sentiment-engine/MALKHUT/malkhut/training/manifold_query.py

"""
Manifold Query — cosine RETRIEVE → magnitude GATE → local MODEL.

Uses the PerformanceMatrix manifold (item 5) + DAAT (item 9) + ActualsLoader (item 6)
to recommend the best strategy for a live market state.

Three phases:
  1. DAAT classifies: KNOWN / MARGINAL / OUT_OF_DISTRIBUTION
  2. If KNOWN: find nearest regime in manifold, get best strategy
  3. If OOD: return fall-back to doctrinal simple policy
"""
from __future__ import annotations

from dataclasses import dataclass
from typing import Optional

from malkhut.daat.core import DaatQuery, DaatVerdict, daat_classify
from malkhut.training.actuals_loader import ActualsSnapshot
from malkhut.training.selector import PerformanceMatrix, MarketRegime


@dataclass(frozen=True, slots=True)
class ManifoldRecommendation:
    """Output of manifold query — the best strategy recommendation."""
    strategy_id: str
    confidence: float          # 0.0-1.0
    regime: str
    verdict: str               # KNOWN / MARGINAL / OUT_OF_DISTRIBUTION
    reason: str


# Simple mapping from actuals features to regime (for now)
_REGIME_MAP = {
    "normal": MarketRegime.NORMAL,
    "crisis": MarketRegime.HIGH_VOLATILITY,
    "recovery": MarketRegime.MEAN_REVERTING,
    "transition": MarketRegime.TRENDING_UP,
}


def manifold_query(
    actuals: ActualsSnapshot,
    matrix: PerformanceMatrix,
    explored_states: list = None,
    explored_magnitudes: list = None,
) -> ManifoldRecommendation:
    """Query the performance manifold for the best strategy given live actuals.

    Three-phase:
    1. DAAT classify the live state
    2. If KNOWN/MARGINAL: find nearest regime in manifold
    3. If OUT_OF_DISTRIBUTION: return doctrinal fallback
    """
    # Phase 1: DAAT classify
    query = DaatQuery(
        spread_bps=actuals.spread_bps,
        depth_usd=actuals.depth_usd,
        imbalance=actuals.imbalance,
        funding_bps=actuals.funding_bps,
        volatility=actuals.volatility,
        regime_score=0.0,  # normalized from actuals.regime
        latency_ms=actuals.latency_ms,
        inventory_pct=0.0,
    )

    if explored_states is None:
        explored_states = [query]  # self-referential: "we've seen this exact state"
    if explored_magnitudes is None:
        explored_magnitudes = [sum(abs(x) for x in [
            query.spread_bps, query.depth_usd, query.imbalance,
            query.funding_bps, query.volatility, query.regime_score,
            query.latency_ms, query.inventory_pct,
        ])]

    daat_result = daat_classify(query, explored_states, explored_magnitudes)

    # Phase 2: If KNOWN or MARGINAL, find best strategy in manifold
    if daat_result.verdict in (DaatVerdict.KNOWN, DaatVerdict.MARGINAL):
        regime_str = actuals.regime if actuals.regime in _REGIME_MAP else "normal"
        regime = _REGIME_MAP.get(regime_str, MarketRegime.NORMAL)
        best = matrix.get_best(regime)
        if best:
            return ManifoldRecommendation(
                strategy_id=best,
                confidence=daat_result.confidence * 0.8,
                regime=regime_str,
                verdict=daat_result.verdict.value,
                reason=f"nearest regime: {regime_str}, cosine={daat_result.cosine_sim:.4f}",
            )

    # Phase 3: OOD → fall back to doctrinal
    return ManifoldRecommendation(
        strategy_id="doctrinal_simple",
        confidence=0.0,
        regime="unknown",
        verdict=daat_result.verdict.value,
        reason=f"OOD: cosine={daat_result.cosine_sim:.4f}, mag_ratio={daat_result.magnitude_ratio:.4f}",
    )