feat(regime): regime-conditional surrogate (#105)

[jc-macdonald]

Closes #105. Builds on #82.

Summary

Adds trade_study.regime so studies that currently gate on a regime feature (e.g. n_samples, noise_level) can:

1. Borrow strength across regimes. Settings that work at n=1k inform predictions at n=3k instead of being thrown away.
2. Interpolate across continuous regime axes. No more hard buckets — query the surrogate at any regime value.

API

from trade_study import Factor, FactorType, fit_regime_surrogate

surrogate = fit_regime_surrogate(
    results,
    regime_factors=[Factor("n_samples", FactorType.CONTINUOUS, bounds=(1_000, 10_000))],
    factors=[Factor("lr", FactorType.CONTINUOUS, bounds=(1e-4, 1e-1))],
    method="gp",
)
surrogate.predict({"n_samples": 2200}, {"lr": 1e-3})
surrogate.uncertainty({"n_samples": 2200}, {"lr": 1e-3})       # GP only
surrogate.recommend({"n_samples": 2200}, objective="val_loss")  # Sobol' search

• Internally: a single SurrogateModel over the joint regime_factors + factors input space, so the regime descriptors are additional input dimensions rather than separate models per bucket.
• recommend() samples n_candidates design configs via scrambled Sobol' (or accepts an explicit candidates= pool) and returns the best surrogate prediction under mode in {"min", "max"}.
• Inherits the GP-only uncertainty contract from SurrogateModel.

Validation

• regime_factors must be non-empty and disjoint from factors.
• recommend() validates mode, objective, and non-empty candidate pools.

Docs

New docs/api/regime.md page in the mkdocs nav, right after Surrogate.

Tests

13 new tests covering validation, predict/predict_batch, GP uncertainty, RF uncertainty raising, recommend() tracking the optimum across regime values, mode inversion, and input validation. regime.py coverage: 100%. Suite: 307 passing, 99.5% project coverage.

Composes with

• Surrogate modeling: GP/RF interpolation from results table #82 (base surrogate) — RegimeSurrogate wraps a SurrogateModel.
• Study checkpointing and resumption #75 (checkpointing, when landed) — the surrogate can be refit incrementally as new trials accumulate.