Import generic optics helpers from aocore

[jacotay7]

Summary

aocore owns the AO stack's conventions and its generic optics primitives. CONVENTIONS.md §9 says consumers import them rather than re-implement them. This PR moves makewfs onto them and makes aocore>=0.1.2,<0.2 a core dependency. It was previously in the dev extra. makewfs already requires Python >= 3.10, which matches aocore.

Replaced

makewfs local code	now uses
api._ARCSEC_PER_RADIAN = 206264.806… (plate scale)	aocore.RAD_TO_ARCSEC (same value)
math.radians(x / 3600) in source.py (field angle, angular kernel, FWHM)	x * aocore.ARCSEC_TO_RAD
(arange(n) - (n-1)/2) in wavefront._coordinates, the SH field-stop mask and the DFT detector grid	aocore.centered_coordinates (scaled on the backend in the working dtype)
phase * λ / 2π for phase input	aocore.phase_to_opd
2π · opd / λ in the SH and pyramid phasors	aocore.opd_to_phase
general-factor branch of sampling.block_sum	aocore.block_sum

makewfs.sampling.block_sum keeps its public signature, error messages and factor-two fast path. aocore has no such fast path, and factor 2 is the common SH oversampling.

Kept local, and why (each module docstring now says so)

• pupil.make_pupil: aocore.Pupil cannot express the configured geometry. That geometry has angular-wedge spiders (width_fraction of π), square-grid segment gaps, pupil_rotation_deg and rectangular grids, and it is built on the selected CuPy/NumPy backend in the working dtype.
• backend.ArrayBackend and its centred FFT helpers: the sensors choose a dtype per array, pass FFT workers and overwrite_x, and need ndimage helpers (map_coordinates, convolve, gaussian_filter). aocore.Backend has one precision per backend and its own worker policy. The centred FFTs use the fftshift convention internally, and the engines convert to CONVENTIONS 1.3 centring, which tests/test_conformance.py checks.
• Quadrature offsets ((i + 0.5)/s - 0.5, used for pupil supersampling and the lenslet fill factor): these are sub-pixel sample offsets, not a coordinate grid.
• Metadata wfs_input_opd_rms_m: this is a whole-input-grid quadratic mean with piston included, reduced on the device. aocore.rms is pupil-weighted, removes piston and runs on the host, so it is a different quantity. It is unchanged here (see below).
• Validation and example centroids: validation/run.py and examples/keck_haka keep their batched, index-origin centroids so checked-in artifacts stay comparable.

Behaviour

No CONVENTIONS violation was found, so the physics is unchanged. I rendered every examples/configs/*.toml on main and on this branch with the same random OPD:

• float32 (as shipped): bit-for-bit identical;
• forced to float64: max relative difference ≤ 6e-16. This is rounding from the changed operation order in the angle and phase conversions and, for factors above two, in binning.

Public names are unchanged. tests/test_conformance.py now imports aocore directly instead of importorskip.

Docs

CHANGELOG (Unreleased), AGENTS.md (product boundary and code organisation), README and docs/units-and-coordinates.md are updated.

Test plan

Local 3_13 conda env:

• ruff check . and ruff format --check .
• python -m mypy
• pytest --cov: 208 passed, coverage 92 %. The 2 failures in tests/test_oopao_validation.py are pre-existing; they fail identically on main (local OOPAO environment).
• pytest -m gpu on a Quadro P620: 17 passed
• python benchmarks/run.py --representative and check_regression.py: OK
• validation/run.py, mkdocs build --strict and python -m build all succeed

🤖 Generated with Claude Code