Files
siloqy/prod/clean_arch/dita/decision.py
Codex 2c9da8f592 PINK Phase 0: FET -$5,990 fix batch — leverage-free PnL, true fill prices, reconcile baseline anchors
Defects fix (FET -$5,990 replay, 2026-06-11):
- realized_pnl() and mark_price(): PnL = qty × Δprice, side-signed; no ×leverage inflation (was 3× every leg).
- BingX MARKET fill events carry true fill price (avgPrice/lastFillPrice), never the order's nominal price (protective bound ±20-25% from mark, poisoned PnL to -$5,990 on a +$164 round-trip).
- Fill routing by ORDER IDENTITY first, FSM state second — late entry-remainder fills during EXIT_WORKING no longer misclassify as exits.
- Entry basis = VWAP across entry fills, not last fill price.
- reconcile_from_slots / restore_state: re-anchor _last_settled_pnl / _slot_was_closed to adopted slot state (cross-restart double-book of carried PnL).
- ACCOUNT_UPDATE with wallet_balance=0 dropped (margin-only frames no longer zero e_available_margin).
- Foreign-fill skip on shared VST account (PRODGREEN collision filter).
- exec_router TTL: entry-requote venue-truth gate (recent own fill + live exchange position probes prevent double-entry).
- bingx_direct: openOrders fetched BEFORE positions (sequential ordering prevents dangerous tear → double-entries).
- Dual-leverage translation via map_internal_conviction_to_exchange_leverage() (strategy conviction → integer at-exchange leverage, bankers rounding).
- BLUE-parity alpha components wired: asset picker (IRP universe ranking) + alpha sizer (cubic-convex dynamic leverage, 0.5-8.0 range).
- ch_writer: date_time_input_format=best_effort on insert URLs; flush error logging at WARNING with counter.
- blue_parity.price_of(): hyphen-tolerant fallback (FET-USDT → FETUSDT).
- Fill test updated to incremental filled_size semantics (BingX WS lastFilledQty).
- Env-override base URLs, supervisord autorestart, per-asset DC histories, single-slot invariant, fill-attribution filter.

Co-authored-by: CommandCodeBot <noreply@commandcode.ai>
2026-06-11 20:53:49 +02:00

237 lines
10 KiB
Python

"""Pure decision engine."""
from __future__ import annotations
from dataclasses import dataclass
from datetime import datetime
from typing import Optional
from prod.clean_arch.tp_curve import compute_our_leverage, compute_soft_tp_pct
from prod.clean_arch.ports.data_feed import MarketSnapshot
from .contracts import Decision, DecisionAction, DecisionConfig, DecisionContext, TradePosition, TradeSide, TradeStage
@dataclass(frozen=True)
class _SnapshotFields:
price: float
vdiv: float
irp: float
ts: datetime
class DecisionEngine:
"""BLUE-compatible decision engine.
Decision only answers whether the system should enter/hold/exit.
It does not size orders or own exchange state.
"""
def __init__(self, config: Optional[DecisionConfig] = None, sizer: Optional[object] = None):
self.config = config or DecisionConfig()
# Optional BLUE-parity sizer (PinkAlphaSizer / AlphaBetSizer-shaped:
# calculate_size(capital=..., vel_div=...) → {fraction, leverage, ...}).
# None preserves the legacy linear-confidence sizing exactly — other
# consumers of this engine (main.py, trading_engine.py) are unaffected.
self.sizer = sizer
def decide(
self,
snapshot: MarketSnapshot,
context: DecisionContext,
position: Optional[TradePosition] = None,
) -> Decision:
fields = self._extract(snapshot)
if (
not snapshot.is_valid()
or fields.price <= 0
or not self._finite(fields.price)
or not self._finite(fields.vdiv)
or not self._finite(fields.irp)
):
return Decision(
timestamp=fields.ts,
decision_id=self._decision_id(snapshot.symbol, context.trade_seq),
asset=snapshot.symbol,
action=DecisionAction.HOLD,
side=TradeSide.FLAT,
reason="INVALID_SNAPSHOT",
confidence=0.0,
velocity_divergence=fields.vdiv,
irp_alignment=fields.irp,
reference_price=fields.price,
target_size=0.0,
leverage=1.0,
metadata={"policy_version": self.config.policy_version},
)
if position is not None and not position.closed:
return self._decide_exit(snapshot, position, context, fields)
return self._decide_entry(snapshot, context, fields)
def _decide_entry(self, snapshot: MarketSnapshot, context: DecisionContext, fields: _SnapshotFields) -> Decision:
if context.open_positions >= 1:
return self._hold(snapshot, context, fields, reason="CAPACITY_FULL")
if not self.config.allow_short:
return self._hold(snapshot, context, fields, reason="SHORT_DISABLED")
if fields.vdiv >= self.config.vel_div_threshold or fields.irp < self.config.min_irp_alignment:
return self._hold(snapshot, context, fields, reason="NO_SIGNAL")
# vol_ok gate — scan bridge marks low-volume periods; block ENTERs when absent
if snapshot.scan_payload and not snapshot.scan_payload.get("vol_ok", True):
return self._hold(snapshot, context, fields, reason="VOL_GATE")
sizing_meta: dict = {}
if self.sizer is not None:
# BLUE-parity sizing (SYSTEM BIBLE §6): cubic-convex dynamic
# leverage + alpha-layer fraction via AlphaBetSizer kernels.
size_result = self.sizer.calculate_size(capital=context.capital, vel_div=fields.vdiv)
leverage = float(size_result["leverage"])
fraction = float(size_result["fraction"])
target_exposure = context.capital * fraction * leverage
breakdown = size_result.get("breakdown") or {}
confidence = min(1.0, max(0.05, float(breakdown.get("strength_score", 0.0))))
sizing_meta = {
"eff_fraction": fraction,
"strength_score": breakdown.get("strength_score"),
"signal_bucket": breakdown.get("signal_bucket"),
"bucket_idx": size_result.get("bucket_idx"),
"sizing": "alpha_bet_sizer_cubic_v1",
}
else:
# Legacy DITAv2 formula. NOTE: an ENTER requires vdiv < threshold,
# so this confidence is always ≥ 1.0 → clamped → leverage pinned at
# max_leverage. Kept verbatim for non-PINK consumers.
confidence = min(1.0, max(0.05, abs(fields.vdiv / self.config.vel_div_threshold)))
leverage = min(self.config.max_leverage, max(1.0, 1.0 + confidence * (self.config.max_leverage - 1.0)))
target_exposure = context.capital * self.config.capital_fraction * leverage
target_size = target_exposure / fields.price if fields.price > 0 else 0.0
our_leverage = compute_our_leverage(notional=target_exposure, capital=context.capital)
tp_base_pct = float(self.config.fixed_tp_pct)
tp_effective_pct = compute_soft_tp_pct(tp_base_pct, our_leverage)
return Decision(
timestamp=fields.ts,
decision_id=self._decision_id(snapshot.symbol, context.trade_seq),
asset=snapshot.symbol,
action=DecisionAction.ENTER,
side=TradeSide.SHORT,
reason="STRUCTURAL_DISLOCATION",
confidence=confidence,
velocity_divergence=fields.vdiv,
irp_alignment=fields.irp,
reference_price=fields.price,
target_size=target_size,
leverage=leverage,
metadata={
"policy_version": self.config.policy_version,
"tp_base_pct": tp_base_pct,
"tp_effective_pct": tp_effective_pct,
"our_leverage": our_leverage,
"tp_curve": "soft_leverage_curve_v1",
**sizing_meta,
},
)
def _decide_exit(
self,
snapshot: MarketSnapshot,
position: TradePosition,
context: DecisionContext,
fields: _SnapshotFields,
) -> Decision:
action = DecisionAction.HOLD
reason = "HOLD"
position_notional = position.size * fields.price if fields.price > 0 else position.size * position.entry_price
our_leverage = compute_our_leverage(notional=position_notional, capital=context.capital)
tp_base_pct = float(self.config.fixed_tp_pct)
tp_effective_pct = compute_soft_tp_pct(tp_base_pct, our_leverage)
if position.side == TradeSide.SHORT:
tp_price = position.entry_price * (1.0 - tp_effective_pct)
if fields.price <= tp_price:
action = DecisionAction.EXIT
reason = "TAKE_PROFIT"
elif fields.price >= position.entry_price * (1.0 + (self.config.catastrophic_loss_pct / max(position.leverage, 1.0))):
action = DecisionAction.EXIT
reason = "CATASTROPHIC_LOSS"
elif position.bars_held >= self.config.max_hold_bars:
action = DecisionAction.EXIT
reason = "MAX_HOLD"
elif fields.vdiv >= 0.0:
action = DecisionAction.EXIT
reason = "MEAN_REVERSION"
if position.side == TradeSide.LONG:
tp_price = position.entry_price * (1.0 + tp_effective_pct)
if fields.price >= tp_price:
action = DecisionAction.EXIT
reason = "TAKE_PROFIT"
elif fields.price <= position.entry_price * (1.0 - (self.config.catastrophic_loss_pct / max(position.leverage, 1.0))):
action = DecisionAction.EXIT
reason = "CATASTROPHIC_LOSS"
elif position.bars_held >= self.config.max_hold_bars:
action = DecisionAction.EXIT
reason = "MAX_HOLD"
elif fields.vdiv <= 0.0:
action = DecisionAction.EXIT
reason = "MEAN_REVERSION"
return Decision(
timestamp=fields.ts,
decision_id=position.trade_id,
asset=position.asset,
action=action,
side=position.side,
reason=reason,
confidence=max(0.0, min(1.0, position.entry_irp_alignment)),
velocity_divergence=fields.vdiv,
irp_alignment=fields.irp,
reference_price=fields.price,
target_size=position.size,
leverage=position.leverage,
bars_held=position.bars_held,
stage=TradeStage.EXIT_REQUESTED if action == DecisionAction.EXIT else TradeStage.POSITION_UPDATED,
metadata={
"policy_version": self.config.policy_version,
"tp_base_pct": tp_base_pct,
"tp_effective_pct": tp_effective_pct,
"our_leverage": our_leverage,
"tp_curve": "soft_leverage_curve_v1",
},
)
def _hold(self, snapshot: MarketSnapshot, context: DecisionContext, fields: _SnapshotFields, reason: str) -> Decision:
return Decision(
timestamp=fields.ts,
decision_id=self._decision_id(snapshot.symbol, context.trade_seq),
asset=snapshot.symbol,
action=DecisionAction.HOLD,
side=TradeSide.FLAT,
reason=reason,
confidence=0.0,
velocity_divergence=fields.vdiv,
irp_alignment=fields.irp,
reference_price=fields.price,
target_size=0.0,
leverage=1.0,
metadata={"policy_version": self.config.policy_version},
)
@staticmethod
def _extract(snapshot: MarketSnapshot) -> _SnapshotFields:
ts = snapshot.timestamp if isinstance(snapshot.timestamp, datetime) else datetime.utcnow()
return _SnapshotFields(
price=float(snapshot.price or 0.0),
vdiv=float(snapshot.velocity_divergence or 0.0),
irp=float(snapshot.irp_alignment or 0.0),
ts=ts,
)
@staticmethod
def _decision_id(symbol: str, seq: int) -> str:
return f"{symbol}-D-{seq:012d}"
@staticmethod
def _finite(value: float) -> bool:
return value == value and value not in (float("inf"), float("-inf"))