Drop the pynufft legs from the NUFFT parity script - #261

Summary

Workspace-test follow-up to PyAutoLabs/PyAutoArray#475 (merged), which removed TransformerNUFFTPyNUFFT.

scripts/interferometer/nufft.py constructed that class directly, so it could not run at all after the removal. Its original purpose — proving nufftax matched pynufft closely enough to replace it — has been served: that swap shipped, and pynufft is now gone from PyAutoArray entirely.

What remains is the standing correctness check, with TransformerDFT (a direct, unapproximated Fourier transform) as the sole reference:

• tests (a) and (b) drop the pynufft leg and gain an explicit library-vs-DFT assertion, so the shipped TransformerNUFFT is now checked against the exact reference rather than only against the local nufftax recipe.
• test (c) used pynufft as its mapping-matrix reference; it now uses TransformerDFT.transform_mapping_matrix. The tolerance tightens from 1e-1 to 1e-9 — the old one was pynufft's gridding precision, the new reference is exact.
• test (d) needed no functional change; only its pynufft prose is gone.
• the residuals plot drops from three panels to two: the shipped transformer and the local recipe agree bit-for-bit (exactly 0.0), so the third panel was an empty log-scale axis emitting a matplotlib warning. The equality is still asserted.
• import warnings removed — every catch_warnings block existed to silence pynufft.

The full-resolution requirement is re-justified rather than dropped. It existed because pynufft's gridding error scaled with N, so the tolerances only held at 256x256. nufftax is accurate at any N (measured 3.0e-14 vs 1.7e-14 relative at 256x256 and 16x16), so the ENV: full_datasets declaration and the geometry guard stay for a different reason: 256x256 / 0.1" is the production geometry the JAX likelihood scripts use, and the guard is now the only thing that catches a silently PYAUTO_SMALL_DATASETS-capped run.

In scripts/interferometer/jax_likelihood/rectangular_sparse.py, the recorded reason for avoiding apply_sparse_operator with TransformerNUFFT was specific to pynufft's kernel-deconvolved adjoint scale. That backend is gone and I have not re-verified whether the incompatibility still holds against the nufftax adjoint, so the dead attribution is removed without asserting a replacement claim. That path deliberately does not depend on the answer. Worth a look separately.

Scripts Changed

• scripts/interferometer/nufft.py — the substantive rewrite (73 insertions, 105 deletions)
• scripts/interferometer/jax_likelihood/rectangular_sparse.py — comment only
• scripts/CLAUDE.md — the script-catalogue entry, which described the file as a pynufft-vs-nufftax parity test

Test Plan

• python scripts/interferometer/nufft.py runs end to end and passes all four tests, no warnings:
  • (a) 5x5 all-ones — max |Δ| nufftax - DFT = 1.5e-13, library - DFT = 1.5e-13
  • (b) 256x256 lensed Sersic, SMA uv — 2.2e-12 absolute, 3.0e-14 relative for both nufftax and the shipped library
  • (c) mapping matrix — 3.0e-14 relative vs DFT; library vs local recipe exactly 0.0
  • (d) adjoint identity 1.1e-15; round-trip peak distance 3.61 px (tolerance 6.0)
• Residuals plot still written to scripts/interferometer/images/nufft_residuals.png

Generated by Claude Code