v1.15.0 — simplify execution engine: remove SNR/EMA sigma, formula-based SL only

What's changed

Bug fixes

• Fix immediate StopLoss triggered by dynamic_resistance/support within
  0.3% of entry.
  SL now uses formula-based price only (entry × sl ratio).
• Fix ADX-based max_hold boost removed — was compounding position hold
  times unpredictably. max_hold fixed to 720h (30-day safety net).

Breaking changes

• build_dynamic_params() removed from GenericBacktestEngine.
  All models now use get_fixed_params().
• use_dynamic_params and use_ema_sigma flags removed from
  ModelRegistry and WalkForwardRunner.
• --no-ema-sigma CLI flag removed from qr-backtest.
• fit() in MdrsSdeCryptoAdapter now returns dict with keys
  trace, summary, estimates instead of a plain estimates dict.
  Params are saved as .pkl instead of .json.
• [parameter_scaling] section removed from backtest_settings.toml.

Internal changes

• SNR scaling (alpha/sigma_1) removed — clip ranges caused 41~48% of
  windows to hit boundaries, making SNR effectively constant.
• EMA sigma reference infrastructure removed
  (_precompute_sigma_reference, _precompute_reference_sigma).
• get_fixed_params() simplified — no vol_quality or max_hold scaling.
• Events loaded before model type check in backtest.py.