feat(jax_compile): make warm compile machine-identifiable via cache_state

[Jammy2211]

Part of #103 (phase 2 of the compile-axis arc). This is the schema leg — the
prerequisite that unblocks pins and the dashboard; those remain open on the issue.

Why

Tracking warm compile per cell is the point of the arc: the persistent cache
turned 117.0 s into 2.3 s (CPU MGE vag) and 5517.8 s into 937.1 s (A100
end-to-end), and nothing watches for that reverting.

But warmness was recoverable only from the free-text tag, which carries ~40
ad-hoc spellings across this corpus — census-warm, census-warm2,
prodigy-census-warm-retry, cache-warm, mb_homo_cold — so the dashboard that
has to track warm compile could not tell which rows were warm.

What changed

probe.py derives cache_state from what the compile did: count cache entries
either side of lowered.compile().

value	meaning
cold	wrote a new entry — a MISS
warm	wrote nothing into a populated cache — a HIT
none	no --cache-dir
unknown	configured, empty, and nothing written — not silently warm

The trap this avoids, and tests against: cache_dir is non-empty on cold rows
too, because the cold run is the one that populates the cache. Any substring check
on the tag or a truthiness check on cache_dir gets this backwards.

Per transform, not per run. Each transform compiles its own module, so one
invocation can legitimately miss on one and hit on another — which the old per-run
tag could not express at all.

host_state (cpu_count + 1m load average) is added because XLA compiles on the
host cores. The README's measurement-discipline section records the first
numbers being wrong by up to 7× (851 s vs 117 s for the same compile) purely
from host load, and the corpus already mixes a 32-core RAL allocation with laptop
rows. Without it, "is this compile slower?" is unanswerable after the fact.

record_kind="compile_probe" makes the schema self-describing — export_probe.py
and trace_profile.py append their own shape into the same results tree.

Backfill

backfill_cache_state.py migrates the 89 existing compile records: 33 cold, 34
warm, 19 none, 3 unknown.

• Exact where cache_dir was empty → none.
• Inferred only from an end-anchored cold/warm tag.
• Everything else left unknown rather than guessed.

The anchoring is the point: mb_homo_cold_laxmap_gpu contains "cold" mid-tag and
is left unknown rather than mislabelled — precisely the false match an unanchored
parse would have made, and precisely the error cache_state exists to remove. The
4 sibling-instrument records are left untouched.

Verification

• 9 new tests in scripts/misc/test/test_jax_compile_cache_state.py, JAX-free.
• ruff check . and ruff format --check . clean across the repo.
• build_readme.py --check passes (dashboard idempotence gate).

CI caveat, worth knowing before merge: scripts/misc/test/ is not run by any
workflow — lint.yml has no pytest step — so these tests pass locally but are not
a gate here. Relatedly, test_vram_config.py::test_aggregate_matches_cell_prefixed_stems
is currently failing on main unnoticed. Both are flagged on #103 rather than
folded into this PR.

---

Generated by Claude Code