Fix live/backtest parity, SELL-side losses, and indicator repairs (+42% total R)

Live pipeline:
- Drop still-forming candle before analysis (signals no longer repaint)
- Direction-aware MTF agreement + opposition gate (no trade against 2-of-3 TF consensus)
- SELL volatility gate: block shorts in low/normal vol regimes (config sell_vol_regimes)
- Divergence: 40-bar lookback, price/indicator swings must align within ±2 bars
  (fired on 0.5% of trades before, 29% after), trending weight 2.0→1.0
- RSI regime_aware by default; transitioning regime only signals at extremes
- ADX ±DM tie no longer nudges bearish; TP2 clamped beyond TP1

Backtester realism:
- Charge $0.30 round-trip spread (--spread), gap-through SL/TP fills at bar open
- Resolve trades on RAW price path (cache stores raw bars, cleaning runs on load)
- Weekend/session filters use bar-close decision time to match live wall clock

Validated per-step on M15 2025-03-30→2026-05-31 under identical pricing:
+46.9R → +66.8R, win rate 69.9% → 73.0%, SELL side -2.3R → +9.6R.
Tested-and-reverted: ROC weight removal (-13R), BB fade suppression (-11.6R),
candlestick de-weight (-2R) — composite thresholds are co-tuned with weight mass.

CLAUDE.md

@@ -39,10 +39,16 @@ telegram_bot.py  main.py
 **Signal Pipeline:** OHLCV → indicators → regime classification → regime-weighted composite score → MTF agreement check → news/event adjustment → diminishing penalty cap → final signal (STRONG_BUY/BUY/NEUTRAL/SELL/STRONG_SELL) → quality grade (A+/A/B/C/D/F) → trade setup if grade C+.
 
 **Regime-Adaptive Weights:** Market classified as `trending_up`, `trending_down`, `ranging`, or `transitioning` via ADX + EMA + ADX slope. Gold-specific weights per regime:
-- **Trending:** MACD (2.0), EMA (2.0), Divergence (2.0), Keltner (1.5), ROC (1.5)
+- **Trending:** MACD (2.0), EMA (2.0), Keltner (1.5), ROC (1.5), Divergence (1.0)
 - **Ranging:** Divergence (2.5), Keltner (2.0), RSI (1.5), BB (1.5), MACD (1.5)
 - **Transitioning:** Keltner (2.5), MACD (2.5), ROC (2.0), Divergence (2.0), EMA (1.5)
 
+ROC is directionally redundant with MACD/Keltner/EMA (0.8–0.94 correlation) but its weight is load-bearing: composite thresholds are calibrated with it — removing it backtested 13R worse. Don't zero it without retuning thresholds.
+
+**RSI (regime-aware, default):** In trends RSI buys pullback zones instead of fading strength (classic mode backtested worse for gold). In `transitioning` only extremes (≥70/≤30) signal.
+
+**Divergence:** 40-bar lookback; price and indicator swings must align within ±2 bars to count as one divergence observation.
+
 **Transitioning Regime:** ADX < 25 but ADX slope > 3.0 = trend emerging. Catches trend birth 5-10 bars before ADX crosses 25. Boosts quality grade for early entries.
 
 **Diminishing Penalty Cap:** Penalties sorted by magnitude, applied with exponential decay (1st at 100%, 2nd at 50%, 3rd at 25%). Hard cap at 30% total. Replaces old additive stacking that could gut 30-50% confidence.
@@ -99,6 +105,10 @@ python backtest/gpu_backtest.py --pair XAUUSD --days 365
 
 The `backtest/` directory contains `gpu_backtest.py` — a GPU-accelerated backtester (CuPy with NumPy fallback) that replays historical OHLCV through the real SignalAggregator. Pre-computes all indicator series including ROC, Keltner, ADX slope for 100x+ speedup. Uses parquet caching for data persistence.
 
+Fill realism: charges a round-trip spread per trade (`--spread`, default $0.30 for XAU), fills gap-through SL/TP at the bar open, and resolves trades against the RAW price path (the parquet cache stores raw bars; validation/cleaning runs on load — signals see cleaned data, trade resolution sees real spikes). Entry is at the next bar's open after the signal bar. Weekend/session filters evaluate bar-close (decision) time to match the live bot's wall clock.
+
+GPU note: CuPy needs the NVIDIA runtime pip wheels (`nvidia-cuda-runtime-cu12`, `nvidia-cuda-nvrtc-cu12`, `nvidia-curand-cu12`, `nvidia-nvjitlink-cu12`) or it silently falls back to NumPy — check the "Backend:" line in the run log. The bar-replay loop itself is single-threaded Python either way (~340 bars/s); GPU only accelerates indicator precompute and SL/TP scanning.
+
 ## Code Conventions
 
 - Python 3.12, type hints on public methods