Move the Phase 7 risk-summary JSON parsing into an exported, side-effect-free
function so the schema handling is unit-testable in isolation. Byte-equivalent
behavior for the legacy risk_categories/categories schema (existing 33 kg-phase
tests stay green); no functional change.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

The current risk-aggregator emits risk-summary.json keyed on exposure_by_category
with string exposures ("$433.75M") and string probability ("8% fail"); the parser
only handled the legacy risk_categories numeric schema, yielding 0 risk nodes.

- Add exposure_by_category as the primary category source (legacy keys preserved).
- Synthesize $-amounts from weighted_exposure/exposure_low/high when no numeric
  p10/p50/p90 bits exist (legacy numeric path unchanged).
- Pass string probability through (already carries % for the downstream regex).
- Emit Exposure BEFORE Mitigation so amounts.slice(0,5) leads with real exposures,
  not the RRTF figure embedded in mitigation prose.

Fixes #231 (parser half).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Contract anchor for #231: pins both the legacy risk_categories numeric schema
and the current exposure_by_category string schema, plus the
exposure-before-mitigation ordering, malformed-JSON fallback, and precedence.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

The live PostToolUse hook only persisted .md files, so the risk-aggregator's
risk-summary.json never reached the reports table and the KG risk layer + the
CRITICAL_REPORTS gate (risk-summary) silently failed.

- Add JSON_REPORT_FILENAMES allowlist (hookDBBridgeConfig.js) — exact basenames,
  NOT all .json, so *-state.json / banker-*.json / entities.json are excluded.
- Broaden the persist gate to the allowlist. persistReport is content-agnostic;
  extractReportType maps /review-outputs/ → review and extractReportKey strips
  .json → report_key 'risk-summary'. Fail-soft (all DB writes are try/caught).

Fixes #231 (live-persistence half).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Bring local backfill to parity with live persistence for recovery of
unpersisted sessions:
- walkMarkdown now also ingests allowlisted JSON deliverables (risk-summary.json)
  so the KG risk layer can be rebuilt (#231); scoped by basename, state sidecars
  excluded.
- Scan review-outputs/ (not just qa-outputs/) for *-state.json, so
  fact-validator / coverage-gap-analyzer / risk-aggregator states are captured.
- (Carries the banker_intake/banker_qa report-type matchers added during session
  recovery so banker artifacts get their canonical types.)

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

One-command recovery for a completed local session directory that never got a
sessions row: bootstraps the row (idempotent), derives transaction_name from
banker-deal-context.json, delegates to backfill-local-to-db.mjs, and optionally
fires the admin rebuild endpoints. Closes the gap that backfill-local-to-db.mjs
aborts when no sessions row exists.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Local, faithful reproduction of POST /api/admin/sessions/:key/rebuild-kg for
when no server / admin JWT is available: entity-synthesis + citation-synthesis
pre-steps then buildSessionKnowledgeGraph, honoring BANKER_QA_OUTPUT for the
banker KG phases. Upsert-only (mirrors the endpoint); --clean is opt-in.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

/#2)

Adversarial review found the node loop re-regexed the rendered block, so
mitigation $-figures leaked into exposure_amounts (R-ANT-001 captured the
$1,237,262,000 RRTF) and the title's % masqueraded as probability (65-80%
instead of the 8% fail prob; 53.8%→8% decimal truncation).

- buildRiskBlocksFromJson now emits structured exposureAmounts (from exposure
  bits only) and probability (first %-token of the probability field, else null).
- The node loop prefers these; markdown Path B (no structured fields) still
  falls back to whole-block regex — unchanged.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…tion

Pins audit findings #1/#2 with faithful R-ANT-001 shapes: mitigation RRTF figure
must not enter exposureAmounts; probability comes from the probability field not
a title %; non-quantified probability → null; legacy numeric path unchanged.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>