docs(agents): record Actions queue and merge procedure

AGENTS.md

@@ -42,14 +42,34 @@ The materialization contract is also covered by [`docs/doctoring/exact-artifact-
   target repository, and pull request number with `cancel-in-progress: true`;
   do not include the head SHA, because that prevents a new head from cancelling
   its predecessor. Non-PR triggers need an explicit collision-safe fallback.
+- Do not let a rerun of an older run ID re-enter the live PR group and cancel
+  newer evidence. Use the PR number only for the first attempt and fall back to
+  `github.run_id` for reruns, or reject the rerun through exact-live-head
+  admission before it can displace current work.
 - Put concurrency at workflow scope when queued jobs must be coalesced before a
   runner is admitted. Job-level concurrency cannot relieve a saturated runner
   queue because it is evaluated only after job admission.
+- Subscribe only to pull-request actions that can produce useful work. The
+  default review set is `opened`, `synchronize`, `reopened`, and
+  `ready_for_review`; do not add `converted_to_draft` or `closed` merely to run
+  a job-level false condition, because the workflow run still enters the
+  organization queue. Add a lifecycle action only when that workflow performs
+  an explicit, tested cleanup or state transition for it. A draft or close
+  event whose purpose is to retire active evidence for that PR must share the
+  PR evidence concurrency group and publish the required exempt/terminal state.
+  An auxiliary cleanup that scans and cancels runs by API must instead use a
+  lifecycle-specific group so it cannot preempt current-head evidence; compare
+  the live PR state and head immediately before every cancellation. After a
+  cancellation reaches its terminal state, re-read both the PR and target run;
+  if the cancelled run now matches the live head, enqueue replacement evidence
+  instead of treating cleanup as successful.
 - Keep cleanup repository-local and event-driven. Do not restore an
-  organization-wide queue sweep, polling `sleep`, or another scheduled scan to
-  compensate for incorrect concurrency. Cancel only runs proven to belong to a
-  superseded head of the same PR, then verify each accepted cancellation
-  reaches `completed/cancelled`.
+  organization-wide queue sweep, unbounded sleep-based polling, or another
+  scheduled scan to compensate for incorrect concurrency. Cancel only runs
+  proven to belong to a superseded head of the same PR, then poll
+  `actions/runs/{run_id}` with a bounded retry. Treat cancellation as complete
+  only when `status == "completed"` and `conclusion == "cancelled"`; an HTTP
+  202 from cancel or force-cancel is not completion.
 - Classify a run's PR head by event-specific evidence before cancellation.
   `pull_request` may use the run's top-level `head_sha`, but
   `pull_request_target` records the trusted base there; use its PR association