gradient README kernel-CDF rows + rect-adapt/kernel alignment verification

[Jammy2211]

Summary

Documentation + verification leg of PyAutoLabs/PyAutoArray#373 (kernel-CDF meshes, library PR PyAutoLabs/PyAutoArray#374), also carrying the visual verification of the #372/#375 rect-adapt plot fix:

• jax_profiling/gradient/README.md: four kernel-CDF rows in the status table (imaging os_pix=1 / os_pix=4 / production shape, interferometer sparse — all strict-FD certified 2026-07-10 with measured FoM-parity numbers), staircase-implications section updated (the "continuous density transform" it anticipated now exists as #374's opt-in meshes; the linear meshes' staircase documentation stands), new kernel-mesh caveats (O(M×N) forward memory — ~60 GB at 15k px × os 4², chunking is follow-up work; measure-thin solver branch flips poisoning single-step FD and the FD-step-sweep answer), findings-log entry.
• plotting_alignment/kernel_cdf_alignment.py (new): end-to-end visual + numeric verification of the #372 edge fix through the real plot_inversion_reconstruction path for the linear AND kernel adaptive meshes — node-containment and mapper-faithfulness asserts (drawn == encoded to 0.01–0.07 local cells; the residual 0.09–0.24-cell displacement is the mapper's own bilinear interpolation, unasserted physics), full-extent reconstruction renders vs the simulator truth. Verification results posted on #372.
• plotting_alignment/simulator.py: repaired against the current API (al.output_to_fits, functional aplt output kwargs) — it previously crashed before writing anything; the auto-simulated dataset is now gitignored.

Test Plan

• python plotting_alignment/simulator.py — runs end-to-end, writes the dataset
• python plotting_alignment/kernel_cdf_alignment.py — all containment + faithfulness asserts pass on 4 adaptive meshes (Delaunay reference skipped: pre-existing #332 crash, reported in-script)
• README numbers match the certification logs (autolens_workspace_test PR)

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

• Effective level: supervised (header: supervised; human directed "continue --auto" in-session after plan approval + library-ship sign-off; rect-adapt verification human-directed mid-run)
• Plan: approved in-session and on issue #373; rect-adapt fold-in per explicit user instruction
• Gate: tests n/a (doc + verification scripts; both executed end-to-end) · smoke n/a (no downstream script surface) · review CLEAN (simulator-tail finding fixed in c3f0622) · 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 (and ideally #375).

🤖 Generated with Claude Code