fix: key bootstrap model selection by basis identity (#16)

[jkitchin]

Fixes #16.

The problem

bootstrap_model_selection keyed feature_frequencies by selected_features_, which for a parametric basis is the name with the fitted value substituted in. The nonlinear parameter is re-optimised in every replicate, so each replicate contributed a uniquely-named feature — nothing aggregated, and stability_score reported 0.0 even when every replicate had selected the same basis.

The fix

Key by basis identity. New BasisLibrary.canonical_name(index) and BasisLibrary.canonical_names return the name a basis was registered with ("exp(-a*x)") rather than the fitted rendering ("exp(-0.4913*x)"); non-parametric bases are unaffected. Both the frequency counts and the feature-set comparison behind stability_score now use it. Fitted values are reported the way a coefficient is — as a distribution over replicates — under a new "parameter_distributions" key, as the issue proposed.

On a reproducer of the WLF-style case from the issue (15 replicates):

feature_frequencies : {'q*y_x': 1.0, 'y_x/(c2+q)^2': 1.0, 'y_x': 0.33, '1': 0.067}
parameter_dists     : {'y_x/(c2+q)^2': {'c2': {'mean': 0.569, 'sd': 0.033, 'q05': 0.541, 'q95': 0.622, 'n': 15}}}
stability_score     : 0.6

Two related defects found in the same function

• It mutated the caller's model. Every replicate was fitted on model.basis_library itself, and _resolve_parametric_params rebinds basis names and evaluation closures in place — so after the call the user's fitted model silently predicted with the last replicate's parameter values. Replicates now fit on a new BasisLibrary.copy(), with a regression test asserting the original model's predictions and names are unchanged.
• The clone dropped constructor arguments. It enumerated eight of thirteen, losing param_optimizer, param_optimization_budget, prune_tol and the constraint settings — the first two govern exactly the parametric fits this issue is about. It now clones through get_params().

Also in this change

• Input validation: unfitted template model, X/y length mismatch, n_bootstrap < 1.
• New "n_successful" return key, so the denominator behind the reported frequencies is legible.
• feature_frequencies is ordered most-selected first.
• Tests: 10 for bootstrap_model_selection (aggregation, parameter distributions, non-mutation, reproducibility, validation) and 7 for the new BasisLibrary methods.
• Docs: uncertainty guide section on parametric bases, re-synced into src/jaxsr/skill/; CHANGELOG under Unreleased.

Scope

The issue's cross-reference to #17 (what gets resampled) is untouched — that is a separate concern, and these frequencies still mean "fraction of pairs-bootstrap replicates", nothing stronger.

Checks

• pytest tests/ — 638 passed, 10 skipped
• black --check src/ tests/ and ruff check src/ tests/ — clean
• scripts/test_under_numpy.py on the touched test files — 98 passed

🤖 Generated with Claude Code

https://claude.ai/code/session_01ARRqz2SjSoYH5N3t66HDjm

---

Generated by Claude Code