[Jammy2211]

Summary

Add the shared point-source image-plane likelihood breakdown required before the CPU and A100 optimization campaigns. The instrument opens all eight PointSolver refinements into cumulative ray-trace, containment, selection, neighbourhood, and up-sampling boundaries while retaining the fused production likelihood as the authoritative runtime.

The committed CPU fp64 baseline measures the solved likelihood at 62.687 ms/call and the separately labelled free-centre control at 74.346 ms/call. Eager/JIT likelihoods agree, vmap(2) agrees in both lanes, the gradient is finite, and four finite model positions survive filtering.

Scripts Changed

• scripts/point_source/likelihood_breakdown/image_plane.py — adds the device-aware shared breakdown, source/library provenance, JIT-phase timing, numerical controls, and two-panel absolute-prefix/signed-difference plot.
• results/breakdown/point_source/image_plane_local_cpu_fp64.{json,png} — records the first canonical CPU fp64 reference row.
• results/notes/point_source_shared_likelihood_breakdown.md — explains the measurement contract, fusion caveat, CPU result, and A100 handoff.
• README.md and scripts/misc/likelihood_breakdown/README.md — add the generated point_source/image_plane result row.

Test Plan

• Full CPU fp64 breakdown completed successfully on committed instrument revision 3052443 (626 s).
• Eager/JIT solved and plain likelihood parity.
• JIT/vmap parity and finite full-model gradient.
• ruff check .
• ruff format --check .
• New cell smoke import plus the five maintained CI smoke commands.
• python scripts/misc/tooling/build_readme.py --check
• git diff --check
• Independent review CLEAN; empirical commit claims are backed by the committed run/artifacts.

Heart readiness was STALE with no RED/YELLOW reasons: release validation incomplete: no rehearsal for current source. This permits a development PR and remains a release blocker.

Closes #291.

Generated by the PyAutoLabs agent workflow.