fix: isolate invalid Delaunay callback lanes - #411

Summary

• isolate non-finite Delaunay mesh points inside each sequential JAX callback lane, avoiding a qhull exception that aborts the whole vmapped batch
• reject padded sentinel simplices in the visibility walk so invalid meshes retain a live nearest-vertex mapping instead of laundering into finite pixel-0 weights
• propagate invalid JAX meshes as NaN pixel weights while leaving the eager SciPy/NumPy exception path unchanged

Closes #410.

API Changes

None — internal changes only.

Test Plan

• python -m pytest -x — 906 passed
• imaging Delaunay production _vmap: poisoned lane resampled, finite sibling lanes bit-identical to the clean vmapped batch, raw fit log-likelihood NaN
• interferometer Delaunay production _vmap: same assertions through the sparse inversion path
• partial non-last vertex, last vertex, and all-NaN mesh poisoning under jit(vmap(value_and_grad)); finite-lane values and nonzero gradients bit-identical
• unchanged scripts/imaging/jax_grad/delaunay.py FD certification passed (rtol=1e-2)
• autolens_workspace_test curated smoke: 20/21 passed; imaging/subhalo_recovery.py exceeded its existing 300-second cap and was human-accepted as unrelated/non-causal

Validation checklist (--auto run)

• Effective level: supervised (bug work-type cap); human approved the design and explicitly released the ship checkpoint
• Plan: recorded on issue fix: isolate non-finite Delaunay callback lanes #410 before implementation
• Gate: tests 906 passed · task-specific dense/sparse parity passed · scoped smoke 20/21 with unrelated timeout human-accepted · review CLEAN · Heart YELLOW exact reason set human-acknowledged
• Human: diff matches the approved plan without scope creep
• Human: merge, amend, or reject

Generated by the PyAutoLabs agent workflow.