bug: record the shared_preloads.py stall that refutes its 08-24 retime

[Jammy2211]

Closes #328.

Summary

multi_dataset/jax_likelihood/shared_preloads.py hit TIMEOUT (300s) in PyAutoHeart
Workspace Smoke run 32902243623 (job 97978549465, 2026-08-25) while all nine
siblings in the same job passed in 8.6–49.3s.

The point is not the tally mark. The 2026-08-24 retime removed this entry's SLOW
marker and returned it to coverage; it stalled on the next weekly run.

Why this is more than another occurrence

Verified 2026-08-25 against autolens_workspace_test at 7fc497d:

List	State of shared_preloads.py	Effect
config/build/no_run.yaml	absent — SLOW marker removed 2026-08-24	in mega-run / weekly coverage
smoke_tests.txt line 15	still commented out (2026-07-22, "the heaviest entry here")	out of the PR gate

The retime (5 repeats per Python leg, 300s cap; runs 32741371308 + 32741386752)
removed eight PyAutoHeart#74 entries because they "completed 5/5 on both legs
(slowest 54.5s) — marker refuted".

~10 executions at ≤54.5s cannot exclude a low-probability hang. The retime was
right that this script is not slow; it was not powered to see that it stalls.
This lands § Task step 1's slow-vs-stall question on the opposite side from where
the retime put this entry — and the same uncertainty attaches to every other entry
that sweep readmitted.

What changed

• draft/bug/ci/jax_vmap_jit_compile_stall.md — a ## New occurrence — 2026-08-25
  section mirroring the existing 2026-08-24 precedent: run/job ids, the ten-row
  sibling timing table, the retime refutation, the exclusion-list disagreement, and
  the affected-set consequence.
• epics.md — the jax-compile-stall entry gains a NEW EVIDENCE 2026-08-25 note
  so the epic's state carries it.
• Dashboard regenerated. lifecycle check: OK.

Two findings worth flagging to a reader of the epic

The channel is new. Every prior occurrence in the ledger comes from a workspace
repo's own Smoke Tests gate or from release-integrate. This is the first via
PyAutoHeart's weekly workspace-validation.yml sweep — a different, PyAutoHeart-owned,
script_matrix.py-driven harness that honours no_run.yaml and not smoke_tests.txt.
Cross-harness corroboration that the stall is not one repo's runner configuration.

The common factor has widened. shared_preloads.py is not an MGE variant, so the
ledger's "composite MGE vmap graph" shape no longer covers the whole affected set.
What the members share is heavy multi_dataset vmap composition, not MGE specifically.

Deliberately not done

Nothing is parked, disabled, or re-marked. The ledger's own § "This is why
quarantining is the wrong end state" argues against whack-a-mole parking that strips
coverage of exactly the heaviest JAX paths, and the 2026-08-24 precedent for
multi_dataset/jax_likelihood/mge.py was to record and deliberately not park.
Re-marking is phase 3's call with a root cause in hand.

autolens_workspace_test was read for evidence only and is not modified by this PR.

Follow-up noticed

pyauto-brain bug returns owner: unresolved / fix locus: unresolved for a prompt
whose only repo is @PyAutoMind, because Mind is not in its library-repo owner map.
The classification was otherwise sound. Not fixed here.

---

Generated by Claude Code