[addisdev]

CI throughput plan (CI-OPTIMIZATION-PLAN-2026-10.md) Phase 0.1, 0.2, 0.6 and 3.3. Owner decision WB-617: heavy 4, Simulator 2, light up to 12.

What changes

Lanes in the admission hook (FLEET_ADMIT_LANES=off|observe|enforce, default off)

• Each job is simulator (existing runner rule; also takes a heavy slot), heavy or light, in this order: Simulator runner, FLEET_ADMIT_HEAVY_PATTERNS, FLEET_ADMIT_LIGHT_PATTERNS (space-separated repo:job shell globs on GITHUB_JOB, repo with or without owner), then the dashboard's GET /api/job-class (median of the job's last 30 ran_s from the admission log's release lines; light under FLEET_ADMIT_SHORT_JOB_S, ≥ 5 samples), else heavy. The hook caches the answer in .admission/job-class/ for 300 s and gives the daemon 3 s.
• Limits: FLEET_ADMIT_HEAVY_MAX_CONCURRENT (4), FLEET_ADMIT_LIGHT_MAX_CONCURRENT (12; the dead FLEET_ADMIT_SHORT_MAX_CONCURRENT is its alias), Simulator unchanged, under FLEET_ADMIT_TOTAL_MAX_CONCURRENT (default FLEET_ADMIT_MAX_CONCURRENT).
• FIFO per lane: the oldest light waiter is never behind an older heavy one. Waiter files are named <ts>-<pid>-<light|heavy>-<runner> so the lockless check finds its own line's head from one listing; old-format names read as heavy. Mutex, reaping, cancel polling, disk floors, timeout semantics and the strict Simulator rule are unchanged.
• observe: the single lane decides exactly as today; every row also records lane, lane_source, lane counts and limits, lane_would_hold, lane_reason and lane_wait_s (when the lanes would first have admitted a held job). off writes byte-for-byte the old line.
• Hold reasons name the lane (4 heavy job(s) already running, at the heavy limit of 4, an older job is waiting for the next heavy slot); the legacy reasons keep their words, so the lake's fct_admission_hold classifier still files them.

Dashboard: /api/job-class, /api/load-by-busy, 11 lane columns on admission_events (addColumn) plus an index on (repo, job, ts), a lanes block in /api/admission, lane rows and a lane column on the Capacity tab.

cockpit load: host load1 p50/p90 (and p90/core) by busy-runner count over --days N, from host_samples.

Phase 0.6: cleanup.sh erases an idle runner's _work/<repo>/DerivedData-*, spm-*, .build-cache-* with nothing written in FLEET_CLEANUP_RUNNER_CACHE_DAYS (7), all of them under pressure (never a building runner's, never _work/_*, never a git checkout), and writes any missing _work/.metadata_never_index. register.sh (and so mirror-runners.sh) already wrote it; ephemeral-runner.sh now does too.

Tests

• hooks/tests/run.sh: 150 pass locally (55 new: classification incl. cache and fallback, heavy/light/total limits, per-lane FIFO, Simulator counts as heavy, legacy alias, observe vs enforce, lockless per-lane check).
• scripts/test-cleanup.sh: 77 pass (15 new).
• Node: new job-class.test.js, lane cases in admission.test.js, route tests for both endpoints; related suites pass.
• Swift LoadTableTests run in CI only (laptop under load).

Rollout

Merge, deploy to ultra in observe (single lane keeps deciding at 8), review the observe data, then FLEET_ADMIT_LANES=enforce.

🤖 Generated with Claude Code