perf(pr-merge): batch main's whole-suite runs and stop sleeping after the dispatch

.changes/pr-pipeline/perf-main-whole-suite-batching-9548bbe5.md

@@ -0,0 +1,10 @@
+---
+bump: patch
+---
+
+### Changed
+
+- pr-merge starts at most one full run of main per 30 minutes, of main's tip, covering every merge since the last. dispatch-full-run and post-merge-dispatch read the newest pr-require-checks.yml dispatches on main first (`Read main's full-run cover`): within the cap they start nothing, and post-merge-dispatch starts one on its next scheduled run for a tip no full run covers, whoever merged, while the lane is live (kill switch and report-only off). A read of the cover refused for a rate limit stops the job red and sends nothing more (`.github/workflows/pr-merge.yml`).
+- dispatch-full-run ends once it dispatches, where it held a runner up to 39 minutes polling the run; its timeout drops from 45 to 10 minutes. post-merge-dispatch now sends a red full run to post-merge-demotion-alert.yml on its next scheduled run while the lane is live (kill switch and report-only off), up to 15 minutes after the run ends, once per run (`pr-merge-alerted-<id>`, counted only when its own run on main uploaded it, by `babysit_postmerge.py --lane-alerted`), and skips a run that never started.
+- pr-require-checks.yml's dispatches on main (attempt 1) share one concurrency group with no `queue` key and `cancel-in-progress` false: one runs, one waits, and a newer one replaces the waiting one. A re-run and every other run keep their own group and cancel.
+- post-merge-demotion-alert.yml files nothing for a cancelled full run with no job, one the concurrency group replaced before it started.

.changes/source-control/perf-main-whole-suite-batching-3c4daaf0.md

@@ -0,0 +1,11 @@
+---
+bump: patch
+---
+
+### Fixed
+
+- The post-merge check (`babysit_postmerge.py`) skips a full run cancelled with no job on attempt 1, one pr-require-checks.yml's main-dispatch concurrency group replaced before it started: it no longer holds a merge as attribution-pending, asks for its re-run, or serves as the baseline of a later red run. A run cancelled after its jobs started still reads red.
+
+### Added
+
+- `babysit_postmerge.py --lane-alerted RUN_ID --branch <branch>` exits 0 when pr-merge.yml's post-merge-dispatch on the branch already sent that full run to post-merge-demotion-alert.yml, 4 when not, 3 when unreadable; 4 is a code no crash produces, and pr-merge.yml sends only on it. It counts the `pr-merge-alerted-<id>` record only from a sender the dispatch and re-run records' rule accepts (`_lane_sender`), so a same-name artifact from any other run is no record.

.github/workflows/post-merge-demotion-alert.yml

@@ -37,10 +37,10 @@ name: post-merge-demotion-alert
 # second intake dispatch. The record keeps needs-human, and closing a fix issue
 # changes nothing on it.
 #
-# A full run pr-merge.yml starts after each merge uses GITHUB_TOKEN, so its
-# completion raises no workflow_run event. pr-merge's dispatch-full-run job
-# sends a red one here through workflow_dispatch with its run id and head SHA
-# instead; the first step reads that run and refuses anything but a red
+# A full run pr-merge.yml starts for main's merges uses GITHUB_TOKEN, so its
+# completion raises no workflow_run event. pr-merge's post-merge-dispatch job
+# sends a red one here on its next scheduled run, through workflow_dispatch
+# with its run id and head SHA, instead; the first step reads that run and refuses anything but a red
 # pr-require-checks run on this repository's main that github-actions[bot]
 # dispatched, at that head SHA. Red is
 # any conclusion in babysit_postmerge.py's FAILED_CONCLUSIONS. A red run
@@ -76,7 +76,7 @@ on:
     workflows: [pr-require-checks, post-merge-detect-silent-revert]
     types: [completed]
     branches: [main]
-  # pr-merge.yml's dispatch-full-run job, with GITHUB_TOKEN, or a listed
+  # pr-merge.yml's post-merge-dispatch job, with GITHUB_TOKEN, or a listed
   # human (trusted-actors.json on main) names a red full run that raised no
   # workflow_run event. Both inputs are checked before use, and the head SHA
   # must be the run's own.
@@ -287,8 +287,11 @@ jobs:
               exit 0
               ;;
           esac
-          # Neither signal's runs are cancelled by concurrency, so a cancel
-          # comes from an actor with actions: write and counts as a failure.
+          # A cancel counts as a failure: pr-require-checks.yml cancels no
+          # running full run, so it came from an actor with actions: write or
+          # a time-budget overrun. A dispatch the main-dispatch concurrency
+          # group replaced before it started is cancelled too, with no job:
+          # that is no signal, and the check below files nothing for it.
           # A re-run that passes is recorded on the record; it clears nothing.
           # A dispatched run also files on action_required, as the Read step
           # accepts every conclusion in babysit_postmerge.py's FAILED_CONCLUSIONS.
@@ -325,6 +328,13 @@ jobs:
             echo "::error::run attempt '$RUN_ATTEMPT' is not a positive integer"
             exit 1
           fi
+          # A failed read is a run that started: it files as before.
+          if [[ "$RUN_CONCLUSION" == cancelled && "$signal" != canary ]] &&
+            jobs="$(gh api "repos/$GITHUB_REPOSITORY/actions/runs/$RUN_ID/attempts/$RUN_ATTEMPT/jobs?per_page=1" \
+              --jq '.total_count // (.jobs | length)')" && [[ "$jobs" == 0 ]]; then
+            echo "$RUN_WORKFLOW was cancelled before any job started (a newer dispatch replaced it): nothing to file"
+            exit 0
+          fi
           newly=''
           if [[ -z "$passed" && "$signal" != canary ]]; then
             # Re-run the failed jobs once and wait here for that attempt: a