fix(interferometer): resample non-PD inversion instead of crashing the search

[Jammy2211]

Summary

Fixes a hard crash where the interferometer (and point-source) analyses kill the whole non-linear search when a sampled model produces an invalid inversion.

AnalysisInterferometer.log_likelihood_function returned the fit's figure_of_merit unwrapped, so a non-positive-definite inversion matrix at a sampled model raised a raw numpy.linalg.LinAlgError: Matrix is not positive definite (Cholesky log-det evidence term, PyAutoArray/.../abstract.py:743) that propagated through Nautilus's multiprocessing pool and killed the search (~38 s) on the NumPy path (release profile PYAUTO_DISABLE_JAX=1). The JAX path returns NaN → resampled, silently masking it. This is not jax-0.10.2 numerical drift (the original hypothesis) — it is a NumPy/JAX backend divergence.

The imaging analysis already guards this (imaging/model/analysis.py:132-144): split on self._use_jax, wrap the NumPy path in try/except → af.exc.FitException, which autofit/non_linear/fitness.py:235 catches and resamples. This PR mirrors that guard in the interferometer analysis, and applies the same guard to the point-source analysis, which had the identical unguarded return. Both method docstrings already promised this discard-on-exception behaviour.

Resolves release-validation tail item G (PyAutoHeart#72). Refs #606 (issue kept open pending corrective validation — fresh wheels + release-integration rerun + new Heart verdict).

API Changes

None — internal behaviour change only. log_likelihood_function keeps its signature; on the NumPy backend it now raises af.exc.FitException (a resample signal the search already handles) instead of propagating a raw LinAlgError for a failed inversion. The JAX branch is byte-for-byte unchanged.

Test Plan

• Repro: NumPy-path autolens_workspace_test/scripts/interferometer/model_fit.py (PYAUTO_DISABLE_JAX=1 JAX_ENABLE_X64=True, real search) — completes (bad point resampled) vs. LinAlgError crash before the fix.
• JAX-path interferometer/model_fit.py (no disable) — still completes (unchanged branch).
• pytest test_autolens/interferometer/ — 22 passed.
• pytest test_autolens/point/ — 58 passed.
• pytest test_autolens/ — full suite: 381 passed.

Heart-RED corrective-PR exception (human-authorized)

Opened under the AUTONOMY.md corrective-PR exception while Heart is RED.

• RED reason cleared (contributor): release validation FAILED (stage integrate) — this PR repairs the interferometer model_fit.py stage-integrate crash (tail item G).
• Authorization: contemporaneous human authorization on issue fix(interferometer): resample non-PD inversion instead of crashing the search #606 quoting the exact reason (2026-07-14).
• Causal map: Heart RED release validation FAILED (stage integrate) → fix(interferometer): resample non-PD inversion instead of crashing the search #606 diagnosis (NumPy-path unguarded Cholesky log-det crash) → NumPy-path try/except → FitException guard mirroring imaging → this PR.
• Sibling RED reasons NOT addressed (Heart stays RED, release stays blocked until each clears by its own path): workspace validation not passing (3 failed, 2026-07-09), 58 stale parked script(s), PyAutoMind: open PR 14d old, install verification not run.
• Scope: commit/push/open-one-PR only. No merge, no issue close, no release, no unrelated changes. Diff is exactly the two-file analysis guard.

Validation checklist (--auto run — corrective-red)

• Effective level: supervised (header supervised, cap bug → supervised); shipped via the human-authorized Heart-RED corrective-PR exception, not the standard autonomous gate.
• Plan: on issue fix(interferometer): resample non-PD inversion instead of crashing the search #606, written at start, unmodified since.
• Gate: tests = 381 passed (full test_autolens/) incl. 22 interferometer + 58 point · smoke = the NumPy-path model_fit.py real-search repro (now completes) · review self-CLEAN (two files, mirrors imaging) · Heart RED — corrective exception, reason quoted above.
• Human: diff matches plan (no scope creep beyond the two-file guard)?
• Human: merge decision (merge stays a human act; release stays blocked while Heart RED).

Generated by the PyAutoLabs agent workflow.