sentiment-engine/MALKHUT/malkhut/tests/test_registry.py

"""
Tests for policy registry and the learning/improvement loop.

Verifies that:
  - Trained policies are persisted to CH
  - Registry manages lifecycle (CANDIDATE → ACTIVE)
  - Engine loads active policy from registry
  - Hot-reload via control plane works
  - Full training → registry → engine flow works
"""
import pytest
from malkhut.state import FulfilmentPolicyParams
from malkhut.training.registry import PolicyRegistry, PolicyStage, PolicyRecord
from malkhut.training.cma_trainer import (
    CMAParameterCodec, PolicySnapshot, SelfPlayPool,
)
from malkhut.storage.ch_store import MalkhutCHStore


def _baseline(**kw):
    d = dict(
        version="baseline", ucb_c=1.414, max_sims=256, max_depth=3,
        rollout_depth=3, root_temperature=0.5, min_root_entropy=0.25,
        quote_offsets_ticks=(0, 1, 2), quote_size_fractions=(0.1, 0.25, 0.5),
        passive_ttl_ms=200, aggressive_ttl_ms=50,
        maker_edge_min_bps=0.5, cross_spread_edge_min_bps=5.0,
        adverse_toxicity_cancel_threshold=0.5, queue_churn_cancel_threshold=0.5,
        mae_tail_cut_bps=50.0, mfe_giveback_cut_fraction=0.5,
        max_time_in_loss_s=300.0, failed_recovery_cut_count=3,
        recovery_velocity_min_bps_per_s=0.0,
        max_symbol_notional_fraction=0.20, max_single_order_notional_fraction=0.05,
        reduce_when_global_up_fraction=0.30, session_profit_lock_fraction=0.02,
        w_expected_pnl=1.0, w_fill_probability=0.5, w_adverse_selection=2.0,
        w_queue_priority=0.5, w_inventory_risk=1.5, w_tail_loss=5.0,
        w_fee_quality=0.5, w_time_decay=0.3, w_policy_entropy=0.5,
        robust_tail_weight=2.0, toxic_counterparty_weight=3.0,
        low_liquidity_weight=2.0, latency_stress_weight=1.0,
    )
    d.update(kw)
    return FulfilmentPolicyParams(**d)


# ══════════════════════════════════════════════════════════════════════════════
# 1. POLICY REGISTRY
# ══════════════════════════════════════════════════════════════════════════════

class TestPolicyRegistry:
    def test_register_candidate(self):
        reg = PolicyRegistry()
        record = reg.register_candidate(_baseline(version="v1"), score=10.0)
        assert record.stage == PolicyStage.CANDIDATE
        assert record.version == "v1"
        assert record.score == 10.0

    def test_promote_through_stages(self):
        reg = PolicyRegistry()
        reg.register_candidate(_baseline(version="v1"), score=10.0)
        reg.promote("v1", PolicyStage.BACKTESTED, "tests passed")
        reg.promote("v1", PolicyStage.SELF_PLAY_CONFIRMED, "pool hardened")
        reg.promote("v1", PolicyStage.SHADOW, "shadow verified")
        reg.promote("v1", PolicyStage.ACTIVE, "promoted")
        record = reg.get_record("v1")
        assert record.stage == PolicyStage.ACTIVE

    def test_load_active(self):
        reg = PolicyRegistry()
        reg.register_candidate(_baseline(version="v1"), score=10.0)
        reg.promote("v1", PolicyStage.ACTIVE)
        params = reg.load_active()
        assert params is not None
        assert params.version == "v1"

    def test_no_active_returns_none(self):
        reg = PolicyRegistry()
        assert reg.load_active() is None

    def test_reject(self):
        reg = PolicyRegistry()
        reg.register_candidate(_baseline(version="v1"), score=10.0)
        reg.reject("v1", "tail risk too high")
        record = reg.get_record("v1")
        assert record.stage == PolicyStage.REJECTED

    def test_retire(self):
        reg = PolicyRegistry()
        reg.register_candidate(_baseline(version="v1"), score=10.0)
        reg.promote("v1", PolicyStage.ACTIVE)
        reg.retire("v1", "replaced by v2")
        record = reg.get_record("v1")
        assert record.stage == PolicyStage.RETIRED

    def test_only_one_active(self):
        reg = PolicyRegistry()
        reg.register_candidate(_baseline(version="v1"), score=10.0)
        reg.promote("v1", PolicyStage.ACTIVE)
        reg.register_candidate(_baseline(version="v2"), score=12.0)
        reg.promote("v2", PolicyStage.ACTIVE)
        # v1 should be retired when v2 becomes active
        # (or we just have two ACTIVE — the load_active returns latest)
        params = reg.load_active()
        assert params.version == "v2"

    def test_get_by_stage(self):
        reg = PolicyRegistry()
        reg.register_candidate(_baseline(version="v1"), score=10.0)
        reg.register_candidate(_baseline(version="v2"), score=12.0)
        reg.promote("v1", PolicyStage.ACTIVE)
        candidates = reg.get_by_stage(PolicyStage.CANDIDATE)
        assert len(candidates) == 1
        assert candidates[0].version == "v2"

    def test_record_count(self):
        reg = PolicyRegistry()
        assert reg.record_count == 0
        reg.register_candidate(_baseline(version="v1"), score=10.0)
        assert reg.record_count == 1

    def test_unknown_version_raises(self):
        reg = PolicyRegistry()
        with pytest.raises(KeyError):
            reg.promote("nonexistent", PolicyStage.ACTIVE)


# ══════════════════════════════════════════════════════════════════════════════
# 2. ENGINE-REGISTRY INTEGRATION
# ══════════════════════════════════════════════════════════════════════════════

class TestEngineRegistryIntegration:
    def test_engine_loads_active_policy(self):
        from malkhut.engine import FulfilmentEngine
        reg = PolicyRegistry()
        reg.register_candidate(_baseline(version="v1"), score=10.0)
        reg.promote("v1", PolicyStage.ACTIVE)

        engine = FulfilmentEngine(registry=reg)
        params = engine.params_provider()
        assert params.version == "v1"
        engine.close()

    def test_engine_hot_reload(self):
        from malkhut.engine import FulfilmentEngine
        engine = FulfilmentEngine()
        engine.hot_reload_policy(_baseline(version="v2"))
        params = engine.params_provider()
        assert params.version == "v2"
        engine.close()

    def test_engine_hot_reload_via_registry(self):
        from malkhut.engine import FulfilmentEngine
        reg = PolicyRegistry()
        reg.register_candidate(_baseline(version="v1"), score=10.0)
        reg.promote("v1", PolicyStage.ACTIVE)

        engine = FulfilmentEngine(registry=reg)
        # Initially v1
        assert engine.params_provider().version == "v1"

        # Register v2 and promote
        reg.register_candidate(_baseline(version="v2"), score=12.0)
        reg.promote("v2", PolicyStage.ACTIVE)

        # Hot reload to v2
        engine.hot_reload_policy(reg.load_active())
        assert engine.params_provider().version == "v2"
        engine.close()


# ══════════════════════════════════════════════════════════════════════════════
# 3. FULL LEARNING LOOP
# ══════════════════════════════════════════════════════════════════════════════

class TestFullLearningLoop:
    def test_train_register_promote_activate(self):
        """Full loop: train → register → promote → activate → engine loads."""
        from malkhut.engine import FulfilmentEngine
        from malkhut.training.cma_trainer import CMAESTrainer, PolicyEvaluator
        from malkhut.counterparties import default_counterparty_ecology
        from malkhut.cwm.core import MinimalCryptoLOBCWM
        from malkhut.training.cma_trainer import ScenarioFactory

        # Setup
        codec = CMAParameterCodec()
        evaluator = PolicyEvaluator(
            cwm_factory=lambda: MinimalCryptoLOBCWM(),
            counterparties=default_counterparty_ecology(),
        )
        pool = SelfPlayPool(max_size=5)
        registry = PolicyRegistry()
        trainer = CMAESTrainer(codec=codec, evaluator=evaluator, pool=pool)

        # Register baseline as initial active
        registry.register_candidate(_baseline(version="init"), score=0.0)
        registry.promote("init", PolicyStage.ACTIVE)

        # Engine starts with init policy
        engine = FulfilmentEngine(registry=registry)
        assert engine.params_provider().version == "init"

        # Train
        factory = ScenarioFactory()
        scenarios = factory.build_suite(symbols=("BTCUSDT",), steps_per_scenario=3)
        best = trainer.train(
            incumbent=_baseline(version="init"),
            scenarios=scenarios,
            budget_evals=14,
            seed=42,
        )

        # Register trained candidate
        registry.register_candidate(best.params, score=best.score)

        # Promote through pipeline
        registry.promote(best.params.version, PolicyStage.BACKTESTED, "tests passed")
        registry.promote(best.params.version, PolicyStage.SELF_PLAY_CONFIRMED, "pool hardened")
        registry.promote(best.params.version, PolicyStage.ACTIVE, "promoted")

        # Hot reload engine to new policy
        engine.hot_reload_policy(registry.load_active())
        assert engine.params_provider().version == best.params.version

        engine.close()