fix(exact-review): retain actionable dispatch 422 detail

[brokemac79]

Summary

Preserve actionable, redacted GitHub validation detail for exact-review dispatch failures, and avoid silently parking a normal pull_request:synchronize update on an opaque HTTP 422.

Problem

On 2026-07-25 at 03:17:12Z, PR #842 received a user-authorized empty synchronize commit (21f465f05a7d382256de1b6c7a16abaf4d042cae -> 81e4abc894e7d3ec1fddbd856378b6aadb4392f3). At 03:17:15.782Z the exact-review queue parked revision 3 with:

state=parked
parked_reason=dispatch_rejected
status=422
class=permanent_rejection
fingerprint=dispatch-72215b4b

No exact-review worker run was created. The contemporaneous GitHub-activity run (30141979462) was intentionally skipped by its synchronize filter; an unrelated exact repository-dispatch review succeeded at 03:21:17Z, so this was not a general Actions outage.

#839 (cf01fe706b) classified every 400/404/422 as an item-scoped permanent rejection and retained only status/class/fingerprint. Its request helper consumed the GitHub response body but did not preserve a safe, actionable diagnosis.

Implementation

• Move source_head_sha under the existing queue_claim object. This keeps the exact dispatch client payload at 10 top-level keys even when review_options is present, while the workflow continues to read legacy top-level payloads.
• Parse only allowlisted 422 validation field/code tokens. Persist and expose those tokens, never the raw GitHub message, values, paths, headers, or request body.
• Park only structured client_payload validation as a genuine item-specific permanent rejection.
• Treat opaque or differently scoped 422s as validation_unknown: retain the item, block the dispatcher with bounded global backoff, and expose the redacted diagnostic.
• Preserve secondary-rate-limit 422 handling as global rate-limit backoff. Keep per-target admission 422s item-scoped.

Validation

pnpm run build:all
node --test --test-name-pattern='(bounds item-specific terminal-state check failures|parks an item after repeated item-specific target-state failures|structured client-payload validation|retains a synchronize update|secondary-limit 422|globally backs off GitHub outages)' test/dashboard-worker.test.ts
node --test --test-name-pattern='exact event item' test/sweep-workflow.test.ts
git diff --check

All commands above passed. actionlint is not installed on this Windows host. The focused workflow parser test passed. Broader workflow tests have existing native-Windows Bash availability failures unrelated to this patch.

Regression coverage proves:

• a fix(exact-review): supersede stale PR heads before dispatch #842-shaped command-plus-synchronize payload has exactly 10 client_payload keys, retains source-head semantics, and exposes an actionable global 422 outcome instead of parking silently;
• structured malformed client_payload validation parks with redacted field/code detail;
• secondary-limit 422 remains global rate-limit backoff;
• non-422 GitHub outage handling remains retryable; and
• no raw validation message, URL, value, or token-shaped test data reaches durable state.

Codex review closeout

• Dirty review: codex review --uncommitted initially found one accepted issue: shared 422 classification changed per-target admission failures to global scope. The patch now forces admission 422s back to item scope; focused admission tests were rerun.
• Final dirty review: codex review --uncommitted reported no actionable defects.
• Branch review: codex review --base origin/main reported no actionable correctness issues.

Risks and rollout

The change affects exact-review dispatch and sweep.yml payload compatibility. New producers emit queue_claim.source_head_sha; the workflow reads both nested and legacy top-level locations. Opaque 422s deliberately pause the shared dispatcher with capped backoff to prevent a retry storm, leaving the affected update visible for operators rather than discarding it.

The response body from the #842 incident is not recoverable from authorized read-only run logs, artifacts, or the public queue API; this patch records only a bounded allowlist needed for future diagnosis.

Links

• Incident target: fix(exact-review): supersede stale PR heads before dispatch #842
• Causal predecessor: fix(exact-review): classify dispatch failures and bound retries #839
• E2E-001 is planning-only; no sandbox or production event, state, comment, label, workflow, Worker, gate, or queue mutation was performed for this PR.

[brokemac79]

@clawsweeper review

[clawsweeper]

🦞🧹
ClawSweeper re-review requested.

I asked ClawSweeper to review this item again.
Action: item re-review queued (workflow sweep.yml, event repository_dispatch).
Result: when the review finishes, ClawSweeper will create the durable review comment if needed or update the existing comment in place.

[clawsweeper]

ClawSweeper status: review started.

I am starting a fresh review of this pull request: fix(exact-review): retain actionable dispatch 422 detail This is item 1/1 in the current shard. Shard 0/1.

This placeholder means the worker is alive and reading the current context. I will edit this same comment with the actual review when the claws are done clicking.

Crustacean status: shell secured, claws on keyboard, evidence pebbles being sorted.