Diagnostics

Filters, trackers, and samplers should expose diagnostics through explicit return paths rather than printing or relying on private attributes. The module pyrecest.diagnostics defines lightweight dataclasses for common diagnostic families:

• FilterDiagnostics for innovation, NIS, NEES, log likelihood, and covariance trace;
• ParticleDiagnostics for effective sample size, resampling decisions, and weight entropy;
• AssociationDiagnostics for cost matrices, gated measurements, assignments, births, and deaths.

A recommended pattern is:

state, diagnostics = filter.update(measurement, return_diagnostics=True)

APIs that do not yet support diagnostics should avoid ad-hoc return tuples and instead add a documented diagnostics object when the feature is introduced.