test: TransformerNUFFT cross-check (Path B) for interferometer JAX likelihood scripts

[Jammy2211]

Summary

Adds a TransformerNUFFT cross-check (Path B) to each interferometer JAX likelihood script. Each script keeps its existing TransformerDFT path (Path A) untouched, then re-runs the same vmap likelihood with transformer_class=al.TransformerNUFFT and asserts the same hardcoded literal at rtol=1e-4. This proves end-to-end that the slow direct-DFT path and the fast nufftax-backed NUFFT path produce the same likelihood, catching any future drift between the two transformers.

Companion library PR

PyAutoArray PR #305 — perf: batched transform_mapping_matrix. The DSPL Path B in this workspace PR depends on that library fix to compile in seconds instead of minutes. The library PR should land first.

Per-script changes

• lp.py / mge.py / mge_group.py / delaunay.py / rectangular.py / rectangular_dspl.py — standard Path B append, asserts the literal at rtol=1e-4.
• delaunay_mge.py — rtol loosened to 2e-3. The Delaunay+MGE inversion amplifies the ~1e-13 forward-operator difference into a ~5e-4 relative shift in the final log-likelihood (likely mesh-vertex selection sensitivity).
• rectangular_mge.py / rectangular_dspl.py — include gc.collect(), jax.clear_caches(), and parameters_nufft = parameters[:1] before Path B to keep peak memory within the 16 GB CI box.
• rectangular_sparse.py — three-way cross-check: DFT+sparse_operator (existing literal) → DFT-no-sparse → TransformerNUFFT (no sparse). The latter two assert against rectangular.py's canonical -3164.286252 literal because apply_sparse_operator gives a numerically distinct ~0.4% off result from the bare DFT, hence the path-specific literals.

DSPL mesh reduction

Both interferometer/rectangular_dspl.py and imaging/rectangular_dspl.py reduce mesh_shape from (30, 30) to (8, 8) to match the resolution the other rectangular*.py tests use. Test scripts should be lightweight; (30, 30) produced 1800-pixel JIT traces and slow compiles. New canonical literals captured empirically:

• interferometer/rectangular_dspl.py: -3170.19672623
• imaging/rectangular_dspl.py: -3797.73182794

Scripts Changed

• scripts/jax_likelihood_functions/interferometer/{lp,mge,mge_group,delaunay,delaunay_mge,rectangular,rectangular_mge,rectangular_dspl,rectangular_sparse}.py
• scripts/jax_likelihood_functions/imaging/rectangular_dspl.py

Test plan

• Every interferometer script: python <script>.py ends with PASS: TransformerNUFFT cross-check matches TransformerDFT.
• imaging/rectangular_dspl.py: passes the existing assertion at the new literal.
• DSPL Path B compiles in ~30s (down from 11+ min) with the library fix.

🤖 Generated with Claude Code