Work Classification

Both: PyAutoArray library repair first, then autolens_workspace_test parity coverage.

Affected Repositories

• PyAutoArray (primary)
• autolens_workspace_test (downstream parity assertions)

Branch Survey

The isolated task worktree avoids the unrelated workspace-test changes.

Suggested branch: feature/delaunay-nan-callback

Worktree root: ~/Code/PyAutoLabs-wt/delaunay-nan-callback/

Implementation Steps

1. In autoarray/inversion/mesh/interpolator/delaunay.py, add a host-side np.isfinite(points_np).all() guard to scipy_delaunay_tri_only. For invalid points, return -1-filled int32 arrays at the documented (2N, 3), (2N, 3), and (N,) shapes without calling qhull. Because _jax_delaunay_tables uses vmap_method="sequential", the check remains lane-local.
2. In pix_indexes_delaunay_walk_from, track whether the current simplex row contains only non-negative vertex indices. An invalid/padded row must not set done; it must terminate through the existing [seed, -1, -1] fallback. Apply the invariant to both primary query mapping and jax_delaunay split-point mapping.
3. In pixel_weights_delaunay_from, after barycentric/nearest-neighbour selection, use the threaded xp backend to make JAX weights NaN when any mesh vertex is non-finite. Keep the eager SciPy/NumPy production route unchanged for its existing exception-based diagnostics. The gate must be a bitwise no-op for finite meshes.
4. Add PyAutoArray unit coverage for sentinel table shapes/dtypes, invalid-simplex fallback mappings with at least one non-negative index per query, and finite-input no-op behavior. Cover an all-NaN mesh, a non-last poisoned vertex, and the last poisoned vertex to prevent negative-index IEEE-order laundering.
5. Update scripts/imaging/jax_likelihood/delaunay.py and scripts/interferometer/jax_likelihood/delaunay.py with production Fitness._vmap assertions: the poisoned lane produces a NaN FOM without raising, and finite lanes equal their unbatched/baseline results bit-for-bit. This covers dense and sparse inversion paths and the linked intermittent interferometer qhull failure surface.
6. Run focused PyAutoArray tests and the full PyAutoArray suite. Run the two likelihood parity scripts and scripts/imaging/jax_grad/delaunay.py unchanged; confirm existing finite likelihood literals and frozen-table FD gradients remain unchanged.

Key Files

• autoarray/inversion/mesh/interpolator/delaunay.py — callback, visibility walk, split mappings, and final interpolation weights.
• test_autoarray/inversion/pixelization/interpolator/test_delaunay_walk.py — focused NumPy-side geometry invariants.
• scripts/imaging/jax_likelihood/delaunay.py — dense production-vmap regression.
• scripts/interferometer/jax_likelihood/delaunay.py — sparse production-vmap regression.
• scripts/imaging/jax_grad/delaunay.py — unchanged frozen-table gradient certification.

Acceptance Invariants

• No non-finite callback lane calls qhull or aborts sibling lanes.
• Every invalid callback mapping row retains at least one non-negative index, so NaN weights remain a live mapping contribution.
• Poisoned dense and sparse lanes reach NaN at the figure of merit.
• Finite lanes are bit-identical to solo/baseline evaluation and finite-input gradients remain FD-certified.
• No NumPy production exception contract is changed.

Delaunay tables callback crashes the whole JAX run when traced points contain NaN

Type: bug
Target: autoarray
Repos:

• PyAutoArray
  Difficulty: medium
  Autonomy: supervised
  Priority: high
  Status: draft

The bug (hit live on RAL, 2026-07-27, wsdev#117 campaign)

Under broad-start multi-start gradient descent (af.MultiStartProdigy, resurrect=True), a start's parameter vector can go non-finite mid-trajectory. For pure-JAX meshes that propagates through the likelihood for the search machinery to handle. For Delaunay, non-finite traced mesh points reach host qhull through _jax_delaunay_tables; qhull raises ValueError: Points cannot contain NaN, surfaced as a JaxRuntimeError, and one poisoned lane aborts all 16 starts.

The same crash occurred with resurrect=False, showing that ordinary broad-start descent can create the invalid mass configurations. Persisted NaN parameters also make dependent/resumed jobs fail immediately.

Fix sketch

Detect non-finite input inside the sequential callback lane and return fixed-shape sentinel tables without calling qhull. Ensure the downstream likelihood evaluates to NaN for that lane rather than laundering the sentinel result into finite weights. Validate via batched JAX likelihood evaluation with poisoned and finite lanes, plus the existing frozen-table finite-difference gradient certification.

Related: the intermittent interferometer Delaunay qhull NaN is likely the same missing guard on the sparse path.