooth.controller_count "$bt_count" 'bluetoothctl list' observed 1.0 number
+            else
+                record_collector_note 'truncated:bluetooth_list (controller count unknown)'
+            fi
+        else
+            record_collector_note "bluetooth discovery unavailable:bluetooth_list (exit $LCTX_CAPTURE_RC; controller count unknown)"
         fi
-        emit_fact workstation.bluetooth.controller_count "$bt_count" 'bluetoothctl list' observed 1.0 number
+        release_probe
         # No controller is a normal state (for example bluez tools installed on a
         # desktop with Bluetooth disabled/absent), not an evidence failure.
         if (( bluetooth_focused && bt_count > 0 )); then