prompt: complete jax-grad-local-vs-ci-assertions — PyAutoArray#471 merged

[Jammy2211]

The fix merged (PyAutoLabs/PyAutoArray#471 → 2a6a416), so the task advances active/ → complete/2026/08/ and leaves active.md.

Written by lifecycle.py record, not by hand. lifecycle check: OK, lifecycle index: OK, dashboard regenerated.

What the record preserves

• The falsification. The prompt's prime suspect (numpy 2.2.6 vs 2.4.6) is dead: lp.py is byte-identical across numpy 2.2.6 / 2.4.6 / 2.5.2 and across 1-core vs 4-core. The likelihood runs through JAX/XLA; numpy only does the FD bookkeeping.
• Why "not the small-datasets cap" was ruled out wrongly — that check verified the resolved env, which was correct; the damage was baked into the FITS on disk by an earlier run, where no env check can see it.
• Four traps found during adversarial review, two of which would have shipped a worse bug than the one being fixed (a data.fits-by-name check is mandatory — a glob would delete every PSF-carrying dataset every run; and the predicate must be == not <=).
• That no tolerance was changed, and specifically that assert_eager_jit_consistent's rtol=1e-10 — the obvious thing to widen — is vindicated rather than under-specified.
• The imaging-only scope limit, so the next person doesn't assume the whole bug class is closed.
• Three follow-ups, one still unfiled (the stale jax<0.7 pin in smoke_install.sh, where CI is getting the right jax by accident).

🤖 Generated with Claude Code

https://claude.ai/code/session_01VEHLT33XpVcRt5YCJGLRMJ

---

Generated by Claude Code