fix(types): refuse NaN and out-of-range probabilities; stop the Brier floor indexing past a distribution

[TMHSDigital]

Fixes #31. Fixes #32.

#32: Prediction accepted impossible probabilities.

• The check was that the distribution summed to about 1. It never checked the entries. {"a": nan, "b": 1.0} and {"a": 1.5, "b": -0.5} both passed, and so did a NaN latency.
• Each distribution entry must now be finite and in [0, 1], and the error names the option. latency_ms must be finite.

#31: multiclass_brier_floor could index past the last option.

• It drew the gold label with searchsorted over the raw cumulative sum. A distribution summing to 0.9995, which Prediction allows, then indexed one past the last option whenever the uniform draw landed above the sum.
• The cumulative sum is now normalized and the index clamped. A distribution that sums to 1 draws exactly as before, so no existing figure moves. The site build confirms the example report still matches its command line for line.

Checked

• New tests failed on main and pass now: NaN and negative distribution entries are refused, NaN and inf latency are refused, and the floor accepts a 0.9995 distribution.
• ruff, mypy --strict, the full pytest suite, and the site build all pass.

🤖 Generated with Claude Code