kaizen: merge_group jobs wait behind PR CI for clean-room runners — reserve merge-queue capacity (queue-wait p90 33 min) #5394

Kaizen countermeasure #1 (build-process timing, hourly loop; measured 2026-10-15 ~11:10Z by aprender-flow, GraphQL reads only).

Measurement

Last 8 merge_group runs on main (10-14 23:17Z → 10-15 07:31Z) plus the live queue:

Loss on the critical path, summed over those 8 merges: runner queue-wait 99 min, shard over-run past x86-main 92 min, actions/checkout ≥3.5 min on 16 jobs (55 job-min).
Live, same cause: #5379's shards sat QUEUED 09:50 → 10:41Z (51 min). #5378's shards were QUEUED from 10:41Z and still waiting at 11:02Z, with 11 of 11 aprender-usable [Linux,X64,clean-room] runners busy. No merge into main 08:52Z → 11:0xZ (MERGE-DROUGHT andon D45/D90/QW).

Five whys

1. Why no merge for 2 h+? MQ jobs waited QUEUED for a runner (51 min and more).
2. Why? Every runner that can take [self-hosted, Linux, X64, clean-room] for aprender was busy.
3. Why? PR-push CI and merge_group CI ask for the same label set (ci.yml runs-on has no event-dependent form), so queue batches wait behind PR runs.
4. Why does that cost merges? GitHub has no job priority. Only a dedicated runner group or label reserves capacity, and none exists for merge_group.
5. Why none? The X64 clean-room pool was sized for throughput and never split by event. Merges are the KPI (≥5 PRs/h), and they get no reserved capacity.

Countermeasure (one)

Reserve merge-queue capacity: ≥2 X64 clean-room runners that only merge_group jobs can take (a dedicated label such as merge-queue, already known to actionlint, on runners that do not carry the plain PR label set). merge_group jobs (x86-main, workspace-test-shard) select it, e.g. via an event-dependent runs-on. The runner side is infra's; the ci.yml side needs the usual CI-workflow sign-off.

Done when (re-measure)

Over the next 8 merge_group runs after it lands: queue-wait p90 ≤ 5 min (from 33), and the MERGE-DROUGHT QW andon (MQ job QUEUED >15 min) silent for 24 h. aprender-flow re-measures hourly and posts the before/after in the kaizen | inbox line.

Owner: aprender-a3 (ci.yml side; runner side coordinated with the infra cop). Related: #5324 (CI pipeline umbrella), #3676 (CI diet).