fix(review): gate every AI-spend call site on the repo's paused/frozen state

[JSONbored]

Summary

• Root cause of a real, confirmed financial incident: resolveRepoActionMode() is already computed once per pass in maybePublishPrPublicSurface and correctly threaded into every GitHub-write call (labels, comments, check-runs) — but was never threaded into the functions that actually spend real tokens: runAiReviewForAdvisory, runAiSlopForAdvisory, runLinkedIssueSatisfactionForAdvisory, runVisualVisionForAdvisory, plus the maintainer-invoked agent-run summary (attachPrivateAiSummary). Each was gated only by its own independent feature flag (aiReviewMode, slop.aiAdvisory, linkedIssueSatisfactionGateMode), completely orthogonal to the fleet-wide env brake, the DB freeze, and per-repo pause.
• Confirmed live: reactivating a repo with a large stale PR backlog let already-queued and newly-triggered review jobs keep spending for as long as the backlog took to drain — even after immediately re-engaging the freeze override. Two repos burned through 500 fresh AI-review/slop calls across 98 distinct PRs over several hours before the queue emptied, because the freeze stopped new sweep dispatches but never stopped an already-dequeued job from reaching the actual LLM call.
• Fix: thread the already-resolved mode into all five spend paths, with mode === "paused" as the very first check in each — gating at the point of spend, not just the point of dispatch, so a freeze that engages between enqueue and execution is still honored. "dry_run" still computes (so a maintainer can validate review decisions locally without publishing); only "paused" stops spend outright.

Scope

• The PR title follows type(scope): short summary Conventional Commit format.
• This PR is focused on one concern (AI-spend freeze gating), though it touches many call sites + their tests since the same required field was added consistently everywhere the audit found a gap.
• This follows CONTRIBUTING.md and does not reintroduce GitHub Pages, VitePress, site/, or CNAME.
• I linked a currently open issue this PR resolves — owner-authored PR fixing a live-confirmed financial incident; no separate issue was filed for this emergency follow-up fix.

Validation

• npm run typecheck
• npx vitest run test/unit/ai-review-advisory.test.ts test/unit/ai-slop.test.ts test/unit/linked-issue-satisfaction-run.test.ts test/unit/visual-vision-wiring.test.ts test/unit/rag-wiring.test.ts test/unit/repo-culture-profile-wiring.test.ts test/unit/reputation-wiring.test.ts test/unit/enrichment-wiring.test.ts test/unit/grounding-wiring.test.ts test/unit/impact-map-processor-wiring.test.ts test/unit/advisory-ai-routing-call-sites.test.ts test/unit/repository-settings-enforcement.test.ts test/unit/agent-orchestrator.test.ts — 279/279 pass, including 5 new dedicated regression tests (one per spend path) proving a "paused" mode never reaches the LLM/vision provider call even when every feature-specific flag would otherwise allow it.
• npx vitest run test/unit/queue.test.ts — full suite, 754/754 pass (this is the primary integration suite exercising the changed processors.ts call sites end-to-end).
• npm run actionlint / npm run test:workers / npm run build:mcp / npm run test:mcp-pack / npm run ui:openapi:check / npm run ui:lint / npm run ui:typecheck / npm run ui:build / npm audit — not run locally for this focused change; relying on CI (validate) for the full gate.

If any required check was skipped, explain why:

• This is a backend-only logic fix (no schema/API/UI surface); the full npm run test:ci gate is left to CI per the repo's own established practice for this size of change.

Safety

• No secrets, wallet details, hotkeys, coldkeys, user PATs, private keys, raw trust scores, private rankings, or private maintainer evidence are exposed.
• Auth, cookie, CORS, GitHub App, Cloudflare, or session changes include negative-path tests. — N/A, no auth/session surface touched.
• API/OpenAPI/MCP behavior is updated and tested where needed. — N/A, internal spend-gating logic only, no external surface.
• UI changes use live API data or real empty/error/loading states. — N/A, no UI changes.
• Visible UI changes include a UI Evidence section. — N/A, no UI changes.
• Public docs/changelogs are updated where needed. — N/A, internal automation behavior only.

Notes

• Root-caused live during an active incident: two repos (recently reactivated from an extended fleet-wide freeze) burned through real AI provider spend for hours despite the operator repeatedly re-engaging the freeze override, because the freeze mechanism only ever gated autonomous GitHub actions, never the AI compute itself.
• A few complementary hardening items surfaced during the investigation but are intentionally left for a follow-up (to keep this PR focused on the confirmed root cause): decoupling the auto_pause_after_reviewed_commits cost counter from GitHub-publish success (it currently undercounts when a publish step is slow/rate-limited), and a coarser per-(repo, PR) minimum-reinterval throttle spanning all three independent trigger sources (scheduled sweep, CI-completion webhook, live push webhook).