feat(profiling): add a compile-time axis to campaign coverage

[Jammy2211]

Closes #218. Phase 1 of 3 of the compile-axis arc.

What this fixes

The Profiling Agent could not see the compile-time corpus autolens_profiling is
already producing. AGENTS.md listed "JAX compilation-time profiling of likelihood
functions" under Future modes, _profiling.py had zero occurrences of "compile",
and all three modes read only results/runtime/ — while scripts/misc/jax_compile/
held 93 committed probe records that nothing cross-referenced against the science
grid.

campaign --axis compile now answers how much of the grid has compile data on a tier:

Cell/transform runs:  11 done · 136 missing (21 grid cells × 7 transforms)
Off-grid records (real measurements, not grid cells):
  datacube_img/mge/hst: 4 · imaging/delaunay_matern/hst: 12 · imaging/knn/hst: 11
Other hardware (neither tier):
  local_gpu_NVIDIA_GeForce_RTX_2060_with_Max-Q_Design: 2 record(s)
Malformed records:    4

That is the number the arc's other two phases are justified by: 2 of 21 grid cells
have any compile coverage, hst only, with no interferometer, datacube, jwst or ao
row anywhere.

Design decisions worth review

Records are placed by their own fields, not by path. Results are filed under
<hardware>/<model_type>, which drops the dataset class and instrument entirely, so
placement uses the in-record (dataset_class, model_type, instrument). A test files an
interferometer record inside mge.json to pin this.

The tier map is deliberately not TIER_CONFIGS. That map keys off sweep config
names which fold precision into the name (local_cpu_fp64 / local_cpu_mp), whereas a
compile record carries a raw hardware string plus a separate mixed_precision
bool. Reusing it would mis-bucket every row. A test pins that an fp64 and an mp record
on one machine are one cell/transform run.

other is a real answer, not a fallback. The corpus holds RTX-2060 rows belonging
to neither tier; folding them into local would report coverage on hardware nobody
asked about. Same reasoning for the off-grid bucket — knn, delaunay_matern and
the datacube_img* multi-band classes are genuine Prodigy-census and multi-band
measurements, so they are neither grid coverage nor noise.

Malformed records are surfaced, not skipped — with file and index, exposing the 4
that carry hardware, dataset_class and instrument all null.

Coverage only; it never compares two timings. jax_compile/README.md records the
first measurements being wrong by up to 7× (851 s vs 117 s for the same compile)
purely from host load, because XLA compiles on the host cores. Rows are comparable only
within (hardware, jax_version, mixed_precision, cache state), and comparison waits on
the pins in phase 2. A test asserts no timing value can leak into a coverage decision.

ingest / triage reject --axis compile with exit 5 rather than ignoring the
flag, so a compile flag can never silently return a runtime answer.

The transform axis is read from the workspace, via the same ast route load_grid
uses for CELLS, so the Brain cannot drift from probe.py's TRANSFORMS.

Verification

• 15 new tests in tests/test_profiling_conductor.py. Profiling was the only
  conductor without a test file, so the "runtime axis unchanged" requirement had nothing
  to assert against — the file is part of the work, not a bonus. Hermetic synthetic
  workspace fixtures on the test_hygiene_conductor.py pattern; stdlib + pytest only.
• Runtime axis verified byte-identical against the real workspace (stashed diff on
  both campaign and ingest), plus tests asserting the runtime decision shape gains no
  axis key and that the two tier vocabularies never cross.
• Full suite: 304 passed. Two test_skill_install.py failures reproduce on clean
  main in this environment (they assert Codex-skill install behaviour that a
  web-github / ci-only environment reports differently) and are untouched by this
  change.

Not in scope

Any judgement about whether a compile time is good (phase 3), any pin/baseline/
dashboard (phase 2), and executing probe.py — the agent reasons and emits a plan,
never runs the workspace.

Follow-ups

• Phase 2 — compile_warm_baseline_dashboard.md: warm rows are not machine-
  identifiable today (warmness lives only in a free-text tag, ~40 ad-hoc values, and
  cache_dir is non-empty on cold rows too), so probe.py needs an explicit
  cache_state before any warm-compile tracking is possible.
• Phase 3 — compile_axis_triage_drift.md: classify drift, and close the arc by
  moving compile-time profiling out of Future modes.

---

Generated by Claude Code