fix(review): reputation-based AI-review skip must fail closed

[JSONbored]

Closes #9015.

LOOPOVER_REVIEW_REPUTATION=true is live on edge-nl-01, so this was an active production exposure, not a theoretical one.

The defect: when shouldSkipAiForReputation fired (low signal, or the submissions>=8 && merged<1 burst), aiReviewWillRun went false with no compensating hold. The contributor-controlled skip beside it fails CLOSED via maybeAddRequiredAutoReviewSkipHold; the reputation skip set no autoReviewSkipReason, so no hold branch fired. The PR then proceeded on deterministic checks only — CI, coverage, guardrails, secrets, linked-issue — none of which read code semantics — and could auto-merge with zero AI defect detection. A suspected abuser therefore received LESS scrutiny than a trusted contributor, exactly inverting the feature's intent. The issue verified three submitters sitting in the burst state at report time.

Fix 1 — fail closed (maybeAddReputationSkipHold): the exact sibling of the contributor-controlled hold, sharing its shouldRequirePublicAiReviewForAdvisory predicate, adding an ai_review_inconclusive finding so the gate holds for a human instead of passing silently. No-ops when no skip fired (the overwhelmingly common path) and when AI review is off for the repo (nothing was suppressed).

Fix 2 — restore the half-blind leg: getSubmitterCadence queried review_targets, which stopped receiving writes at the 2026-06-22 self-host cutover (newest row frozen there), so the machine-paced cadence signal never fired for anyone. It now reads the live pull_requests ledger (repo_full_name + case-insensitive author_login, created_at as the submission timestamp), making the second leg of the feature actually functional.

Tests: the hold fires under a reputation skip and stays silent on both no-skip and review-off arms; a query-shape assertion pins the cadence source to pull_requests and away from review_targets so this cannot silently regress at the next cutover. Changed-line coverage: all statements and branches.

Acceptance (from the issue): a burst-flagged author's PR now either gets a review or a named human hold — never a silent deterministic-only auto-merge.

[superagent-security]

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

[codecov]

❌ 1 Tests Failed:

Tests completed	Failed	Passed	Skipped
21561	1	21560	21

View the top 1 failed test(s) by shortest run time

test/unit/reputation-wiring.test.ts > shouldSkipAiForReputation (helper) > submission-cadence signal (#4514) > FLAG-ON: true for a machine-paced submitter even though every submission itself looks fine (quality-neutral)

Stack Traces | 0.295s run time

AssertionError: expected false to be true // Object.is equality

- Expected
+ Received

- true
+ false

 ❯ test/unit/reputation-wiring.test.ts:299:105

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