Expand test coverage

[AUTProgram]

There is currently at least one test (test_power_a) where the combination of floated parameters is not actually picked up by the estimate_parameters function of the model. We need to check if this is the only test where this is the case and fix them all.

[hartytp]

Agreed: we should expand the test coverage to cover every heuristic path.

It would also be good to get some tests which use noisy, sparsely-sampled data rather than relying on the noise-free, uniformly-sampled data we have in these tests. We did some of this kind of testing when developing the original fit functions (e.g. checking how FFT-based heuristics behave with sparsely-sampled data) but they never made it into the test suite.

It would also be nice to have some data which is not generated from exactly the same model we're fitting to -- e.g. real lab data where the model is a reasonable approximation of the underlying distribution, but there are still some noticeable systematics. We should still expect robust fits in this case.

A related question is more generally what we want to do about testing heuristics. At present we just check that the fit converges and take a view that "heuristics working means they get us close enough for the main solver to pass." That's a pretty decent functional definition, however it allows for the possibility that heuristics are considered working when they functioning as intended but aren't sufficiently broken to cause the fit to fail to converge. That seems rather unsatisfactory and allows bugs to hide undetected. Ideally we would have some tests specific to heuristics.

That is probably easy to implement but we need to decide what we want first. IME it's hard to put a tolerance on heuristics that is tight enough to catch errors but not so tight that they fail regularly. I'm not sure what the best approach here is.

[hartytp]

would be nice to set some heuristic tolerances as well. That will catch gross errors in the heuristics

[hartytp]

• [ ] One thing we really need is a way of running tests where the datasets are passed in
• [ ] unify check_single_param_set and check_multiple_param_sets
• [ ] way of not passing in fixed parameters to check_single_param_set
• [ ] test on derived parameters?
• [ ] checking of parameters for periodic parameters?
• [ ] add test for sorting and scaling done by fitter including the various corner cases (repeated points, infinite points, etc)
• [ ] test tooling currently assumes the input data is sorted (otherwise comparisons break)
• [ ] test tooling support 2d models (e.g. plotting)
• [ ] add tests for fit validator
• [ ] make check_multiple_param_sets not force us to specify values for fixed parameters