Affected Repositories

• PyAutoGalaxy (primary)
• PyAutoLens (docstring + mirror test)

Branch Survey

Suggested branch: feature/interferometer-primary-beam-persistence

Implementation Steps

See the prompt's "What" section (steps 1-4 with file:line anchors).

Key Files

• autogalaxy/interferometer/model/analysis.py — interferometer_hdu_list_from (:76-218), save_attributes (:400-455)
• autogalaxy/aggregator/interferometer/interferometer.py — _interferometer_from (:161-235), _has_hdu (:50-59)
• test_autogalaxy/interferometer/model/test_analysis_interferometer.py — round-trip patterns (:148, :172-255, :258-285)
• autolens/interferometer/model/analysis.py (:336-339), test_autolens/interferometer/model/test_analysis_interferometer.py (:220)

Persist the interferometer primary beam through analysis save and aggregator rebuild

Type: feature
Target: PyAutoGalaxy
Repos:

• PyAutoGalaxy
• PyAutoLens
  Difficulty: small
  Autonomy: safe
  Priority: medium
  Parent: complete/2026/10/interferometer-primary-beam.md
  Status: formalised
  Consequence: glance
  Witness: a test asserting an Interferometer with a primary beam saved by AnalysisInterferometer and rebuilt by the aggregator carries an identical primary_beam_slim
  Review-minutes: 3
  Unattended: ready
  Filed: 2026-10-21

What (survey of main 2026-10-21, PyAutoArray at 034a616f incl. #644/#645)

The beam is saved nowhere today. interferometer_hdu_list_from (autogalaxy/interferometer/model/analysis.py:76-218) writes one FITS: in-memory branch HDUs MASK, DATA, NOISE_MAP, UV_WAVELENGTHS (:117-126); array-free branch MASK, NUFFT_PRECISION_OPERATOR, DIRTY_IMAGE, DIRTY_BEAM, SPARSE_TERMS_SCALARS (:129-217). save_attributes (:400-455) writes image/dataset.fits and transformer_class. The aggregator _interferometer_from (autogalaxy/aggregator/interferometer/interferometer.py:161-235) rebuilds via from_sparse_terms (:206-213) or aa.Interferometer(...) (:215-231); _hdu() falls back to legacy positional indices, so a new HDU must be detected with _has_hdu. PyAutoLens reuses both functions (no source change). FitInterferometer passes the dataset's transformer and sparse_operator to DatasetInterface, so passing primary_beam= back into the constructor is sufficient.

1. analysis.py: helper _primary_beam_native_from(dataset) → None or the float64 native array (masked pixels 0) of primary_beam_slim on real_space_mask; append it as a trailing PRIMARY_BEAM HDU in both branches (legacy indices 0-3 unchanged); docstrings.
2. aggregator/interferometer/interferometer.py: _primary_beam_from(hdu_list) via _has_hdu; pass primary_beam= into from_sparse_terms and aa.Interferometer; docstrings.
3. Tests in test_autogalaxy/interferometer/model/test_analysis_interferometer.py (extend the round-trip patterns at :148, :172-255, :258-285 with a Gaussian beam spanning ~0.3-1 on interferometer_7): (a) in-memory beamed round trip — PRIMARY_BEAM present, primary_beam_slim and transformer.primary_beam_slim identical, FitInterferometer.log_likelihood equal (rel 1e-10) for light-profile and pixelization galaxies, and different from the beamless fit; (b) array-free beamed round trip via from_stream(primary_beam=) — sparse_operator.primary_beam_slim identical, log_evidence / figure_of_merit equal; (c) beamless writes no PRIMARY_BEAM HDU, HDU count pinned at 4, reload gives primary_beam is None; (d) old outputs without the HDU still load with primary_beam=None for both layouts.
4. PyAutoLens: docstring line at autolens/interferometer/model/analysis.py:336-339 and a mirror beamed round-trip test next to test_autolens/interferometer/model/test_analysis_interferometer.py:220 (optional, do it).

Not in scope: the pre-existing gap that an in-memory dataset with apply_sparse_operator comes back without the operator (likelihood still matches via the beamed transformer); the forced-numba fallback and repeated P=0 warning (PyAutoArray side, separate small prompt if wanted).