perf(backtest): vectorize engine + parallel LLM propose + multi-fold validation tool

- backtest/engine.py: state machine numpy invece di pd.iterrows()
  - 16.8x speedup su 2y (470ms -> 28ms), 11.3x su 7y (1744ms -> 154ms)
  - 7 parity test parametrici vs reference iterrows assicurano equivalenza
- orchestrator/run.py + run_phase1.py: --llm-concurrency N
  - ThreadPoolExecutor parallelizza hypothesis_agent.propose() per generazione
  - 5-8x speedup wall time GA loop (OpenRouter qwen-2.5 regge 6-10 concorrenti)
  - default 1 = backward-compat sequenziale
- scripts/validate_run.py: validation multi-fold standalone
  - prende run_id + top-K + N-folds expanding-window su dataset esteso (7y)
  - rivela overfitter mascherati da fitness IS alta (vedi
    phase1-extended-001: elite IS Sharpe 1.93 collassa OOS a -1.00)
  - ranking per robust_score = min(fitness_oos) su tutti i fold

Test 250/250 pass.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

scripts/run_phase1.py

@@ -114,6 +114,16 @@ def parse_args() -> argparse.Namespace:
             "Schema: {styles: {<name>: {directive: <testo>}}}"
         ),
     )
+    p.add_argument(
+        "--llm-concurrency",
+        type=int,
+        default=1,
+        help=(
+            "Numero di propose() LLM concorrenti per generazione (default 1 = "
+            "serial). 6-10 tipicamente accettati da OpenRouter qwen-2.5 senza "
+            "rate-limit; riduce wall time GA loop di 5-8x."
+        ),
+    )
     return p.parse_args()
 
 
@@ -187,6 +197,7 @@ def main() -> None:
         eval_oos_during_loop=args.eval_oos_during_loop,
         fitness_combined_alpha=args.fitness_combined_alpha,
         prompt_library=prompt_library,
+        llm_concurrency=args.llm_concurrency,
     )
 
     run_id = run_phase1(cfg, ohlcv=ohlcv, llm=llm)