test: document JAX finite-difference branch flip

[Jammy2211]

Summary

• identify parameter 9 as galaxies.lens.mass.einstein_radius in the RectangularKernelAdaptDensity (os_pix=1, bandwidth=0.1) gradient check
• confirm JAX 0.9.2 and 0.10.2 return the same autodiff value, ruling out a JAX 0.10.2 autodiff regression
• document the JAX 0.10.2 finite-difference branch flip and exclude only that named comparison while retaining its printed measurement and every other gradient, liveness, and FoM assertion

Scripts Changed

• scripts/jax_grad/imaging_pixelization.py — resolve the FD-unreliable parameter by name for the affected os_pix=1 variant, document the cross-version/ULP evidence, and report the exclusion explicitly
• scripts/jax_grad/util.py — clarify that skip_indices covers documented unreliable comparisons, including finite differences that cross a discontinuity

Corrective-PR exception

• Exact authorized Heart RED reason: release validation FAILED (stage integrate)
• Authorization record: fix: investigate JAX gradient mismatch at parameter 9 #164 (comment)
• Causal mapping: Heart reason → release-integration jax_grad failure → issue fix: investigate JAX gradient mismatch at parameter 9 #164 plan → this two-script, parameter-specific diff
• Remaining sibling RED reason, not addressed here: PyAutoFit: 1 commit(s) behind origin
• This exception permits this one pending-release PR only. Merge, issue close, release, and release rehearsal remain forbidden while Heart is RED.

Test Plan

• Run scripts/jax_grad/imaging_pixelization.py end-to-end with JAX/JAXLIB 0.10.2 (pass; documented parameter-9 FD exclusion exercised)
• Run the same imaging script end-to-end with isolated JAX/JAXLIB 0.9.2 (pass; autodiff value matches 0.10.2)
• Run sibling scripts/jax_grad/interferometer.py with JAX/JAXLIB 0.10.2 (pass)
• Run all six workspace smoke sets (52 passed, 0 failed, 3 declared skips)
• Compile both changed scripts and run git diff --check
• Independent review verdict: CLEAN

Validation checklist (--auto run — plan was not pre-approved)

• Effective level: supervised (header: supervised, cap: bug → supervised)
• Plan: on issue fix: investigate JAX gradient mismatch at parameter 9 #164, written at start and unchanged in scope
• Gate: tests n/a (workspace has no unit-test directory; exact regression scripts passed) · smoke 52 passed / 0 failed / 3 declared skips · review CLEAN · Heart RED with human-authorized corrective exception for release validation FAILED (stage integrate)
• Human: plan sound in hindsight?
• Human: diff matches plan (no scope creep)?
• Human: merge, amend, or reject — then log the outcome

Generated by the PyAutoLabs agent workflow.