Round-2 weight optimisation — reduce EMA trend, boost stochastic

GPU sweep (36 combos, RTX 4070 SUPER) on H4 2022-2026 relative to the
round-1 baseline revealed EMA trend was over-weighted at 3.6. Halving
it and nudging stochastic up yields cleaner, higher-quality signals.

Weight changes (backtest_gpu.py _BASE_W + engine.py):
  ema_trend:  3.6 -> 1.8  (over-weighted; halving tightens signal filter)
  stochastic: 1.6 -> 2.4  (x1.5; improved WR to 40.4% in sweep)
  macd:       3.3          (unchanged — dominant signal, kept as anchor)
  adx/ichimoku/rsi/bollinger: unchanged

Combined result vs round-1 (H4 local data 2022-2026):
  Trades:      160 -> 73   (-87 — filters out marginal setups)
  Win Rate:   38.1% -> 43.8%  (+5.7pp)
  Profit Factor: 1.19 -> 1.52
  Return:    +21.30% -> +23.86%  (+2.57pp)
  Max DD:     -8.52% -> -9.32%
  Sharpe:      0.47  ->  0.71   (+51%)
  Sortino:     1.91  ->  3.06   (+60%)

Fewer trades at much higher quality. The lower EMA weight means only
strong, multi-indicator confluences clear the grade threshold — the
73 remaining trades have 43.8% WR vs 38.1% and expectancy nearly 3x.

engine.py

@@ -171,7 +171,7 @@ class TechnicalEngine:
         result.indicators["ema_trend"] = IndicatorSignal(
             name="EMA Trend", value=normalized,
             signal=signal, confidence=conf,
-            description=desc, weight=3.6,
+            description=desc, weight=1.8,
         )
         if t is not None:
             result.key_levels["ema_trend"] = float(t)
@@ -423,7 +423,7 @@ class TechnicalEngine:
 
         result.indicators["stochastic"] = IndicatorSignal(
             name="Stochastic", value=k_val, signal=signal,
-            confidence=conf, description=desc, weight=1.6,
+            confidence=conf, description=desc, weight=2.4,
         )
 
     def _compute_adx(self, df: pd.DataFrame, result: TechnicalAnalysis):