fix(comply): verify the tamper-evidence chain instead of asserting it

[rylinjames]

Audit §3.10 / Part 1 #17 — the comply module's flagship integrity claim.

Problem

summarize_audit_log computed its own rolling hash and reported it as the "tamper-evidence head". The technical file and the EU AI Act Art-12 mapping then cited "tamper-evidence hash-chain" as evidence — but nothing was ever verified. An edited / reordered / truncated audit log produced a clean summary and a signed conformity bundle. Meanwhile the runtime recorder already writes a real prev_record_hash/record_hash chain and ships a tested verifier (record.verify_record_chain) that comply never called.

Fix

• summarize_audit_log now verifies each file's chain with verify_record_chain (per-file — the recorder resets the chain per session). Adds chain_verified (True / False / None), chained/unchained file counts, and the first broken file+index.
• _iter_records counts skipped lines (parse_error_count, non_dict_count, unreadable_file_count) instead of silently dropping them. A broken chain or any dropped line flips the summary status to "tampered".
• The technical file renders the verdict (PASS / FAIL — chain broken at <file> #<idx> / n/a), not just an opaque head hash.
• Art-12 record-keeping flips to gap when the chain is broken or lines were dropped — tampered logs no longer count as covered evidence.

Tests

tests/test_comply_chain_verification.py (6 cases, all pass): valid chain verifies; edit / reorder / mid-file drop each break it with location; corrupt line counted + status flip; unchained ActionGuard logs report None (no false claim).

Existing test_comply.py (signed-bundle export + CLI export→verify round-trip) still passes. ruff clean.

🤖 Generated with Claude Code