fix: initial_peak_guesses refuses unknown models instead of falling back (v0.1.5)

[garrekstemo]

What

initial_peak_guesses carried two unknown-model fallbacks — npp = ... : 3 and positional p$i parameter names — that could never run. _peaks_to_p0 calls _n_peak_params, which throws "Unknown model. Provide n_peak_params manually." before either fallback is reached, so a custom lineshape always errored, just with a message naming a keyword that does not exist anywhere in the package.

The function now validates _is_known_model up front, names the five built-in lineshapes, and points custom-model callers at fit_peaks' explicit p0. Both dead branches are gone and the docstring states the constraint.

Failing fast is the right resolution rather than making the fallback work: automatic guesses need the model's parameter layout — the parameter count, which slot holds the width, and how a detected FWHM maps onto it via width_scale. None of that is inferable from an arbitrary function, and assuming a 3-parameter amplitude/center/width layout would silently build a wrong-length or wrong-order p0 for a 4-parameter custom model.

What is deliberately unchanged

The same-looking fallbacks in fit_peaks (npp, _model_name, param_names) are not dead, and are left alone. With an explicit p0, fit_peaks never calls _peaks_to_p0, so a custom lineshape fits normally and reports positional parameter names. Verified before touching anything:

custom_peak(p, x) = @. p[1] / (1 + ((x - p[2]) / p[3])^2)
f = fit_peaks(x, y; n_peaks=2, model=custom_peak, p0=[0.5, 2020.0, 8.0, 0.3, 2080.0, 6.0, 0.01, 0.0])
f[1].params      # [:p1, :p2, :p3]
f[1].model       # "custom_peak"
f.r_squared      # 1.0

Applying the same up-front check there would have removed a working feature.

Tests

Extends the initial_peak_guesses testset with both halves: the refusal (ArgumentError, message mentions "known lineshape") and the custom-model fit that still works through p0.

Full suite: 1940 passed, 0 failed.

Note, not addressed here

_n_peak_params' message still says "Provide n_peak_params manually", but no function in the package takes such a keyword. After this change its only remaining live caller is the TA peak-spec fit in fitting.jl:1037, so a custom peak function there is still pointed at a keyword that does not exist. Out of scope for this fix.

🤖 Generated with Claude Code