Add market regime detection for conditional model selection

[Copilot]

Description

Adds three regime signals as first-class ML features and a regime-routing trading strategy. The pipeline previously applied identical logic regardless of market conditions; now each bar is labelled with volatility, trend, and GMM-based regime state so models and strategies can condition on market context.

Changes in this pull request

• src/feature/regime.py (new) — three detectors exposed via classify_regime:

  • detect_volatility_regime — rolling vol bucketed into low/medium/high via quantile thresholds
  • detect_trend_regime — bull/bear/sideways from price vs 50d/200d MAs
  • detect_gmm_regime — Gaussian Mixture Model on (rolling mean return, rolling vol); components sorted so state 0 is always the most bearish; includes assert gmm.means_ is not None guard to satisfy static type checking

• src/feature/feature_engineering.py — _add_regime_features() appends Regime_Volatility, Regime_Trend, Regime_GMM to every feature matrix (shifted with all other indicators)

• src/feature/__init__.py — exports classify_regime and regime constants

• src/simulation/strategy/regime_aware.py (new) — RegimeAwareStrategy registered as "regime_aware", routes each bar to one of three sub-logics:

  • Bull → momentum (enter on positive pred, exit when pred turns negative)
  • Bear → conservative (close longs only, no new entries)
  • Sideways → mean-reversion (higher entry/exit threshold to suppress noise)

• src/simulation/strategy/__init__.py — imports the new strategy

• src/pipeline/base_pipeline.py — _log_regime_coverage logs train/test regime distributions per symbol; warns when the test split is single-regime

• tests/feature/test_regime.py — 28 unit tests covering all detectors, edge cases (tiny series fallback, state-count clamping), and FeatureEngineering integration

• tests/simulation/test_regime_strategy.py — 21 unit tests covering all strategy branches and a full simulation loop

[ch55secake]

@copilot fix the typecheck

[Copilot]

@copilot fix the typecheck

Fixed in 3a3a0b8. Added assert gmm.means_ is not None after fit_predict in detect_gmm_regime — the attribute is always populated post-fit but ty rightly flags it as potentially None without an explicit guard.