PR #30057 - manaflow-ai/cmux

Summary

On the fleet Macs, 3 acpmux tests failed only on macOS, on the STILL-RED list (STATUS 2359):

stop::a_slow_stop_does_not_fold_the_next_prompt_into_the_stopped_turn
forum a_rule_steered_into_a_running_turn_counts_as_told
hub::steer_end::race_tests::an_unsent_steer_gets_its_own_turn_after_another_steer_finishes

Cause (test under load).

Since #27600, once a session's start is over, the daemon puts its harness in PRIO_DARWIN_BG (cpu_class.rs, default background).
The stand-in agent (tests/fake_agent.py) times its Codex-steer turn with short sleeps: it polls 300 x 10 ms for a joined prompt, then waits 0.2 s.
On a loaded fleet Mac (load1 13 to 21 on 10 performance cores, from the daemon's own host-budget warning), a background-class stand-in did not end that turn in 18 to 30 s.
The diagnostic event dumps show the turn's first chunk ("turn: X"), then nothing: jobs 7057c820 (stop), ad77491f (steer race) and 8432fc15 (forum: no answer to "warm up").
The scratch commit 093a4d29 changed only the default class to normal. With it, all 3 pass on macOS: a35af9a6, 0b8d6072, 1935c6fb.

Fix (test environment only; the product default stays background).

daemon_env::vars (the hermetic environment of every acpmux test process and of the daemons the tests start) sets ACPMUX_SESSION_CPU_CLASS=normal.
tests/cpu_class.rs, which checks the product's class, asks for background in its own process and in its real daemon.
The steer race unit test (library, no hermetic environment) names normal in its stand-in's profile env.
New check stop::a_stand_in_harness_runs_at_normal_class_in_tests (macOS).

Cause detail (review). The class change alone is proved (093a4d29 changes only DEFAULT). Host load is only inferred: the daemon logged high load in the diagnostic runs, but not in the 2 reds at 33f01ce9. Darwin background timer coalescing of the stand-in's 10 ms sleeps is another possible mechanism, and it does not need load. To decide, one run of the steer race test at background class on an idle mini (owner follow-up).

Note for the owner of cpu_class.rs. A real harness that times its work with short sleeps can also run slowly at background class on a loaded Mac. That is the product's choice (#27600, cmuxterm-hq#4947), and this PR does not change it. It only stops the tests from measuring it by accident.

Testing

run	ref	result
red: new check	33f01ce93ed (test only)	macOS 3256ec1b (m4pro-7): FAIL, `Background` != `Normal`
red: stop / steer race at the test-only commit	33f01ce93ed	macOS 2fd97413 (m4pro-7) FAIL (hub_integration.rs:1323 timeout); a29da46e (m4pro-4) FAIL (steer_end_race_tests.rs:73)
green `stop::a_slow_stop_does_not_fold…` 5 of 5	37defdb6cae	macOS f2ccd79c m4pro-3, 66d4bc66 m4pro-8, 69f5abda m4pro-3, 5f443a81 m4pro-7, a77653ba m4pro-2 (each compiled acpmux, 1 passed, ~9.9 s)
green forum `a_rule_steered…` 5 of 5	37defdb6cae	macOS b8dfe9bf m4pro-7, 7d2cba69 m4pro-2, b0da60be m4pro-6, 14a9237c m4pro-7, 899bf74d m4pro-8
green steer race 5 of 5	37defdb6cae	macOS a49ada96 m4pro-2, 26b79ad4 m4pro-7, f78ba0e6 m4pro-3, 793bafaf m4pro-7, 2e8a6d05 m4pro-9
new check	37defdb6cae	macOS efe04dc8: pass
`cpu_class` tests (product class still background)	37defdb6cae	macOS f0c3f4af: 9/9
clippy	37defdb6cae	89db498a red only on base `hub/processes.rs:256` (fixed on feat by #30009)
after merging feat (0a45903b)		clippy Linux d0a4e04f: pass; `cpu_class` macOS 5dd8ef18: see the check comment

Observed / Not run

Observed: everything above. Not run: Linux (Linux has no CPU class: for_session returns normal there).

Changelog

none

Proof

No capture: a test-environment change.

^{Need help on this PR? Tag @codesmith-bot with what you need.}

Summary by cubic

Fixes three acpmux tests that timed out only on loaded fleet Macs because their stand-in harnesses ran at background CPU class once a session's start was over. This test-only change keeps the stand-ins at normal class; the product default (background) is unchanged.

The hermetic environment (daemon_env::vars) now sets ACPMUX_SESSION_CPU_CLASS=normal for test processes and the daemons they start.
tests/cpu_class.rs explicitly requests background in its own process and real daemon, so it still verifies the product's actual class.
The steer race unit test sets normal on its stand-in's profile env.
Adds a macOS check that a stand-in harness runs at normal class in tests.