Park multi_dataset/jax_likelihood/mge.py out of the smoke gate - #262

Summary

multi_dataset/jax_likelihood/mge.py hangs to the 300s smoke cap, blocking unrelated PRs. This comments it out of smoke_tests.txt with the reason recorded inline, matching how its two siblings were already handled.

This is a new instance of a known, documented bug — not a new diagnosis. config/build/no_run.yaml:87 already quarantines its sibling:

multi_dataset/jax_likelihood/delaunay.py # NEEDS_FIX 2026-08-01 - hangs to the 1800s release cap
in 3 of 5 release-integrate runs since 07-31 (18s when it passes; intermittent XLA compile stall,
family-wide - ag_test rectangular.py also hit); quarantined ...; see autolens_workspace_test#245

"18s when it passes; intermittent XLA compile stall, family-wide" is precisely the signature here — same folder, same shape. mge.py was the last multi_dataset/jax_likelihood/ entry still in the PR gate; shared_preloads.py was disabled in this file on 2026-07-22.

Evidence

• 4/4 job failures on Drop the pynufft legs from the NUFFT parity script #261 — Python 3.12 and 3.13, twice each including a re-run — always TIMEOUT (300s) multi_dataset/jax_likelihood/mge.py, always 23/24 passed.
• Not caused by Drop the pynufft legs from the NUFFT parity script #261. That PR changes three files (interferometer/nufft.py, interferometer/jax_likelihood/rectangular_sparse.py, scripts/CLAUDE.md); none is in smoke_tests.txt, so this gate never executes them. mge.py contains no reference to transformers, NUFFT or pynufft.
• Not cap margin. Run standalone under its declared ENV: jax full_datasets profile it completes in 32s against a 300s budget, simulating its dataset from scratch exactly as CI does. It stops making progress rather than running long.
• Not dataset contamination from a neighbour. mge.py is the only script in the gate that touches dataset/multi_dataset/lens_sersic, and the runner executes each script as its own subprocess, so no in-process state carries over.

What this deliberately does not do

No config/build/no_run.yaml entry. Every failure observed is at the 300s smoke cap. I have no evidence about behaviour at the 1800s release cap, and quarantining it there on a guess would drop real release coverage for a script that runs in 32s. The release side is untouched.

No attempt to fix the stall. That is #245's job and is family-wide (it has also hit ag_test rectangular.py), not something to solve inside a gate-list change.

Test Plan

• Only smoke_tests.txt changes — one line commented out, no script or config edits
• The remaining 23 gate entries are unchanged
• Smoke gate goes green on this PR, confirming mge.py was the sole failure

Generated by Claude Code