mind: retire jax-likelihood-smoke-pins-stale — fixed before it was filed

complete/2026/08/jax-likelihood-smoke-pins-stale.md

@@ -0,0 +1,130 @@
+# The jax_likelihood smoke-pin bug was already fixed when it was filed
+
+Retired unstarted on 2026-08-26 by a `/start_dev` research pass. Never issued
+as a GitHub issue and never developed: the four failures the prompt describes
+had all been resolved on `autolens_workspace_test` main by 2026-08-22, most of
+them within hours of the prompt being written. Recorded rather than deleted
+because the *reason* it went stale is the reusable finding.
+
+## Verified state at retirement
+
+`autolens_workspace_test` main `8be3d59`, Smoke Tests run
+[32922456742](https://github.com/PyAutoLabs/autolens_workspace_test/actions/runs/32922456742)
+(2026-08-26 02:24 UTC), green on both Python 3.12 and 3.13:
+
+```
+Running 22 listed scripts
+  scripts/imaging/jax_likelihood/lp.py ...              PASS (9.8s)
+  scripts/interferometer/jax_likelihood/rectangular.py  PASS (21.1s)
+  scripts/interferometer/jax_likelihood/mge.py ...      PASS (12.6s)
+  ... 22/22 PASS
+```
+
+The prompt's headline — "every workspace_test smoke sweep reads 20/24 until
+this lands" — is now 22/22. The denominator moved too: two further entries
+left `smoke_tests.txt` for reasons unrelated to pins.
+
+## Disposition of each claim
+
+- **`imaging/jax_likelihood/lp.py`, pin `-1.34797827e09` is double.** Fixed.
+  Repinned to `-6.74165366e08` in `197ce6e` (autolens_workspace_test#257) on
+  2026-08-19 — the same day the prompt was filed.
+- **"The unmerged fix `fb1aefe0b` sits in the orphan worktree
+  `~/Code/PyAutoLabs-wt/positions-lh-penalty-accumulation`; landing it and
+  re-pinning must happen together."** Both had already happened. `fb1aefe0b`
+  is the head SHA of [PyAutoLens#700](https://github.com/PyAutoLabs/PyAutoLens/pull/700)
+  ("sum PositionsLH penalties instead of doubling the last entry", fixes
+  PyAutoLens#699), merged 2026-08-17 — two days *before* the prompt called it
+  unmerged. `197ce6e` is precisely the paired re-pin the prompt asked for, and
+  it repinned five scripts, not four (it also caught
+  `interferometer/jax_likelihood/lp.py` and both `imaging/substructure/subhalo.py`
+  pairs).
+- **`interferometer/jax_likelihood/rectangular.py` misses pin `-3164.286252` by
+  0.0124% vs rtol 1e-4.** Passing. The pin literal is *unchanged* to this day,
+  and `99d63b3` (2026-08-21) re-verified the script green including its
+  TransformerNUFFT cross-check. So the 08-19 miss was environment drift in the
+  control run, not a stale pin — nothing to recalibrate and no tolerance to
+  widen.
+- **`interferometer/jax_likelihood/mge.py`, "same stale-pin family".** Fixed by
+  the same `197ce6e`: `-7.94439429e08` -> `-3.97221282e08`, i.e. it was a
+  PositionsLH-doubling casualty, not an independent drift.
+- **`multi_dataset/jax_likelihood/mge.py`, "same stale-pin family".** Never a
+  pin problem. Commented out of `smoke_tests.txt` on 2026-08-22 (#262) for
+  hanging to the 300s cap — the family-wide XLA compile stall tracked in
+  autolens_workspace_test#245. Its pin (`EXPECTED_VMAP_LOG_LIKELIHOOD =
+  -2173221.43685875`) is untouched since the `multi` -> `multi_dataset` rename
+  (`d7cc60c`) and has never been shown wrong.
+
+## Key traps / findings
+
+- **A control-test finding has a shelf life measured in hours.** This prompt
+  was written from a control run taken during `lazy-heavy-imports` (#1505) and
+  was obsolete before the day was out — `197ce6e` landed 2026-08-19 10:25 EDT.
+  Filing a finding is not the same as checking whether someone is already on
+  it; a pass over recent commits touching the named files would have caught it.
+- **"Orphan worktree" is a claim about a laptop, not about a repository.** A
+  local worktree whose branch has been merged and deleted upstream looks
+  identical to one holding genuinely unlanded work. The check that settles it
+  is the branch's head SHA against the remote's merged PRs — here `fb1aefe0b`
+  resolved straight to a PR merged two days earlier. Adjudicate an orphan from
+  the remote, never from the worktree's own state.
+- **One root cause can wear several symptom labels.** Three of the four listed
+  scripts were the *same* bug (the 2x-last PositionsLH penalty), not a
+  "stale-pin family". Grouping symptoms by folder invented a class of drift
+  that did not exist and would have sent the fix in the wrong direction —
+  widening tolerances on correct pins.
+- **A script leaving the smoke gate changes the denominator.** A pass/total
+  target like "20/24" is not a stable goal: `multi_dataset/jax_likelihood/mge.py`
+  and `multi_dataset/jax_likelihood/shared_preloads.py` both left the list for
+  timing reasons, so the sweep now reads 22/22 without any pin having been
+  touched for that.
+
+## Follow-ups
+
+- **Local debris only.** `~/Code/PyAutoLabs-wt/positions-lh-penalty-accumulation`
+  is a stale worktree over a merged-and-deleted branch on the laptop; it is a
+  `/repo_cleanup` item, not code work, and no cloud session can see or clear
+  it.
+- **Latent, owned elsewhere.** `multi_dataset/jax_likelihood/mge.py`'s pin has
+  never been verified against a post-PyAutoLens#700 library. If
+  autolens_workspace_test#245's XLA stall is fixed and the script returns to
+  the gate, re-derive its pin as part of that re-enable rather than trusting
+  the pre-rename literal. Deliberately not filed as its own prompt — it has no
+  standalone trigger.
+
+## Original prompt
+
+# autolens_workspace_test jax_likelihood pins: 4 scripts fail smoke on main
+
+Type: bug
+Target: autolens
+Repos:
+- @PyAutoLens
+Difficulty: low
+Autonomy: supervised
+Priority: normal
+Status: draft
+Filed: 2026-08-19 (backfilled from git)
+
+## Finding (2026-08-19, control-tested during lazy-heavy-imports #1505)
+
+`pyauto-heart smoke autolens_test` fails 4/24 on canonical main (verified twice,
+worktree and canonical identical):
+
+- `imaging/jax_likelihood/lp.py` — vmap result is ~half the hardcoded pin
+  `-1.34797827e09`. This is the positions-LH penalty-doubling: the unmerged fix
+  `fb1aefe0b` ("sum PositionsLH penalties instead of doubling the last entry")
+  sits in the orphan worktree `~/Code/PyAutoLabs-wt/positions-lh-penalty-accumulation`
+  (PyAutoLens). The pin was calibrated against the doubling behaviour — landing
+  that fix and re-pinning must happen together.
+- `interferometer/jax_likelihood/rectangular.py` — misses pin `-3164.286252` by
+  0.0124% vs rtol 1e-4 (stale-pin drift).
+- `interferometer/jax_likelihood/mge.py`, `multi_dataset/jax_likelihood/mge.py` —
+  same stale-pin family.
+
+## Task
+
+Adjudicate the orphan positions-LH worktree (land or discard `fb1aefe0b`), then
+recalibrate the four scripts' hardcoded pins against current main and widen
+tolerances where the value is legitimately environment-sensitive. Every
+workspace_test smoke sweep reads 20/24 until this lands.

complete/index.md

@@ -6,7 +6,7 @@ Token-light navigation over the finished-work records (schema:
 only then grep a dated bucket. Curators: edit the band between the CURATED
 markers; everything below GENERATED is rebuilt.
 
-1142 records across 7 buckets.
+1143 records across 7 buckets.
 
 <!-- CURATED:START -->
 ## Highlights
@@ -122,6 +122,7 @@ _(curate hard-won records here — survives regeneration.)_
 - [jax-grad-local-vs-ci-assertions](2026/08/jax-grad-local-vs-ci-assertions.md)
 - [jax-grad-smoke-timeout-budget](2026/08/jax-grad-smoke-timeout-budget.md)
 - [jax-likelihood-mass-sensitivity](2026/08/jax-likelihood-mass-sensitivity.md) — Audit + fix for median-literal mass blindness across all jax_likelihood/datacube scripts (phase 2 of the findi…
+- [jax-likelihood-smoke-pins-stale](2026/08/jax-likelihood-smoke-pins-stale.md)
 - [jax-stall-shared-preloads-retime-refutation](2026/08/jax-stall-shared-preloads-retime-refutation.md)
 - [knowledge-board](2026/08/knowledge-board.md) — auto-closed on merge
 - [lacosmic-cr-option-and-star-pass-decoupling](2026/08/lacosmic-cr-option-and-star-pass-decoupling.md) — both deliberately OPEN — closure is the default-flip decision, not this ship