fix: park EP scripts as NEEDS_FIX no_run for release testing

[Jammy2211]

Why

EP is not release-ready. The 2026-08-03 nightly release integrate failed on
scripts/graphical/ep.py (PyAutoBrain run 30787947298 -> PyAutoHeart run
30788224561), reported as "1 failed: autofit_test graphical/ep.py".

This parks the EP scripts rather than papering over the instability in one
script's search settings.

Root cause (reproduced locally under the release profile)

The cloud message "The log_likelihood_function is always returning nan
values" is misleading — it is one of three causes listed in a single
InitializerException. The real trigger is the "all initial samples have the
same figure of merit" branch. No NaN is involved.

1. ep.py truncates its per-factor search (maxcall=1000, maxiter=1000).
   Probing one such search directly gives an effective sample size of 1 — a
   single sample carries all the weight, so the weighted std of every
   parameter is exactly 0. An uncapped search on the same data gives ESS
   ≈ 1420 and healthy stds ≈ 0.13.
2. EP projects that degenerate posterior into a message and feeds it back as
   the next cycle's prior. Instrumenting the EP update shows q* arriving
   already collapsed (q*_sigma=1.5e-06 against a healthy
   cavity_sigma=30), so the collapse enters at the search→message projection,
   not the cavity division.
3. The next search then draws initial points that all yield an identical log
   likelihood, and PyAutoFit raises InitializerException.

Both failure faces were captured: sigma collapsing to delta functions, and the
shared centre diverging to 6.4e7 (data spans ~100 pixels) so the Gaussian
falls entirely outside the data — giving a figure-of-merit spread of exactly
0.0.

It is intermittent (it depends on where the truncated search happens to stop),
which is why graphical passed the five previous nightly runs. The EP source
has not changed in weeks — this is a long-standing flake that finally bit, not
a regression.

Two things that do not fix it

• Damping. At delta=0.5 a prior still collapsed in a single step —
  q* is already degenerate before damping applies.
• check_sigma_collapse. It looks for monotone shrinkage or std < 1e-8;
  this is a one-step drop to ~1e-7, so it reports (none) even mid-collapse.

See PyAutoFit #1332 F10.

Change

Four NEEDS_FIX-tagged entries in config/build/no_run.yaml, so every mega-run
surfaces them with a warning banner until EP is fixed. Sibling PRs park the
remaining two EP scripts in the release matrix (autofit_workspace
features/expectation_propagation.py, autolens_workspace
guides/modeling/advanced/expectation_propagation.py). HowToFit's two EP
tutorials are not in the release matrix and need no entry.

Verification

• Patterns are .py-anchored and checked against the runner's own
  should_skip: exactly 4 scripts newly skipped here, no over-match, no
  inert entries. (An unanchored graphical/ep would have silently swallowed all
  four via substring match — anchoring keeps each reason attached to its script.)
• All entries parse via find_needs_fix_skips with correct dates and reasons.
• The exact CI command for the failing shard now exits 0:
  run_python.py autofit_test scripts/graphical --env-config config/build/profile_release.yaml
  → 4 skipped with reasons, 3 passed.

🤖 Generated with Claude Code