fix(queue): unblock parallel exact result publication

[brokemac79]

Summary

Replace the single global exact-review publisher with bounded publication admission (default 24) while preserving target-serialized comment routing.

Incident evidence

• #592 (30502fdf4e8a, merged 2026-07-15 10:38:45Z) introduced the split exact-review publisher and its global clawsweeper-exact-review-publisher concurrency group.
• #611 restored per-target publishing for the normal review path only; it did not remove the exact publisher's global serialization.
• #612 preserved/requeued durable exact-review lease tuples, but did not increase final-publication capacity.
• The test issue #613 reached the successful producer run (29454686822), proving the read-only review stage can run. Its final publisher path then exposed the separate durable-publication choke point.
• Live queue snapshot at 2026-07-15T23:25:29.922Z: 2,638 pending, 0 dispatching, 16 leased; openclaw/openclaw accounts for 2,342 pending rows. The 64-item handoff fleet has at least 48 unoccupied slots. This PR fixes the publisher bottleneck, not that separate review-admission/execution under-utilization.

Implementation

• Add EXACT_REVIEW_PUBLICATION_MAX_CONCURRENT, defaulting to 24 and clamped to the 64-item exact-review fleet cap.
• Admit and wake up to that many durable publication items; remove the workflow-wide publisher concurrency group that serialized them to one.
• Preserve the existing repair-comment-router.yml target-repository concurrency boundary. Exact publishers now hand every current, nonterminal verdict to that router rather than concurrently writing its shared ledger inline.
• Add queue, workflow, and routing-predicate tests, plus operator documentation.

No permission scope is broadened: the existing write-capable publisher token still owns GitHub mutation, and this change only uses the job's pre-existing actions: write permission to queue the existing router workflow.

Validation

Passed locally:

pnpm run build:all
node --test test/dashboard-worker.test.ts test/repair/event-apply-proof.test.ts test/clawsweeper.test.ts
node --test --test-name-pattern="exact event review hands immutable artifacts to the queue-bounded publisher lane|exact event workflow binds all work to canonical claim" test/sweep-workflow.test.ts
pnpm run lint:dashboard
pnpm run lint:repair
pnpm run lint:scripts
pnpm exec oxfmt --check <11 changed files>
git diff --check

The focused test run passed 200 behavioural tests; the exact-workflow selection and all changed-file format/lint checks passed.

Codex review closeout:

• An initial branch review found a real P1: parallel publishers could race the shared comment-router ledger. This patch removes the inline ledger writer and routes through the existing per-target serialized router.
• Final codex review --uncommitted: clean.
• Final codex review --base origin/main: clean — “No actionable defects were identified in the diff. The bounded publication admission and deferred routing changes are consistently wired through the queue, workflow, and tests.”

Read-only local ClawSweeper review:

pnpm run review -- --local-range --target-repo openclaw/clawsweeper --base origin/main

It reviewed e0e44656f856..eff4017042fdef, found no discrete correctness or security finding, and made no GitHub mutation. Artifact: .git/clawsweeper/reviews/local-range-1784157727786-15652/0.md.

actionlint is not installed on this host. The full workflow-helper suite has five baseline Windows/WSL failures because this Docker Desktop WSL distribution lacks /bin/bash; those tests exercise Linux-hosted Bash helpers outside this change. The targeted exact-workflow assertions above passed.

Risks and rollout

This is a bounded throughput change, not a claim that all current backlog causes are fixed. After deployment, retained exact publication items are automatically eligible for the 24-way lane; no manual replay is needed for that publisher backlog.

Monitor for 10–15 minutes:

1. publisher leases approaching 24 while durable comments/labels appear;
2. repair-comment-router.yml running serially per target without ledger conflicts;
3. total pending and especially openclaw/openclaw pending declining;
4. the separate idle-fleet symptom (16 active / at least 48 free in the cited snapshot). If it persists, treat it as the next root-cause investigation rather than increasing this cap again.

Links

#592 · #611 · #612 · #613 · producer run 29454686822 · publisher run 29455750741