fix(review): add a review-failure-burst signal alongside review-burst

[JSONbored]

Summary

• Closes fix(review): review-burst rule can't see a repeat-failure retry storm, only repeat-success republish #3926. The existing review-burst anomaly (findHottestReviewTargetForRepo) only counts SUCCESSFUL pr_public_surface_published events, so a repeat-failure retry storm (every attempt inconclusive, never reaching a publish) is invisible to it -- the exact incident c707394/fix(review): stop the medium-effort retry storm burning tokens on zero output #3747 fixed.
• Every AI review call already records a structured inconclusive boolean in ai_usage_events.metadata_json (set at record()'s call site in src/services/ai-review.ts) regardless of whether it published. Added findHottestInconclusiveReviewTargetForRepo, a companion query keyed on that field, and wired it into the same detectOutcomeAnomalies detector alongside the existing signal.
• Deliberately does NOT touch the status column (always "ok" for a completed call, inconclusive or not) -- the daily neuron-budget sum filters status='ok', so repurposing that field would have silently excluded genuinely-spent tokens from budget accounting. Uses the existing metadata field instead.
• The existing publish-burst signal is untouched; it still catches a different, valid incident shape (excessive successful re-reviews of an unchanged PR).

Scope

• Conventional Commit title, focused single fix.
• No issue-link checkbox in the template sense -- Closes #3926 above.

Validation

• npm run typecheck
• npx vitest run test/unit/ops-wire.test.ts -- 26/26 pass, including an E2E test reproducing the fix(review): stop the medium-effort retry storm burning tokens on zero output #3747 incident shape (repeated inconclusive calls, zero publishes) and a healthy-mix negative test.
• 100% branch coverage on the changed lines (verified via lcov).

Safety

• No secrets/private data.
• No auth/API/UI changes.
• Read-only observability change; never affects what blocks a live PR.

[superagent-security]

Superagent didn't find any vulnerabilities or security issues in this PR.

[codecov]

❌ 4 Tests Failed:

Tests completed	Failed	Passed	Skipped
11138	4	11134	7

View the top 3 failed test(s) by shortest run time

test/unit/moderation-config-db.test.ts > global moderation config DB round-trip (#selfhost-mod-engine) > a raw DB row with an over-max violation_decay_days is also clamped on READ (not just on write)

Stack Traces | 0.399s run time

AssertionError: expected 100 to be 3650 // Object.is equality

- Expected
+ Received

- 3650
+ 100

 ❯ test/unit/moderation-config-db.test.ts:108:41

test/unit/moderation-config-db.test.ts > global moderation config DB round-trip (#selfhost-mod-engine) > REGRESSION (gate-flagged): violationDecayDays above MAX_MODERATION_VIOLATION_DECAY_DAYS is CLAMPED, not passed through raw -- an unbounded value overflows Date arithmetic on the live close path

Stack Traces | 0.421s run time

AssertionError: expected 100 to be 3650 // Object.is equality

- Expected
+ Received

- 3650
+ 100

 ❯ test/unit/moderation-config-db.test.ts:91:41

test/unit/moderation-config-db.test.ts > global moderation config DB round-trip (#selfhost-mod-engine) > a violationDecayDays AT the max is preserved unclamped (boundary, not just strictly-under)

Stack Traces | 0.422s run time

AssertionError: expected 100 to be 3650 // Object.is equality

- Expected
+ Received

- 3650
+ 100

 ❯ test/unit/moderation-config-db.test.ts:101:41

test/unit/queue.test.ts > queue processors > a #1960 action-command verb with no dispatch handler wired yet (e.g. pause) is bailed out of the Q&A answer-card path, not misrendered as help (#2160)

Stack Traces | 0.792s run time

AssertionError: expected 1 to be +0 // Object.is equality

- Expected
+ Received

- 0
+ 1

 ❯ test/unit/queue.test.ts:22807:28

To view more test analytics, go to the Test Analytics Dashboard
_{📋 Got 3 mins? Take this short survey to help us improve Test Analytics.}