prompt: land the stranded point-source-dataset-cap-guard close-out (#710)

[Jammy2211]

Recovers the Mind close-out for PyAutoLens#710 (point-source-dataset-cap-guard), which was written but never reached main, and finishes the one piece it deferred.

Why this exists

/start_dev draft/bug/autolens/point_jax_vmap_parity_nondeterministic.md found the task already shipped:

Leg	PR	State
PyAutoLens	#711 — make the PointSolver SMALL_DATASETS short-circuit announce itself	MERGED
autolens_workspace_test	#265 — guard the point-source parity scripts against the cap	merged

PyAutoLens#710 is closed as completed. But the prompt was still sitting in draft/, the active.md entry never landed, and the sibling planned.md entry was still live — because the session that did the work left its Mind commits on claude/autolens-point-jax-vmap-parity-htyfi1, which never merged.

That branch had forked from an old main and could not be merged as-is: its diff against current main deletes ~5000 lines of subsequent work (scripts/lifecycle.py, tests/test_lifecycle_dates.py, a dozen complete/2026/08/ records). So its three relevant commits are cherry-picked here instead, onto current main.

Note that lifecycle.py check passed throughout this drift: a shipped-but-still-draft/ prompt with no active.md entry is invisible to the guard, which only checks that no active.md slug has a record. Worth a follow-up, not fixed here.

What lands

Cherry-picked from the stranded branch (a76314a4, 4a7cb760, b761a601)

• Prompt advanced draft/bug/autolens/point_jax_vmap_parity_nondeterministic.md → complete/2026/08/point-source-dataset-cap-guard.md (active.md entry added and removed, net zero, as the real lifecycle would have done).
• draft/bug/autolens/point_solver_error_bisect_health.md no longer absorbs the parity bug — the "same investigation surface" premise was false (a workspace env-declaration defect, not triangle-solve accuracy). Stale point_solver.py:111 reference corrected to :119.

New here (40ea0e11)

• jax-point-source-point-smoke-sentinel retired. The stranded branch marked it "retire this entry and its prompt when #710 ships". #710 has shipped. Its -1e99 symptom was the same poisoning, and prompt: re-home pynufft removal to maintenance, file two follow-ups #265 measured point.py passing on main with the pin exact (-83.38049778) — which settles the "what is still unknown" question that prompt's own 2026-08-09 note left open. Entry removed from planned.md; prompt archived to complete/archive/shelved/ behind a WITHDRAWN banner, provenance only.
• Completion record's reconciliation section rewritten from "retire" to "retired", naming the archive path.
• Filed: 2026-07-22 backfilled onto the folded original prompt — main added it to the draft after the close-out branch forked.

The underlying finding, for the record

The prompt's leading hypothesis was right, and the mechanism was worse than it guessed. Under PYAUTO_SMALL_DATASETS=1, should_simulate() rmtrees the committed, gitignore-allowlisted dataset/point_source/simple — it is JSON-only, so there is no data.fits carrying the SMALLDAT stamp that PyAutoArray#471's guard reads — and PointSolver.solve short-circuits to a fixed, model-independent position pair. One run poisons the next. That is exactly the reported "run A → -1e+99, runs B/C/D → 16.131221" signature. No library numerical bug exists; the prompt's repro command set the cap by hand, overriding the script's own ENV: jax full_datasets declaration, so it could never have passed.

Verified against autolens_workspace_test@main while preparing this: all five parity scripts carry the pre-should_simulate cap guard, point_source/visualization/modeling_visualization_jit.py now declares ENV: full_datasets real_output, and the solver's one-shot logger.warning plus updated Notes docstring are in place on PyAutoLens@main.

Verification

origin/main was merged in (it moved during this work); the three generated-file conflicts (complete/index.md, dashboard.md, dashboard.html) were resolved by regenerating with the owning tooling, not by hand-merging. Net diff against current main is unchanged by that merge — the same 8 files, no collateral.

lifecycle.py check          OK
lifecycle.py index --check  OK
lifecycle.py orphans        none
registry_toc.py --check     fresh

Follow-up (not in this PR)

origin/claude/autolens-point-jax-vmap-parity-htyfi1 is now fully superseded and should be deleted, so nobody merges the stale version.

---

Generated by Claude Code