Mass `trio` deadline-miss failures on byte-identical code turned
out to be a firmware/EC power-cap (AMD PPT/STAPM) clamping the
all-core sustained clock while every static knob (`governor`,
`scaling_max_freq`, EPP, platform-profile) still read "performance"
— invisible to the existing `cpu_scaling_factor()` check. See
`scripts/cpu-perf-check` + the
`ai/conc-anal/trio_033_cancel_cascade_slowdown_depth3_issue.md`
notes.
Deats,
- add `_measure_sustained_headroom()` to `tests/conftest.py`: a
one-shot ~0.9s all-core burn (explicit `fork`-ctx `mp` procs)
sampling achieved-vs-max freq AFTER the boost window; under a 0.6
gate it returns the full inverse fraction (capped 4x), else 1.0;
best-effort 1.0 on non-linux or any error,
- add `cpu_perf_headroom()`: `max()` of the static scaling factor
and the (session-cached) sustained probe,
- inflate deadline budgets by it in `test_dynamic_pub_sub`, both
`test_clustering` cases, the
`test_multi_nested_subactors_error_through_nurseries` pexpect
waits + `test_nested_multierrors`,
- `xfail(strict=False)` `test_nested_multierrors` depth=3 under
throttle: the deep tree trips tractor's INTERNAL reap deadlines
(`soft_kill`/`hard_kill` `terminate_after=1.6`) minting a
`Cancelled` inside the runtime — not fixable by test-budget
inflation; auto-clears once the box un-throttles.
(this patch was generated in some part by [`claude-code`][claude-code-gh])
[claude-code-gh]: https://github.com/anthropics/claude-code
Gud Boi
20cb99ecd4