docs: point-source gradient probe status flip (phase 5 of #657)

[Jammy2211]

Summary

Phase-5 companion (#657; library PR linked below): the point-source gradient probe's documented all-zeros finding flips on the production path. The probe prose now records the implicit-diff custom_jvp and the expected per-row pattern: steps 1–3 (hand-built closures over raw arrays) remain all-zero BY DESIGN — they bypass model registration and exercise the compatibility gate's unregistered fallback — while the full Fitness.call pipeline is PASS with non-zero gradients (measured: grad norm 17096). The regression guard is the full-pipeline row; a flip of that row back to all-zero means the implicit rule was disconnected. Also points at the workspace_test staircase-aware FD certification.

Scripts Changed

• jax_profiling/gradient/point_source/image_plane.py — "Finding" docstring section rewritten for the phase-5 status flip (expected per-step pattern + inverted full-pipeline guard); no code changes.

Upstream PR

PyAutoLabs/PyAutoLens#677

Test Plan

• Probe re-run against the library branch: solver-chained steps report PASS with non-zero gradients

Generated by the PyAutoLabs agent workflow.