feat(g5-observability): T1 — citation_verdicts table + audit endpoint + WORM export (v6.8.6)

[Number531]

Summary

T1 of three-tier G5 citation-verifier observability remediation. Closes the regulator-facing gap surfaced by the recent audit (post-PR #121): per-footnote verdicts from citation-websearch-verifier are now queryable in SQL and exposed in the audit-report endpoint + WORM regulator-handoff bundle, rather than buried in markdown.

Without this PR, a regulator asking "prove footnote [^VII.34] was verified" can see the tool calls (hook_audit_log) and the source (citation_source_links), but not the verdict the verifier ultimately assigned — undermining the 96.8% citation accuracy claim the GTM docs (just shipped in PR #121) lean on for EU AI Act Art. 12/13 transparency.

What ships

Component	Detail
citation_verdicts table	New junction table mirroring Wave 2 citation_source_links pattern. FK to reports(id) and sessions(id), both ON DELETE CASCADE. UNIQUE (report_id, footnote_id) for idempotent re-parse. Three indexes covering session+verdict, method, and report fetch paths.
Dual-path schema	migrations/015_*.sql + CITATION_VERDICTS_DDL in postgres.js + call in ensureHookSchema()
Parser promotion	test/sdk/_lib/certificateParser.mjs (PR #119) copied to src/utils/certificateParser.js. Test harness still imports from _lib/ to keep PR #119 fixtures green.
Fire-and-forget persistReport hook	When reportType==='qa' && reportKey==='citation-verification-certificate' lands in reports, the parser runs in backgroundTasks and writes verdicts via a single batch INSERT ... VALUES (one round-trip for up to ~500 footnotes). Idempotent upsert.
Audit endpoint extension	/api/session/:sessionKey/audit-report now returns citation_verification_certificate (full markdown + summary stats: confirmation rate, confirmed/unconfirmed/error/skip/pass_with_note/paywalled counts) and citation_verdicts (per-footnote array). report_version 1.0 → 1.1. Access logged to access_log (Wave 3).
WORM bundle inclusion	client-audit-export now ships citation_verdicts__csv.gz + citation_verification_certificate__csv.gz in the regulator-handoff bundle (both session-scoped and date-range modes).

Pattern precedent

This is the Wave 2 citation_source_links pattern, line-for-line:

• Table additive, FK ON DELETE CASCADE, UNIQUE for idempotent upsert
• Fire-and-forget via backgroundTasks Set in persistReport
• Try/catch around parser, non-fatal on failure
• Audit endpoint exposure with .catch(() => ({ rows: [] })) graceful fallback
• WORM-bundle CSV export

That pattern has shipped clean across Wave 1, Wave 2, Wave 3, and Exa A3 (v7.6.0-v7.6.2). The only Wave 2 incident was a missing-DDL hotfix caught in 24h; this PR uses dual-path (migration + ensureHookSchema) which prevents that.

Risk: 2/10

• Strictly additive table (no existing-schema mutation)
• Endpoint adds JSON fields (no field removal/rename) — report_version bump signals consumers
• All new queries gracefully fall back to [] on stale schemas
• Parser is pure logic (no I/O); failure is silent + non-fatal
• Batch INSERT avoids per-row round-trip serialization (the only hot-path concern from plan review)
• WORM bundle change is durable (Object Lock means the SKILL.md update ships permanently from this point), but the change is additive — existing bundles unaffected

Test plan

• DDL applies cleanly on fresh DB (CREATE TABLE + 3 indexes idempotent)
• DDL applies cleanly on production-shaped DB with existing reports table (FK ON DELETE CASCADE wires correctly)
• After a session runs, verify SELECT verdict, COUNT(*) FROM citation_verdicts WHERE session_id = ? returns expected distribution
• /api/session/:sessionKey/audit-report returns non-null citation_verification_certificate and populated citation_verdicts array
• client-audit-export --session <key> produces citation_verdicts__csv.gz + citation_verification_certificate__csv.gz in bundle
• Malformed/empty cert → parser returns silently, no rows written, no error log spam
• Stale schema (table missing) → endpoint returns 200 with empty arrays, audit report still generates

Follow-ups (separate PRs)

• T2 (~1d): Prometheus metrics (citation_verifier_confirmation_rate_pct, confirmed_total, unconfirmed_total, errors_total) + logInfo('citation_verifier_completed') + 3 alert rules. Risk score 1/10.
• T3 (~0.5d + measurement): OTel span wrap via SubagentStart/Stop in hookSSEBridge + _v3 tool metrics with agent_type label (dual-emit 7d). Risk score 3/10 — gated on Prometheus headroom measurement.

🤖 Generated with Claude Code