results: #678 phase B evidence campaign — truth-anchored A100 cells + campaign notes

[Jammy2211]

The autolens_profiling half of the point-source defaults campaign
(PyAutoLens#678): the
evidence-campaign harness, 23 truth-anchored A100 result cells, and the written
synthesis the workspace guides cite.

Every cell evaluates its likelihood flavour at the simulator-truth model and
scores the search by delta = max_log_likelihood − truth_log_likelihood, so a
likelihood defect (a wrong model out-ranking truth) is distinguishable from a
search defect (failure to reach the basin). That distinction is what the
campaign's conclusions rest on.

What the evidence established

• All-to-all pairing as the default — on a quad with one true image removed,
  PairRepeatSolved mis-ranks truth by +1.8e5 while PairAllSolved recovers it
  (+1.3). On clean data the two are equivalent, so robustness decides. This is
  what shipped as PyAutoLens#686.
• Tensor source-plane weighting ranks truth first at both tiers; the scalar
  convention mis-maps the μ=367 image's radial noise.
• Solved centres on both chi-squared flavours, with free centres documented
  as needing sampler muscle (the free-centre gradient plateau survives the
  phase-A logsumexp fix).
• Two library defects found and fixed by the cluster gradient cells —
  PyAutoFit#1441 (traced attributes classified as pytree aux) and PyAutoLens#685
  (padded solver rows anchored at NaN-Jacobian centres). Both merged.

Evidence tail in this PR

The final commit closes the campaign out with the last three A100 jobs:

• 331887 re-runs the cluster gradient cell that exposed both library
  defects. It now completes end-to-end (832 s, finite, no NaN), validating the
  merged fixes through a full search. But it still misses the basin
  (delta −1723.6 vs Nautilus +16.9), which closes the one question the notes had
  left open: the solved image-plane cell was the last candidate for a
  cluster-scale gradient-search recommendation, and it fails alongside the
  source-plane cells. At cluster scale, every objective tested is Nautilus's.
  Finite gradients were a correctness fix; they did not buy convergence.
• 331888/331889 replace two cells that had been silent PyAutoFit resumes
  (sub-second sampler wall — identifiers ignore the dataset). Both reproduce
  their prior conclusions within scatter, so the resumes had not misled, but the
  shipped numbers now rest on searches that actually ran.
• 331885/331886 are recorded as a deliberate non-result: both simple_extra
  arms timed out at 8h walls. A spurious position imposes an honest ~−1.4e4
  floor on every model, so Nautilus never converges. The DNF is symmetric,
  so this arm does not discriminate between the pairings — the default rests on
  the missing-image arm, which was already in hand.

Notes

• The notes doc records two harness quirks that are easy to misread as defects
  of a given run: likelihood_evals is 65 for every MultiStartProdigy cluster
  cell (it counts vmapped batch calls), and best_fit is an AttributeError on
  every cluster cell including the Nautilus ones.
• Results-only repo — no library or workspace code changes here. The workspace
  prose that cites these numbers is autolens_workspace#468.

🤖 Generated with Claude Code