[Jammy2211]

The current HST/MGE probe exposes rounded single-process timings and cannot demonstrate persistent-cache reuse. Add an opt-in raw CPU pilot runner for the same fixed-parameter likelihood, with separate fresh cold/seed/warm processes, unrounded phase timings and a dedicated steady-runtime process. This performs no inference.

The command prints a plan by default. Explicit execution preserves partial failures, isolates outputs/caches, caps and reaps each worker, validates actual context, and requires the target executable's warm cache key to match its seed. Dirty or unknown source identities are rejected. All observations remain accepted: false; these records do not enter the baseline reader or historical pins.

Closes #415.

Scripts Changed

• scripts/misc/jax_compile/pilot.py: plan, sequential orchestration, receipt and matched-context summaries.
• scripts/misc/jax_compile/pilot_worker.py: fixed-parameter construction, phase timing, provenance and target cache diagnostics.
• scripts/misc/test/test_jax_compile_pilot.py: protocol, isolation, mismatch, interruption and process tests.
• scripts/misc/jax_compile/PILOT.md and README: method, resource boundaries and qualification gaps.
• Generated catalogue/dashboard metadata and navigation; no timing record or measurement date changed.

Validation

• Full existing suite plus initial pilot tests: 1,452 passed, 6 skipped (before the final provenance fix); final targeted suite: 32 passed.
• Final producer source b5b8160: 11x11 synthetic image, 3x3 PSF, CPU/fp64, one repetition, 60-second per-worker cap completed four distinct cold/seed/warm/runtime processes. Target states miss/miss/hit/miss, seed/warm keys match, ten raw runtime samples, every record scientifically unaccepted. This is development validation, not production performance evidence.
• Independent review CLEAN after fixing dirty/unknown library provenance.
• Ruff, catalogue/Pulse contract, dashboard/readme idempotence, wiki, artifact-layout, timing inventory and wall-submit checks passed. Browser interaction/responsiveness checks passed.
• Existing per-section smoke imports and the new CLI help passed. Markdown link checking remains covered by PR CI.
• Heart STALE only: release validation incomplete: no rehearsal for current source. Development ship permitted; no release.

Remaining qualification work

Independent correctness reference/tolerances, exact production inputs and full resolved settings, reference host/environment lock and background-load method remain unresolved. Raw host load includes campaign work. Scheduler RAM is external; worker timeout/process count are enforced here. The default runner covers 31 processes (15.5 worker-hour upper bound); Pulse's separately reserved correctness process is not implemented.

No production compute, scientific acceptance, default/pin changes, broad baseline migration or inference is authorized. Tier: judge; human merge. Pulse handoff: PyAutoLabs/PyAutoPulse#55 (merge after this PR).