research: final compile-time census + close-out (#77)

[Jammy2211]

Overview

Final census of the defaults-live compile UX (cache #128 + autotune-off #132) and the remaining-speedup close-out for #77. Worst cold cell ≈ 35 s on the A100 (was ~70 min); warm ≈ 5–9 s, almost entirely tracing. Remaining levers: jax.export RULED OUT (deserialize 4 ms but recompiles ~156 s every process — bypasses the persistent cache); compile-parallelism flags inconclusive with a ~10 s once-per-machine ceiling; the tracing floor is jax-internal. Any further reduction comes from upstream JAX, not this stack — the compile-time arc (#71 → #74 → #77) is done.

Scripts Changed

• jax_compile/README.md — final census tables + close-out section
• jax_compile/export_probe.py — NEW: jax.export save/load probe
• jax_compile/results/ — census + export-probe records (CPU + A100)

Validation checklist

• Tests/smoke: n/a (additive research artifacts) · Review: CLEAN (1 commit, +564) · Heart: RED — 5 pre-existing unrelated reasons (list unchanged from prior acks), human-acked for open+merge 2026-07-17

🤖 Generated with Claude Code