orb(gate): the six anti-abuse close paths are all exempt from the precision breaker AND from live recheck

[JSONbored]

downgradeCloseToHold only considers a close downgradable when closeKind === "heuristic" (src/settings/agent-actions.ts ~617-620). Six close paths carry other closeKind values — blacklist, contributor_cap, review_nag, copycat, screenshot_table, linked-issue-hard-rule — so the close-precision circuit breaker returns their plans untouched in both Reason A (project breaker) and Reason B (per-rule precision). They also get no executor-side live recheck: every guard at src/services/agent-action-executor.ts ~416-439 requires closeKind === "heuristic".

The stated justification ("zero-hallucination, deterministic") is sound for blacklist and contributor_cap — those read facts about an account. It is not sound for screenshot_table, whose determinism is a regex heuristic over free-text markdown, and which is our top close reason (170 all-time). A deterministic rule can still be systematically wrong; that is exactly what the breaker exists to catch.

Related ordering hazard

The five anti-abuse short-circuits sit above both the conclusion === "skipped" return (~950) and the CI-pending settle (~994). So a screenshot-table close executes on a PR whose gate was never evaluated and whose CI is still running — "push a commit → closed before CI finishes."

Latent false-positive classes in the screenshot gate (code-confirmed, NOT currently firing)

I checked 8 recently screenshot-closed PRs against GitHub: all had zero images in the body, so those closes were correct and these paths are latent rather than active. Filed so they are fixed before they bite:

• Staleness on any second push (packages/loopover-engine/src/review/screenshot-table-gate.ts ~417-433): same evidence fingerprint + different head ⇒ violated. A rebase or CI retrigger without re-uploading screenshots closes a compliant PR. Amplified because presenceModeEvidenceFingerprint (~350-352) filters urls.length >= 2, so the common one-image-per-row layout fingerprints as "[]" and is permanently stale from push chore(release): prepare public gittensory launch #2.
• hasImageOutsideTable (~160-166) is a bare per-line regex with no code-fence awareness: a correct table plus a CI badge, logo, or fenced markdown example ⇒ violated.
• hasCommittedImageFile (~173-179) is scopedPaths.length === 0 || matchesAny(...), so when the gate is scoped by labels only it checks every changed path — a UI PR that adds any image asset is closed, table or not.

Fix

1. Split the breaker exemption: keep it for identity-based closes (blacklist, cap); subject content-inspection closes (screenshot_table, linked-issue rule 4) to the per-rule precision check with their own rule ids.
2. Move the short-circuits below the CI-pending settle.
3. Fix the three screenshot FP classes: fingerprint single-image rows, only treat evidence as stale when the push touched in-scope visual paths, and require a non-empty whenPaths for the committed-image check.

Refs #9030, #9040.

[linear-code]

JSO-6287