feat(profiling): ingest --axis compile — warm-pin drift detection

[Jammy2211]

Stacked on #219 — base is feature/compile-axis-campaign-coverage, not main,
because this needs that PR's load_compile_corpus / compile_tier_of. Merge #219
first and this retargets to main cleanly. Review the [two-commit diff]
(feature/compile-axis-campaign-coverage...feature/compile-ingest-pins).

Completes the Brain side of phase 2 (autolens_profiling#103). Pairs with
autolens_profiling#104, which owns the pins this reads.

Why

This is the surveillance the whole arc exists for. The persistent compilation cache
(117.0 s → 2.3 s CPU; 5517.8 s → 937.1 s A100) and --xla_gpu_autotune_level=0 (17×
on the pathological cold probe) are settings. A config drift or an XLA_FLAGS
clobber — the PyAutoNerves#127 failure that went undetected for two months — puts the
70-minute worst case straight back with nothing failing.

What it does

Reads the workspace's jax_compile/pins.json and reports warm rows that are
unpinned or have drifted. Every comparison happens strictly inside one
(hardware, hostname, jax_version, mixed_precision, cache_state). A jax_version
bump recompiles once by design, so it surfaces as a new unpinned key, never as
drift.

Against the real corpus: 25 pins, 0 drifted, 0 unpinned.

Two corrections that came from running it, not from designing it

1. Rows predating their pin are not drift. The first run flagged four — all
July-16 measurements that the July-28 pin had been chosen over. It was reporting
the improvement that set the pin as though it were a regression. Drift now requires a
row newer than its pin.

2. That exposed the deeper flaw, fixed in autolens_profiling#104: pins must be
sticky. With "most recent warm row wins", re-deriving pins after a cache
regression would have moved the pin onto the regressed value and every later check
would have reported all-clear forever. An existing pin now never moves without an
explicit --repin.

I would not have found either from the design alone — both only showed up against
real data.

Thresholds

Drift needs both gates: ≥ 2.0× the pin and ≥ 1.0 s absolute. The ratio
alone screams about sub-second cells where 100 ms of jitter is 3×; the floor alone
misses a cheap cell degrading by an order of magnitude. Both are generous because
host load alone has produced 7× errors in this corpus, and an alarm that cries
wolf gets ignored.

Tests

11 more (27 total in the file), including:

• a synthetic warm-reverting-to-cold row proving the alarm does fire;
• each comparability field proving it does not fire across the key;
• a jax_version bump classified as a new key rather than drift;
• both threshold gates asserted independently;
• a guard that the Brain's mirrored key definition matches the workspace's pins.py
  — mirrored rather than imported, since importing the workspace would drag the JAX
  stack into the Brain.

Full suite: 316 passed. The 2 test_skill_install.py failures reproduce on clean
main in this environment and are untouched.

triage --axis compile still exits 5 — classification is phase 3.

---

Generated by Claude Code