malkhut(optim): vectorized reward + Ray parallel eval + VBT post-analysis

1. Vectorized reward path (cwm/core.py):
   - Wired up existing compute_reward_vectorized from numba_core (was unused!)
   - Eliminates FeatureVector dict allocation + Python dict lookups on hot path
   - Numba path used when _HAS_NUMBA=True, Python fallback otherwise
   - Bit-identical: same math operations, just via numba JIT

2. Ray-based parallel eval (training/ray_eval.py):
   - Industrial multi-core execution via Ray (used by OpenAI/Anyscale)
   - ray.put() stores params/scenarios in shared object store (no pickle per worker)
   - Each worker: own CWM + planner, zero shared state, no races
   - Bit-identical: same seed + same params = same results regardless of worker count
   - PolicyEvaluator.evaluate_candidate: new use_ray=True parameter

3. VBT post-analysis (training/vbt_analysis.py):
   - episodes_to_pnl_array, episodes_to_metrics (Sharpe, Sortino, VaR, win_rate, etc.)
   - cross_asset_comparison, parameter_sensitivity
   - format_metrics for human-readable output
   - Analysis tool only — runs AFTER engine produces results

4. numba_core.py: added missing 'import math' for compute_reward_vectorized

13 new tests: vectorized reward bit-identity, Ray determinism, Ray result fields,
VBT metrics structure, cross-asset comparison, parameter sensitivity, edge cases.
Total: 1178 tests, 50 files, all green, zero regressions.

MALKHUT/malkhut/cwm/core.py

@@ -46,6 +46,7 @@ try:
         round_tick as _nb_round_tick,
         round_lot as _nb_round_lot,
         clip_lots as _nb_clip_lots,
+        compute_reward_vectorized,
     )
     _HAS_NUMBA = True
 except ImportError:
@@ -587,6 +588,39 @@ class MinimalCryptoLOBCWM:
         next_state: MarketWorldState,
         params: FulfilmentPolicyParams,
     ) -> float:
+        if _HAS_NUMBA:
+            # Fast path: numba-optimized reward computation
+            # Extract values directly — avoids FeatureVector dict allocation
+            b = next_state.book
+            path = next_state.trade_path
+
+            pnl = path.pnl_bps if path else 0.0
+            toxicity = path.orderflow_toxicity if path else 0.0
+            churn = path.queue_churn_score if path else 0.0
+            time_in_loss = path.time_in_loss_s if path else 0.0
+            spread_bps = b.spread_bps if b.bids and b.asks else 0.0
+
+            inv_risk = self._inventory_risk(next_state)
+            tail_risk = self._tail_risk_proxy(next_state)
+
+            is_maker = (action.order_type and
+                       action.order_type.value in ("POST_ONLY", "LIMIT"))
+            is_cross = action.kind.value == "CROSS_SPREAD"
+            is_cancel = action.kind.value in ("CANCEL", "CANCEL_REPLACE")
+
+            return compute_reward_vectorized(
+                pnl, toxicity, churn, time_in_loss, spread_bps,
+                inv_risk, tail_risk,
+                params.w_expected_pnl, params.w_adverse_selection,
+                params.w_inventory_risk, params.w_tail_loss, params.w_time_decay,
+                is_maker, prev_state.venue.maker_fee_bps,
+                is_cross, prev_state.venue.taker_fee_bps,
+                is_cancel, params.adverse_toxicity_cancel_threshold,
+                params.queue_churn_cancel_threshold,
+                params.w_queue_priority, params.w_adverse_selection,
+            )
+
+        # Fallback: Python path (no numba)
         fv = self.feature_extractor.extract(next_state).values
 
         pnl = fv.get("pnl_bps", 0.0)