Review: remaining unreviewed corners — fit_best, Distribution ABC, serialisation registry, plotting layer, datasets, experimental shim

[derrynknife]

Smaller corners that no review round has covered, grouped for one pass:

• surpyval/fit_best.py — automatic model selection: does it fit every candidate with the requested method, rank by the documented criterion (AIC/BIC/AICc?), and handle fit failures without silently dropping or mis-ranking candidates?
• surpyval/distribution.py — the Distribution ABC itself (default implementations, operator hooks — note the |/& hooks go away with Remove the alpha series/parallel composition models #284)
• surpyval/serialisation.py — the package-level from_dict/from_json registry and schema-version dispatch (individual model round-trips are well tested; the registry/dispatch code itself was never read)
• The plotting layer: univariate/parametric/probability_plotting.py, nonparametric/filliben.py, nonparametric/success_run.py, and the plotting-position heuristic formulas in nonparametric/plotting_positions.py (Blom etc. — only exercised indirectly via MPP fits). Filliben correlation coefficient vs its published definition; success-run formula vs the standard nonparametric reliability bound
• surpyval/datasets/ — loaders return what their docstrings claim
• surpyval/experimental/ — legacy shim after the alpha rename: still importable, warns appropriately, no stale duplicate code paths

Low expected yield per line compared to #290/#291, but cheap to sweep and currently zero coverage.