fix(strategies): align factory regime numbering with training labeler (#15)

[bradsmithmba]

Summary

StrategyFactory and the training-time RegimeType labeler (src/data/regime_labeler.py) used the same regime integers (0–7) to mean different things. The neural net is trained on labels from the labeler, then its predictions are interpreted by the factory. Because the two numberings were misaligned, the system recommended bull-market strategies in bear markets and vice versa across nearly every regime.

Closes #15.

The misalignment (before this PR)

Regime int	Labeler (training labels)	Factory (inference mapping)
0	BULL_TRENDING	Deep Bear Market ← inverted
1	BEAR_TRENDING	Bull Trending ← inverted
2	HIGH_VOLATILITY	High Volatility Trending
3	LOW_VOLATILITY	Uncertain/Transitional
4	SIDEWAYS_RANGING	Low Volatility/Sideways
5	RECOVERY	Moderate Bull/Sideways
6	DISTRIBUTION	Recovery Phase
7	CRISIS	High Volatility Uncertain
8	(never produced)	Extreme Volatility ← unreachable

When the labeler tagged a market BULL_TRENDING (0) and the model learned to output 0, the factory mapped 0 to "Deep Bear Market" and recommended LONG_PUT / BEAR_PUT_SPREAD / BEAR_CALL_SPREAD.

Changes

• src/strategies/factory.py

  • _build_regime_mappings() realigned to RegimeType semantics (0=BULL_TRENDING … 7=CRISIS)
  • Removed the unreachable regime 8 (RegimeDetector emits 8 classes, indices 0–7)
  • _assess_expected_performance() performance map realigned
  • _assess_risk_for_regime() high_risk_regimes corrected to [1, 2, 6, 7] (BEAR_TRENDING, HIGH_VOLATILITY, DISTRIBUTION, CRISIS)
  • Docstring now names RegimeType as the source of truth

• tests/strategies/test_factory.py

  • Fixed fixtures and tests that had encoded the inverted numbering as expected behavior (this is why the bug was never caught — the tests asserted the bug was correct)
  • Added test_regime_mappings_match_labeler_numbering: a regression guard that fails if the factory and labeler diverge again

• tests/strategies/test_integration.py

  • Corrected stale regime comments and condition data to match the canonical numbering

Why the bug survived

The pre-existing test_factory.py asserted that regime 0 should produce bearish strategies and regime 1 bullish — i.e. it validated the inverted mapping. Passing tests that encode a bug make the bug permanent. The new regression test ties the factory's directional output back to the documented RegimeType numbering so this cannot silently re-invert.

Required follow-up: retraining

This is a code fix, not a full remediation. Any model checkpoint trained before this change learned feature → wrong-regime-integer associations and must be discarded and retrained after merge. No checkpoints are committed to the repo, so there is nothing to delete here — this is a heads-up for anyone holding local/remote weights.

Two upstream issues should be resolved before a useful retrain, since they corrupt the training input itself:

• Critical: 7 of 11 volatility features are hardcoded constants with no live signal #2 — volatility features are hardcoded constants (5 of 11 dims)
• Architecture: Live recommendation path bypasses the ML pipeline entirely #5 — RegimeStateVector assembles 21D by default, but RegimeDetector expects 48D

Testing

$ pytest tests/strategies/ -q
141 passed

Verification was scoped to tests/strategies/ because the rest of the suite currently fails to collect due to unrelated, separately-filed issues (#1 missing src.data providers, #6 sys.exit at import, #7 missing gym). This PR touches none of those paths.

🤖 Generated with Claude Code