fix: test-mode bypass tolerates a single-eval FitException (resample-equivalent)

[Jammy2211]

Summary

TEST_MODE=2's _fit_bypass_test_mode (autofit/non_linear/search/abstract_search.py) calls the likelihood once to verify the analysis runs. If that single prior-median instance is pathological — a non-positive-definite inversion, or a degenerate Delaunay mesh that yields NaN vertices — the analysis raises af.exc.FitException. That is exactly the signal a real non-linear search absorbs by resampling. In test mode there is no sampler, so the exception hard-failed the whole run.

This made flaky pixelization scripts intermittently break CI smoke. Concretely, autolens_workspace/scripts/interferometer/features/pixelization/delaunay.py fails ~50% on CI py3.12 with ValueError: Points cannot contain NaN from the Delaunay/qhull triangulation, via source_pix_2 → _fit_bypass_test_mode → log_likelihood_function → inversion → scipy_delaunay (autolens_workspace#307, PyAutoLens#640). The same workspace commit passed on a parallel run — a true nondeterministic (CI-thread/BLAS-dependent) flake.

Change

Wrap the single verification eval: on FitException, log the cause and keep the existing -1.0e99 sentinel (the same effect as a resample-to-reject). Only FitException is caught — genuine code errors (e.g. ValueError, AttributeError) still propagate and fail test mode, so a truly broken likelihood stays loud. Central fix: benefits every analysis (imaging, interferometer, galaxy), not just this one script. Complements PyAutoLens#639 (raise_fit_exception preserves the cause behind the FitException).

Tests

• TestBypassToleratesFitException::test__test_mode_2__fitexception_tolerated__fit_completes_with_sentinel — fit completes, sentinel log-likelihood retained.
• ...::test__test_mode_2__non_fitexception_still_propagates — a ValueError still raises.
• 108 non_linear/{search,result,samples} tests pass; no regressions.

API Changes

None. Behaviour change is confined to the TEST_MODE=2 bypass path; public API, signatures, and normal-search behaviour are unchanged. No downstream workspace migration required.

Follow-ups

• autolens_workspace: un-park the interferometer Delaunay no_run.yaml entry + add to smoke (library-first: merge this PR first).
• Phase 2 (PyAutoArray): fix the actual NaN/non-PD producer (candidate: fnnls.py:134 divide-by-zero; degenerate Hilbert-mesh vertices) so the fit is not merely tolerated but correct.

🤖 Generated with Claude Code

[Jammy2211]

Design validation (parity with existing autofit conventions). This isn't a bespoke workaround — FitException → resample sentinel is how autofit already handles pathological instances in three places, and this change makes the test-mode bypass consistent with them:

1. Real numpy fitness wrapper — fitness.py:255-256:

   except exc.FitException:
       return self.resample_figure_of_merit

   The live sampler already treats a FitException as a resample-to-reject. The bypass is the only likelihood call that didn't.

2. Sentinel value matches the searches these scripts use — Nautilus and Dynesty set resample_figure_of_merit=-1.0e99 (nautilus/search.py:194,215, dynesty/.../abstract.py:184). That is exactly the -1.0e99 the bypass already defaults to (abstract_search.py:872) and that this fix retains — so a resample-rejected instance produces the identical sentinel in test mode and in a real Nautilus fit.

3. Initializer — initializer.py:35,39 treats both FitException and figure_of_merit < -1e98 as "invalid, resample" (return None), and its own test-mode bypass uses -1.0e99 (initializer.py:233).

Scope is deliberately narrow (only FitException, matching fitness.py) so genuine code errors still fail test mode loudly. Complements PyAutoLens#639 (raise_fit_exception preserves the cause), whose body explicitly flagged the "single evaluation not absorbed by a sampler under PYAUTO_TEST_MODE" problem but scoped the tolerance out — this PR completes it.

CI: unittest 3.12 + 3.13 green, docs green.