jax_grad: certify kernel-CDF meshes (strict FD on all params everywhere)

[Jammy2211]

Summary

Certification leg of PyAutoLabs/PyAutoArray#373 (kernel-CDF meshes, library PR PyAutoLabs/PyAutoArray#374): extends the jax_grad gradient-audit scripts with the RectangularKernelAdaptDensity / RectangularKernelAdaptImage variants and asserts the task's success criterion — strict FD on ALL parameters in every configuration, including the two corners where the linear adaptive mesh has zero usable gradients (imaging at pixelization over-sampling 1; the interferometer sparse-operator path).

Both scripts pass end-to-end against the library branch (incl. the #372/#375 rect-adapt edge fix):

• imaging_pixelization.py variants E/F/G — strict FD (default tolerances, no skips) on all 14 params at os_pix=1 (bandwidth=0.1) and os_pix=4 (default bw; full production shape at bw=0.1), mass/shear-live guards, eager-vs-JIT, FoM parity: no-degradation at os_pix=1 (kernel beats linear by 4.1e-2), strict 2.7e-5 (F) and 6.3e-4 ≤ 1e-3 (G).
• interferometer.py variant D — sparse production shape, strict FD on all 7 mass/shear params, eager-vs-JIT, FoM parity 3.9e-5. Existing variants A–D / A–C unchanged and still passing (staircase regressions intact).
• util.py — new opt-in FD-step-sweep mode (rel_steps): single FD steps are pseudo-randomly poisoned by measure-thin solver branch flips (width < 1e-15 in the parameter, ΔLL ~1.6e-3–14, present under reg.Constant and reg.Adapt; pre-existing, probed 2026-07-10 — full evidence on #373). Per parameter the sweep compares AD at the closest of {1e-8, 1e-7, 1e-6} and prints the whole matrix; clean steps converge to AD at 1e-6..1e-9 relative, so a wrong AD fails every step. Existing callers (rel_step) unchanged.

Test Plan

• python scripts/jax_grad/imaging_pixelization.py — all 7 variants pass
• python scripts/jax_grad/interferometer.py — all 5 variants pass
• Run against PyAutoArray feature/rectangular-kernel-cdf-mesh @ 6df3ebf2 (kernel meshes + rect-adapt merge; suite 893 passed)

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

• Effective level: supervised (header: supervised, cap: feature → safe≤medium; human directed "continue --auto" in-session after plan approval + library-ship sign-off)
• Plan: approved in-session (Plan Mode) and on issue #373; deviations disclosed in the batched question on #373 (FD-step-sweep design; os_pix=1 parity → no-degradation; variant-G 1e-3 limit at measured 6.3e-4 floor)
• Gate: tests n/a (workspace repo — the changed scripts ARE the validation, both pass end-to-end; upstream PyAutoArray suite 893 green) · smoke n/a (jax_grad scripts have no downstream script surface) · review CLEAN · Heart YELLOW, reason set identical to the launch-acked 6-reason set
• Human: plan sound in hindsight?
• Human: diff matches plan (no scope creep)?
• Human: merge, amend, or reject — then log the outcome

Merge gate: library-first — do not merge before PyAutoLabs/PyAutoArray#374.

🤖 Generated with Claude Code