glaeda-cmux-runner-hook: give test-e2e's build and test the gui token

[teamleaderleo]

Why

The hook did not know manaflow-ai/cmux test-e2e.yml's build and test, so it treated both as a compile pinned to root 1 (persistent-dd + root). Neither job took the gui token. The cmux PR "ci: run E2E tests on the runner that built them" (branch ci/e2e-single-runner) moves E2E test execution into build, through the e2e-run-tests composite action. That action runs cmuxTests through run-in-console-session.sh, and cmuxUITests in the console user's GUI bootstrap. test stays only as a fallback. As a result, an E2E build on an owned mini ran tests in the shared console session while app-host-unit-tests, tests-build-and-lag or a rerun held the gui token. The old test job had the same gap.

What

• ("test-e2e.yml", "build") gets a new class, compile-gui: 2 units, tokens gui + root.
  • No persistent-dd. The job never runs owned_build_state.py, never sets CMUX_OWNED_STATE_ROOT or CMUX_SEED_LOCAL_CACHE, and adopts seeds from R2. The only DerivedData and CAS it wipes are <root>/derived-data-compile-admission and <root>/compile-admission-cas at the root it holds. The kept admission state lives in the store and is cloned in, so it is not touched.
  • Producer, not a root consumer. It takes the root its token picks, and every root path in the job uses ${CMUX_CI_CANONICAL_ROOT:-/private/tmp/cmux-ci}, so it is no longer pinned to root 1. Its own restore step re-takes that root through glaeda-canonical-root take, which is a no-op.
  • Root 1 first. Unlike compile admission, it skips the own-root preference: it tries root 1, then any free root. Main publishes seeds and caches keyed to root 1, since compile-app-host-test-product.sh adds root= to the fingerprint for any other root, and the E2E build has no per-root state that would favor another root.
  • Tokens are taken in the order gui, then root. When the gui token is taken, the refusal is the one GUI_WAIT_S (60 s) waits for.
• ("test-e2e.yml", "test") becomes gui, like app-host-test-rerun's rerun. With more than one root it is a consumer: its restore step takes the producer's root, which restore-app-host-test-product.sh reads from the receipt.
• Comments and docs/CMUX_MINI_RUNNER.md no longer say that bare build/test name test-e2e's jobs. The stale doc line listing app-host-test-rerun as an unknown job is fixed.

Trade-off

compile-gui takes the gui token at job start and holds it through the compile. The hook cannot see the later step where the tests start. So while an E2E build runs on a mini, other GUI jobs there wait 60 s and are then refused, and the owned-pool rescue moves them (REFUSAL_SECONDS 120 covers test-e2e too). Releasing the token during the compile would need a cmux change, such as a take-gui helper run before the test step, and this PR does not change cmux.

Tests

• New test_capacity_e2e_jobs_hold_the_gui_token (two roots). It checks:
  • build holds gui+root-1 without persistent-dd and exports the root.
  • The build's re-take of its own root is a no-op.
  • A compile admission still runs in root 2.
  • app-host-unit-tests, E2E test and a second E2E build each wait for the gui token, then are refused.
  • E2E test takes no root at job start and then takes the producer's root.
• test_job_class_keys_on_workflow_file_and_job_id is extended, including a check that every class named in the tables has a cost.
• scripts/test-glaeda-cmux-runner.py: 152 OK. scripts/test-glaeda-fleet-status.py: 65 OK.

Review

An independent read-only review of the first head found no correctness problems (persistent-dd, root paths, deadlock and stranding all checked). It found one performance problem: a build on runner instance 1 would compile cold at root 2. That is fixed in 22a2241, which also clarifies the docs for the one-root case. Full suite on 22a2241: 152 OK.

Rollout

After merge: scripts/glaeda-cmux-runner-fleet --apply --hosts <canary>, then the rest. job-started runs the hook fresh for each job, so no restart is needed.

🤖 Generated with Claude Code