feat(review): add a submission-cadence/inter-arrival-time signal to the reputation and slop models to flag machine-paced submitters

[JSONbored]

Context

A maintainer review-stack × Autonomous Miner System (AMS) integration audit (2026-07-09) found that every anti-abuse signal in the review stack counts volume or ratios within a time window, but none of them has a submission-cadence (inter-arrival-time) dimension at all.

Evidence

• src/review/submitter-reputation.ts:144-152 — DEFAULT_REPUTATION_CONFIG has windowDays/minSample/fail-rate thresholds, no rate term.
• src/review/reputation-wire.ts:35-36 — REPUTATION_BURST_SUBMISSION_FLOOR is a raw count within a 90-day window, not a rate; by the module's own comments, a high-volume but well-formed, high-merge-rate submitter is deliberately never flagged as "low."
• src/signals/slop.ts (SLOP_WEIGHTS) — entirely per-PR content signals (whitespace churn, padding, description/commit quality, duplicate-cluster membership), no submitter-level temporal feature.
• A repo-wide search for cadence/velocity/submission-rate concepts anywhere in src/ turns up nothing related to PR-submission anti-abuse (only unrelated queue-backpressure/recap-scheduling hits).

Why this matters for the Autonomous Miner System

An autonomous agent can legitimately clear every existing deterministic/quality bar (good tests, real descriptions, linked issues, low close-rate) while operating at a cadence no human contributor could sustain — and today that cadence itself carries zero weight in any credibility or slop signal. A fast, well-formed, but strategically low-value fleet is invisible to the one dimension (superhuman pace) that is otherwise a strong tell for automated, high-volume gaming — distinct from simply widening the existing count-based thresholds, since a rate signal catches a pattern no count-over-a-window can.

Deliverables

• Add an inter-arrival-time / submission-cadence signal (e.g. rolling PRs-per-hour or median gap between an actor's submissions) to the submitter-reputation and/or slop scoring models
• Decide the response tier for a flagged cadence outlier (advisory-only vs. contributing to the existing burst/low-reputation downgrade) — should not penalize a genuinely fast but well-calibrated actor on its own
• Tests: synthetic superhuman-cadence-but-high-quality actor is now distinguishable from a normal human contributor; existing count-based signals unaffected
• Docs: document the new signal and its interaction with existing reputation/slop thresholds

---

Surfaced by a maintainer review-stack × AMS integration audit (2026-07-09). Related: #2350 (fleet-wide anti-farming detector) is the closest existing work but is scoped to merge-precision/reversal-rate/duplicate-election skew, not submission timing — this issue is a distinct, complementary signal.

[linear-code]

JSO-3359