bug: jit/imaging/delaunay.py — log_evidence rebuild returns -inf vs FitImaging's finite 26288.32

[Jammy2211]

Summary

On clean main (autolens_workspace_developer@4e99bf9, PyAutoLens 2026.5.14.2), jax_profiling/jit/imaging/delaunay.py crashes its self-consistency check at the log_evidence step: both the step-by-step rebuild and the inv-matrices rebuild return -inf, while the eager FitImaging.log_evidence reference returns a finite, expected 26288.32. The bug is in the rebuild, not in FitImaging.

Surfaced by the autolens_profiling Phase 1 follow-up smoke (issue #67); split out of that triage because it's an independent numerical / underflow bug not related to the point_source regression drift.

Reproduction

cd autolens_workspace_developer
source ../activate.sh
python jax_profiling/jit/imaging/delaunay.py

Output (relevant tail)

--- Step 13: Mapped reconstruction + log evidence ---
  [log_evidence_eager] 0.5566 s
  [log_evidence_jit_lower] 0.0364 s
  [log_evidence_jit_compile] 0.2256 s
  [log_evidence_jit_first_call] 0.1473 s
  [log_evidence_jit_steady_x10] 1.2660 s
    -> per-call avg: 0.126602 s
  log_evidence (step-by-step) = -inf
  log_evidence (inv matrices) = -inf
  log_evidence (reference)    = 26288.321397232066
Traceback (most recent call last):
  File ".../jax_profiling/jit/imaging/delaunay.py", line 833, in <module>
    np.testing.assert_allclose(
AssertionError:
Not equal to tolerance rtol=0.0001, atol=0
Log_evidence from inversion matrices does not match FitImaging.log_evidence
-inf location mismatch:
 ACTUAL: array(-inf)
 DESIRED: array(26288.321397)

Diagnosis hint

• FitImaging.log_evidence (the high-level path) produces a sensible finite number.
• Both the step-by-step rebuild and the inv-matrices rebuild produce -inf.
• That means the bug is in one of the building-block computations the script calls directly (regularization-determinant term, NNLS-residual term, or curvature-determinant term), not in FitImaging itself.
• The earlier steps (data vector, curvature matrix, regularization matrix, reconstruction, mapped reconstruction) all produce finite results with sensible shapes — only the final log_evidence aggregation goes to -inf.

Most likely culprits:

• log / slogdet of a near-singular matrix returning -inf from the JAX path under xp=jnp.
• Negative argument to xp.log somewhere in the Bayesian-evidence aggregation.
• The same expression on the eager xp=np path is benefiting from numpy's slight numerical-noise tolerance.

Context

• All other JIT scripts on clean main behave consistently:
  • imaging/mge.py, interferometer/{mge,pixelization,delaunay}.py, datacube/delaunay.py — PASS.
  • imaging/pixelization.py — crashes with mapping-matrix shape mismatch (separate issue).
  • point_source/{image_plane,source_plane}.py — drift in regression constants (triaged in triage: drift in jit/ regression constants (autolens_profiling F1) #67).

interferometer/delaunay.py passes on the same Delaunay pixelization with the same regularization (ConstantSplit), so the imaging-specific Delaunay log_evidence rebuild is the narrow culprit.

Related

• Triage parent: triage: drift in jit/ regression constants (autolens_profiling F1) #67 (autolens_profiling F1 regression-constant drift)
• Phase 1 mirror PR: feat: Phase 1 — mirror JIT likelihood profiling scripts autolens_profiling#2