research: final compile-time census + remaining-speedup analysis

[Jammy2211]

Overview

Coda to #71/#74 (both merged): measure the compile-time UX exactly as users now get it (cache #128 + autotune-off #132 defaults live), and analyze where any remaining speedup would come.

Plan

• Census (jax_compile/probe.py): cold (fresh cache) + warm rows for mge & pixelization × jit/vag (+laxmap_vag mge) on local CPU and A100, wrapper defaults only (no manual flags).
• Remaining levers: (a) can jax.export AOT serialization beat the uncacheable tracing floor (~5–17 s) on repeat runs; (b) do XLA compilation-parallelism flags cut the residual no-autotune compile (MGE vag 117 s CPU / 29 s A100); (c) quantify the warm floor (trace + cache read).
• Deliverable: census table + final addendum in jax_compile/README.md; follow-ups only if a lever clears the bar.

Prompt: PyAutoMind draft/research/autolens_profiling/final_compile_time_census_and_remaining_speedup.md · branch feature/compile-census-final

[Jammy2211]

Merged via PR #78. Final census: worst cold cell ~35 s A100 (was ~70 min), warm 5–9 s (tracing). jax.export ruled out (recompiles ~156 s every process, bypasses the cache); parallelism flags inconclusive with ~10 s ceiling; tracing floor is jax-internal. Compile-time arc #71 → #74 → #77 complete — further speedup is upstream JAX's, not ours. Closing.