regen-on-ci: branch regens first (master cache runs cancel and yield)

[thejackshelton]

What changed (.github/workflows/regen-on-ci.yml only):

• Concurrency groups:
  • Master's cache runs (push) share one group, regen-on-ci-master-warm, with cancel-in-progress, so only the newest master run matters.
  • Branch runs get a group per branch and per label. Before, every pull_request labeled event joined its branch's group, so any other label, such as fulltest, replaced a queued regen run.
• Branch regens first: the macOS runners are one pool of 5 jobs, and GitHub has no job priority. So a master cache run yields: its resolve job counts regen-on-ci runs that are not pushes and are queued, waiting, pending or in progress. If there are any, it records mode=skip and runs nothing; the next master push warms the cache instead. The resolve job gets actions: read to list the runs.

Regen time now (per-step seconds from the rounds' last-run.json):

run	round	wall	steps (seconds)
branch seld-r2-c 37275532666 (40 min)	chrome-1	20.5 min	pixel-capture 319, wpt 182, vectors 172, hit-capture 154, dpr-capture 152, native-gen 150, break-capture 129, capture 126, profile-rows 117, media-sweep 95, dpr-vectors 69, north-star 56, anim-capture 38; the rest are under 15
	host-1	5 min	lanes-host 264
	chrome-2	3.7 min	profile-rows 116, north-star 41
	sweep	9 min	tw-sweep 448
master 37283930399 (37 min, after #161)	chrome-1	19 min	pixel-capture 296, ua 241, wpt 144, native-gen 141, dpr-capture 132, hit-capture 131, profile-rows 99, break-capture 95, vectors 94, media-sweep 65
	host-1	8 min	lanes-host 429
	sweep	3.4 min	tw-sweep 138

• tw-sweep is no longer the long pole: 138 s since tw-sweep: compile on every core; Chrome checks on 8 pages, each capture one JSON string (same snapshot) #161, against 448 s before and about 3,000 s originally. So I did not shard it.
• The long pole is the chrome-1 round (about 19–20 min of Chrome-bound steps, two at a time on 3 cores), then lanes-host.
• Next levers:
  • split chrome-1's independent capture steps over 2 macOS jobs (costs one more macOS slot);
  • lower the per-round job setup, about 1.5 min each.

Tests: actionlint is clean. The yield takes effect on the first master push after this lands.

🤖 Generated with Claude Code