Epic: wave-2 waste/resource-abuse audit — cache gaps, redundant GitHub calls, missing backoff

[JSONbored]

Motivation

Following tonight's incident (the sweep endlessly re-visiting already-reviewed PRs, fixed in #4474, and a downstream cache-fingerprint bug this surfaced, fixed in #4481), we ran a second, broader audit specifically looking for the SAME classes of bug elsewhere in the review engine: unbounded/excessive periodic re-processing, AI-capability cache-coverage gaps, non-deterministic queries silently defeating a hash/fingerprint or eviction policy, redundant GitHub API calls within a single pass, and queue/cron scheduling waste (missing rate-limit admission or job dedup).

Every finding below was independently adversarially verified against the actual current source (not just claimed) before being tracked here — one candidate finding out of 14 was refuted and dropped. This is a real, evidenced backlog, not speculation.

Non-goals

• This epic is about correctness and resource waste, not new features. No new AI capabilities, no new public-facing surfaces.
• Findings that were investigated and found to be already-correctly-guarded (e.g. ai_review/ai_slop's existing frozen-reuse/pause mechanisms) are explicitly NOT re-litigated here.

Sub-issues (priority order — fix top-down for biggest impact first)

• fix(backfill): scheduled per-repo backfill ignores freshness/error-backoff, re-syncs every repo every 30 min forever #4497 — Scheduled backfill ignores freshness/error-backoff, re-syncs every registered repo every 30 min forever — P1, high severity/high confidence
• fix(review): PR mergeable-state + CI aggregate fetched live from GitHub twice per review pass #4498 — PR mergeable-state + CI aggregate fetched live from GitHub twice per review pass — P2, medium/high-confidence
• fix(review): grounding re-fetches every changed file's full body every 30 min with zero caching #4499 — Grounding re-fetches every changed file's full body every 30 min, no cache — P3, medium/high-confidence
• fix(review): impact-map re-embeds + re-queries the vector index up to 20x per pass with no result cache #4500 — Impact-map re-embeds + re-queries the vector index up to 20x per pass, no cache — P4, medium/high-confidence
• fix(db): missing tiebreak sort key defeats deterministic ordering in product-usage rollup + review-suppression eviction #4501 — Missing tiebreak sort key defeats deterministic ordering in two tables (product-usage rollup, review-suppression eviction) — P5, medium/medium-confidence
• fix(review): backlog-convergence-sweep lacks the anti-duplication guards its sibling agent-regate-sweep just got #4502 — backlog-convergence-sweep lacks the anti-duplication guards its sibling agent-regate-sweep just got — P6, medium/medium-confidence
• fix(upstream): fileUpstreamDriftIssues unconditionally PATCHes unchanged GitHub drift issues every 6 hours #4503 — fileUpstreamDriftIssues unconditionally PATCHes unchanged GitHub drift issues every 6h — P7, medium/medium-confidence
• fix(queue): reconcile-open-prs has no GitHub rate-limit admission or job dedup #4505 — reconcile-open-prs has no GitHub rate-limit admission or job dedup (flag off by default today) — P8, medium/high-confidence, no current blast radius
• fix(queue): refresh-installation-health loops per-installation with no GitHub-budget rate-limit admission #4506 — refresh-installation-health loops per-installation with no GitHub-budget rate-limit admission — P9, medium/medium-confidence
• fix(review): reputation signal re-queried twice per pass with zero memoization #4507 — Reputation signal re-queried twice per pass with zero memoization — P10, low severity
• fix(review): review-memory suppression list re-read from D1 on every comment render, no cache #4508 — Review-memory suppression list re-read from D1 on every comment render, no cache — P11, low severity
• fix(review): repo-culture-profile cache works correctly but has zero hit/miss telemetry #4509 — Repo-culture-profile cache works correctly but has zero hit/miss telemetry — P12, low severity, observability-only

Standing requirement for every sub-issue

Every fix in this epic must ship with invariant tests AND a regression test named for the specific bug, not just a happy-path check — this whole epic exists because a prior generation of similar bugs shipped without tests that would have caught the exact failure mode. Each sub-issue's own Requirements section restates this specifically for its bug.

Expected outcome

Every confirmed waste/correctness gap from tonight's audit fixed or explicitly tracked with a clear priority, each with regression coverage so this class of bug cannot silently reappear.

Effort

Epic — 12 independently shippable sub-issues, roughly 1-3 days total depending on how many are tackled.

[linear-code]

JSO-3334