Correction: the jax-compile-stall is not a compile stall - #279

…oArray mapper bug

Reproduced the failure bit-for-bit with a runnable stack built in-session (both
Python legs, source-installed libraries at the retime run's exact versions), and
the prompt's own UPDATE 2026-08-24 diagnosis does not survive it.

Refuted, with evidence:
  1. NOT "Python 3.13 only" — 3.12 fails identically on the same host, so the CI
     split was runner hardware, not CPython.
  2. NOT a PDIP branch flip — pdip_iter is identical eager vs jit, and the gap is
     bit-identical across 5 (nnls_solver_tol, nnls_max_iter) policies.
  3. Therefore pinning solver policy cannot fix it; that direction is dead.

Actual cause: a discrete bilinear cell-assignment flip in PyAutoArray's
rectangular mapper. transform() ends in clip(F_q, 0, 1), so saturated points land
on exactly-integer indices where ix_up = ceil(g) collapses onto ix_down; one ULP
in the traced grid then jumps a point's weight a whole mesh row. Underneath it the
row weights are mirrored — the current code fails a linear-reproduction test by
~a full cell in the row axis while the column axis is exact.

Spawns draft/bug/autoarray/rectangular_mapper_bilinear_row_weights.md as the real
fix (human-required: it changes reconstructions library-wide and 16 workspace
scripts' hardcoded constants). The NEEDS_FIX park stays until that lands.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016azuS2UbS3mFvfxkDFGsKj

…#279 / PyAutoArray#490)

draft/test/workspaces/physical_model_check_when_speeding_up_smoke.md
  A speed-up on a validation script is not done until that script's lens model is
  known to be physical. #490 survived 11 months partly because the jax_likelihood
  scripts assert at deliberately-wrong models: rectangular_rtu.py pins the mass
  centre to (0.3,-0.3) against a (0,0) truth, and the jax_test dataset carries
  Sersic+Exponential lens light the mass-only models have no component for (there,
  truth measurably fits WORSE than the offset model). Records the pixelized-source
  caveat: assert on log_likelihood and source recovery, not the figure of merit,
  which for a pixelization is the evidence and moves for unrelated reasons.

draft/test/autoarray/final_numerics_audit_of_every_mesh_interpolator.md
  One systematic pass over every mesh interpolator using the five checks that
  actually discriminated in #490 — linear reproduction, continuity across integer
  boundaries, convergence under refinement, ground-truth source recovery, and a
  bit-identical known-good control. Carries the specific suspicions: saturating
  transforms feeding a discretisation, corner/weight pairing, n_knots not scaling
  with mesh_pixels (roundtrip drift 0.0055 -> 0.0238 -> 0.101 index units at
  n=16/32/64), and guard-node conventions. Requires every new test to be shown
  failing against a broken variant.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016azuS2UbS3mFvfxkDFGsKj

…regression, not an NNLS branch flip

Both PRs merged library-first and verified as ancestors of their main with 0
unmerged commits:
  PyAutoArray#490            merged 158db384
  autolens_workspace_test#280 merged 8be3d598

The filed prompt's own diagnosis was refuted on three counts by building both
Python legs from source and reproducing the failure bit-for-bit: not 3.13-only
(3.12 fails identically on the same host — the CI split was runner hardware),
not a PDIP branch flip (pdip_iter identical eager vs jit; the gap bit-identical
across five solver policies), and therefore not fixable by pinning solver policy.

Actual cause: mirrored bilinear row weights plus a round-off-dependent cell
assignment in PyAutoArray's adaptive rectangular mapper, live since 2025-09-23.
Both are regressions — the mapper was correct at introduction (2025-06-24) and
the correct formulation still ships for the uniform mesh at
interpolator/rectangular_uniform.py:72-99, which is what the fix restores.

Moves active/ -> complete/2026/08/, releases the active.md claim, refreshes
complete/index.md and regenerates the dashboard in this commit.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016azuS2UbS3mFvfxkDFGsKj