feat(jax_profiling): hannah ALMA preset for datacube + interferometer delaunay

[Jammy2211]

Summary

Add a "hannah" instrument preset to the jax_profiling datacube + interferometer Delaunay profilers that pins Hannah Stacey's realistic ALMA science-case parameters. Step-by-step cube cost at this scale (n_channels=34, 16984 visibilities/channel, 40×40 real-space): 205.92 s per cube evaluation, with an estimated 34.82 s saving available once the shared-Lᵀ W̃ L optimisation lands.

Hannah's quoted settings:

• n_channels = 34
• n_visibilities = 16984 per channel
• pixel_scale = 0.125 "/pixel
• shape_native = (40, 40)
• mask_radius = 2.3 "

Scripts Changed

• jax_profiling/dataset_setup/interferometer.py — add "hannah" preset to the INSTRUMENTS dict; thread transformer_class per preset (NUFFT for ALMA-scale, DFT for sma/alma); dataset auto-simulates to dataset/interferometer/hannah/.
• jax_profiling/jit/interferometer/delaunay.py — add "hannah" to INSTRUMENTS; promote mask_radius to a per-instrument field; convert EXPECTED_LOG_EVIDENCE_SMA to a per-instrument EXPECTED_LOG_EVIDENCE dict (hannah pinned at -204838.07924622478); pass raise_error_dft_visibilities_limit=False so the JAX-traceable DFT path can load >10k-visibility datasets (the JIT path is intentionally DFT until nufftax is wired in).
• jax_profiling/jit/datacube/delaunay.py — add "hannah" to INSTRUMENTS; promote mask_radius; switch default instrument="hannah" and n_channels=34; gate Part C (full-pipeline cube JIT) behind CUBE_FULL_JIT=1 since lower+compile at n_channels=34 is ~70 s on this laptop and was killing the default run; convert regression literal to a per-instrument EXPECTED_LOG_EVIDENCE_PER_CHANNEL dict.
• jax_profiling/dataset/interferometer/hannah/ — new simulated dataset (16984-visibility NUFFT-simulated ALMA cube, 40×40 grid, single-channel reused 34× for the cube profile).
• jax_profiling/results/jit/interferometer/delaunay_likelihood_summary_hannah_v2026.5.14.2.{json,png} — captured single-channel hannah timings + bar chart.
• jax_profiling/results/jit/datacube/delaunay_likelihood_summary_hannah_v2026.5.14.2.{json,png} — captured 34-channel cube timings + bar chart.

Upstream PR

• feat(interferometer): from_fits accepts raise_error_dft_visibilities_limit PyAutoArray#311 — Interferometer.from_fits accepts raise_error_dft_visibilities_limit. This profiler PR consumes that kwarg, so the library PR should merge first.

Test Plan

• Single-channel hannah profiler runs end-to-end: eager-vs-JIT correctness PASSED, per-channel regression pinned at -204838.07924622478 PASSED.
• 34-channel hannah cube profiler runs end-to-end with default (Part C skipped): per-step recompute matches summed FitInterferometer.log_evidence at rtol=1e-4 (= -6964494.694372); eager cube regression vs 34 × -204838.07924622478 PASSED.
• Backward-compat: instrument="sma" remains the default in the single-channel profiler; the existing SMA regression literal is unchanged; n_channels=4 for the cube still works when toggled back.

🤖 Generated with Claude Code