[avds2]

RAM identity previously lived mainly in optional DMI evidence, leaving the canonical model with only OS-visible capacity. This change adds standard-profile typed memory arrays/modules and broader CPU, GPU, firmware, disk and system-battery identity, with explicit discovery gaps instead of inferred DDR/vendor details.

Changes:

• Accept legacy and IEC capacity spellings from dmidecode (GB/GiB, MB/MiB, etc.) so installed modules and firmware array limits remain available across tool versions.
• RAM capacity/type/manufacturer/part number, slots, populated/empty state, rank, ECC policy and rated/configured speeds. Firmware-installed capacity stays distinct from OS-visible RAM. DMI serial/asset identifiers are filtered before persistence; failed/truncated probes cannot produce module facts.
• CPU topology/cache/selected ISA features, microcode and vulnerability/mitigation reports; motherboard/BIOS identity; GPU PCI model/driver/available VRAM; disk model/transport/capacity/flags; system battery full/design capacity; available RAM and swap counters. Loop and kernel RAM disks and peripheral batteries are excluded.
• Omit redundant hardware text during automatic snapshots while preserving focused evidence under named targets/all.
• Add context.ai.json, a lossless dictionary/template view with an embedded decoding guide and an exact round-trip gate. It falls back to canonical encoding when dictionary overhead would increase bytes. It shares final residual scanning, manifest, archive and transactional publication. Existing context.json v5 integrations remain compatible.
• Bound bluetoothctl list to five seconds; failed/truncated discovery leaves the count unknown and records probe telemetry. Stop collector descendants before private staging cleanup on INT/TERM or early exit, including producers in timeout process groups. Add Bluetooth timeout/failure/truncation and INT/TERM cleanup regressions.
• Document hardware limitations, coverage and AI ingestion; add hardware/sysfs/privacy/failed-probe/truncation and AI-view publication regressions.

Validation:

• IEC capacity regression and hardware collector tests plus syntax checks passed locally. Replaying an uploaded GiB-formatted DMI sample now produces two populated 8 GiB modules, 16 GiB installed and a 32 GiB array limit. Full GitHub Actions make check for c099dd9 passed on Python 3.9, 3.11 and 3.13 (run 36916993972).
• Follow-up hang/interrupt fix: new local Bluetooth and cancellation regressions plus syntax checks passed. All runnable local tests passed; ownership/publication and Docker fixtures remain constrained by the local namespace. Full GitHub Actions make check for commit 8f7beb0 passed on Python 3.9, 3.11 and 3.13, including ownership/publication and Docker fixtures (run 36915753279).
• Full GitHub Actions make check passed on Python 3.9, 3.11 and 3.13, including ownership and Docker fixtures (run 36839081843).
• Local syntax, hardware and AI-view tests passed. All non-Debian scripts and non-Docker Debian fixtures passed locally. The local execution namespace maps only UID/GID 0 and forbids Unix sockets, so full ownership/Docker coverage was verified on CI instead. No source tests were relaxed.
• max snapshots validated without structural errors; pre-existing evidence warnings were unchanged. Automatic evidence fell from 48,027 to 43,201 bytes on the restricted environment. The sparse snapshot correctly used canonical fallback for the AI view. A 32-module test model fell from 11,482 to 4,147 bytes with exact reconstruction; these are byte sizes, not tokenizer measurements.

Limits: module details remain dependent on dmidecode, privileges and firmware quality. Firmware slot/max-capacity claims are not guaranteed upgrade advice, disk identity may describe virtual media, and battery full/design ratio is an estimate, not a health test.