Share one spectrum across a centroid's fits, in both optimisers

[brmather]

Cleanup pass over the two spectrum-sharing commits (0107ecb, dee8caf), from /simplify, then the generalisation that review recommended.

1. Provenance rides on the spectrum, not beside it

_spectrum_key was a second instance attribute that had to be kept in step with last_spectrum at every exit, and it only ever knew the provenance of the single most recent spectrum. Reuse survived only because asarray returns an already-float64 array unchanged, so the arrays could be matched by identity — correct, but three lines of comment defending a coincidence.

_Spectrum is a tuple subclass carrying its own provenance. That removes _spectrum_key, the ours flag and the identity zip, and it is strictly more capable: a spectrum held across an intervening call no longer loses its label. It unpacks, indexes and pickles as a plain tuple, so spectrum=, last_spectrum and Global_CPD's fit_spectrum(*...) are unchanged.

A process_subgrid no longer disables multiprocessing for good

The old key stored the taper and process_subgrid callables. That kept their captured scope alive, and made pickle.dumps(grid.optimise) raise for a closure or lambda — so one optimise(..., process_subgrid=...) sent parallelise_routine to serial for every later call on that instance, warning that it could not send func, which was perfectly picklable. Measured before the change: 48 centroids on 4 processors, 0.16 s → 0.48 s, and it stayed there.

parallelise_routine rejects spectrum

Alongside on_error and seed. One spectrum cannot describe a list of centroids: forwarded, it gave every centroid the same answer — a flat map that looks like a result — and since the warning fires per call, whether the user was told depended on how many processors were available.

Corrections and tidying

• residuals claimed np.errstate is thread safe where catch_warnings is not. It only became context-local in NumPy 2.0 and pyproject.toml allows numpy>=1.20. The two arguments that hold on every version are kept.
• Four identical spectrum : parameter blocks collapse to one description plus three cross-references — the pattern already used for dof_factor.
• Tests use monkeypatch, pytest.warns and recwarn in place of hand-rolled restores and warning capture.

2. Tanaka gets the same sharing

The mechanism was already class-agnostic — it never mentions power, which each _spectrum pins below the seam — so it moves to CurieGrid beside window_spectrum.

Tanaka has the stronger case of the two. Two straight-line fits cost almost nothing beside computing the spectrum — 24.8 ms of a 25.9 ms optimise at a 1025-cell window, 96% — and the documented workflow is a band sweep: check_bands needs k, then optimise needs all of it, then sensitivity, then a revised band needs it again. None of those depend on the bands. Measured on optimise / check_bands / optimise: 51.9 ms → 26.6.

Two things were Bouligand-specific, both parameters in disguise. A subclass defining _spectrum now also defines:

name	why
_SPECTRUM_ARGS	its positional argument names, in order, so provenance is picked out by name not position — Tanaka takes seven with beta at position five, Bouligand six
_SPECTRUM_PROVENANCE	the subset a reused spectrum is checked against. Tanaka's includes beta, which subtracts the fractal contribution and so changes Phi — now a warning rather than a silent answer
_SPECTRUM_RETURNS	fixes the arity (four arrays against three) and names them when it is wrong, which matters where the order of (k, Phi, Phi_n, sigma) is not obvious

Call sites are unchanged apart from the name. optimise and sensitivity take spectrum=; parallelise_routine already rejects it and is inherited, so the Tanaka routines get that for free — verified rather than assumed.

Testing

Suite 121 → 129 tests, all passing. New coverage includes: a grid that ran optimise with a lambda process_subgrid still pickles; a routine refuses a shared spectrum; Tanaka's four-array spectrum round-trips and its beta is provenance while its bands are not; and test_grid.py pins the three class attributes against inspect.signature, since a second list to keep in step with a signature is the drift this refactor exists to remove.

🤖 Generated with Claude Code

https://claude.ai/code/session_01PLgEaDtuZw6QXDphSrzcdQ