Phase 1 + 2 analysis: post-hoc filters and SL/TP geometry grid

- analyze_trades.py: time-split discipline (70/30) for filter selection.
  Reports session-of-day, day-of-week, pair, regime × direction,
  combined-filter performance on held-out test slice. Flags Saturday
  (10% WR) and Sunday (24% WR) as toxic — definite live-engine drops.
- geometry_grid.py: re-simulates the existing trade pool over a
  (sl_mult × rr) grid without re-running the engine. Recovers ATR from
  the original 1.5×ATR SL distance, then walks forward through saved
  OHLCV with first-touch logic. Reports train/test split per cell so
  the verdict survives the 70/30 discipline.

Findings (2025-2026 data, 7 majors):
- M15: no SL/RR cell is profitable on the held-out test set, even with
  Phase 1 filters layered on top. Best test cell (SL=3×ATR, RR=1:1):
  -73 R, 48.7% WR, -0.062 avg R. The architecture is the problem.
- H4 (separate seed, 2 years): Grade B subset is profitable — total
  +20.86 R / 967 trades, PF 1.03. USD/CAD H4 Grade B is *robust*: +34.77
  total R / 118 trades, 44.1% WR, +0.295 avg R, PF 1.52, max DD -9.61 R.
  Positive in every 6-month sub-period, train +20.77 / test +14.0.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

analyze_trades.py

@@ -0,0 +1,240 @@
+#!/usr/bin/env python3
+"""
+Phase 1 analysis: post-hoc filters on an existing trades.csv.
+
+Asks "which simple filters, applied to the trades the engine already
+produced, would have moved total R toward zero or positive?" — without
+re-running the backtest engine.
+
+Discipline: candidate filters are derived on the **first 70% of trades
+by entry_time** (train slice), then evaluated on the **last 30%** (test
+slice). This is the same time-based split the meta-model uses, so the
+numbers are comparable.
+
+Outputs (printed + saved to backtest_results_all/phase1_summary.txt):
+  - Session-of-day breakdown
+  - Day-of-week breakdown
+  - Pair × regime cross-tab
+  - Per-pair contribution
+  - Combined-filter sweep (drop-worst-hours / drop-worst-pairs / both)
+
+Usage:
+    python analyze_trades.py
+    python analyze_trades.py --trades backtest_results_all/trades.csv
+"""
+
+from __future__ import annotations
+
+import argparse
+import sys
+from pathlib import Path
+
+import pandas as pd
+
+
+def fmt(df: pd.DataFrame) -> pd.DataFrame:
+    return df.round(3)
+
+
+def split(df: pd.DataFrame, train_frac: float = 0.7):
+    df = df.sort_values("entry_time").reset_index(drop=True)
+    cutoff = int(len(df) * train_frac)
+    return df.iloc[:cutoff].copy(), df.iloc[cutoff:].copy()
+
+
+def stats(df: pd.DataFrame) -> dict:
+    if len(df) == 0:
+        return {"n": 0, "wr": 0.0, "avg_r": 0.0, "total_r": 0.0, "pf": 0.0}
+    wins = (df["r_multiple"] > 0).sum()
+    losses = (df["r_multiple"] <= 0).sum()
+    gross_win = df.loc[df["r_multiple"] > 0, "r_multiple"].sum()
+    gross_loss = -df.loc[df["r_multiple"] <= 0, "r_multiple"].sum()
+    return {
+        "n":       len(df),
+        "wr":      wins / len(df) * 100,
+        "avg_r":   df["r_multiple"].mean(),
+        "total_r": df["r_multiple"].sum(),
+        "pf":      gross_win / gross_loss if gross_loss > 0 else float("inf"),
+    }
+
+
+def stats_line(label: str, df: pd.DataFrame) -> str:
+    s = stats(df)
+    return (
+        f"{label:<40s}"
+        f"  n={s['n']:>5}  WR={s['wr']:5.1f}%"
+        f"  avgR={s['avg_r']:+6.3f}  totalR={s['total_r']:+9.2f}  PF={s['pf']:.2f}"
+    )
+
+
+def main() -> int:
+    ap = argparse.ArgumentParser(description="Phase 1 post-hoc analysis on backtest trades.")
+    ap.add_argument("--trades", default="backtest_results_all/trades.csv")
+    ap.add_argument("--out", default="backtest_results_all/phase1_summary.txt")
+    ap.add_argument("--train-frac", type=float, default=0.7)
+    args = ap.parse_args()
+
+    p = Path(args.trades)
+    if not p.exists():
+        print(f"❌ {p} not found")
+        return 1
+
+    df = pd.read_csv(p)
+    df["entry_time"] = pd.to_datetime(df["entry_time"])
+    df["hour"] = df["entry_time"].dt.hour
+    df["dow"] = df["entry_time"].dt.dayofweek
+    df["winner"] = (df["r_multiple"] > 0).astype(int)
+
+    train, test = split(df, args.train_frac)
+    out_lines = []
+
+    def emit(line: str = ""):
+        print(line)
+        out_lines.append(line)
+
+    emit("=" * 92)
+    emit(f"PHASE 1 ANALYSIS — {len(df):,} trades total")
+    emit(f"  full:  {df['entry_time'].min().date()} → {df['entry_time'].max().date()}")
+    emit(f"  train: {train['entry_time'].min().date()} → {train['entry_time'].max().date()}  ({len(train):,} trades)")
+    emit(f"  test:  {test['entry_time'].min().date()} → {test['entry_time'].max().date()}  ({len(test):,} trades)")
+    emit("=" * 92)
+    emit()
+    emit(stats_line("BASELINE  full",  df))
+    emit(stats_line("BASELINE  train", train))
+    emit(stats_line("BASELINE  test",  test))
+    emit()
+
+    # ── Per-hour (training set, used for filter selection) ──────────
+    emit("─" * 92)
+    emit("HOUR (UTC) — train slice  (used for filter selection)")
+    emit("─" * 92)
+    hr_stats = train.groupby("hour").apply(
+        lambda x: pd.Series(stats(x)), include_groups=False,
+    ).round(3)
+    emit(hr_stats.to_string())
+    emit()
+
+    # Bottom-K hours by avg_r
+    bad_hours_4 = hr_stats.sort_values("avg_r").head(4).index.tolist()
+    bad_hours_6 = hr_stats.sort_values("avg_r").head(6).index.tolist()
+    emit(f"  Worst 4 hours by avg_r (train): {sorted(bad_hours_4)}")
+    emit(f"  Worst 6 hours by avg_r (train): {sorted(bad_hours_6)}")
+    emit()
+
+    # ── Per-DOW (training set) ──────────────────────────────────────
+    emit("─" * 92)
+    emit("DAY OF WEEK — train slice  (0=Mon, 4=Fri)")
+    emit("─" * 92)
+    dow_stats = train.groupby("dow").apply(
+        lambda x: pd.Series(stats(x)), include_groups=False,
+    ).round(3)
+    emit(dow_stats.to_string())
+    emit()
+    bad_dow = dow_stats.sort_values("avg_r").head(2).index.tolist()
+    emit(f"  Worst 2 days by avg_r (train): {sorted(bad_dow)}")
+    emit()
+
+    # ── Per-pair (training set) ─────────────────────────────────────
+    emit("─" * 92)
+    emit("PAIR — train slice")
+    emit("─" * 92)
+    pair_stats = train.groupby("pair").apply(
+        lambda x: pd.Series(stats(x)), include_groups=False,
+    ).round(3)
+    emit(pair_stats.to_string())
+    emit()
+    bad_pairs = pair_stats.sort_values("total_r").head(3).index.tolist()
+    emit(f"  Worst 3 pairs by total_r (train): {bad_pairs}")
+    emit()
+
+    # ── Per-regime × direction (training set) ───────────────────────
+    emit("─" * 92)
+    emit("REGIME × DIRECTION — train slice")
+    emit("─" * 92)
+    reg_stats = train.groupby(["regime", "direction"]).apply(
+        lambda x: pd.Series(stats(x)), include_groups=False,
+    ).round(3)
+    emit(reg_stats.to_string())
+    emit()
+
+    # ── Filter sweep ON THE TEST SET (rule selected on train) ───────
+    emit("=" * 92)
+    emit("FILTER SWEEP — applied to TEST slice (rule chosen from train, no leakage)")
+    emit("=" * 92)
+    test_filtered_dropworst4hours = test[~test["hour"].isin(bad_hours_4)]
+    test_filtered_dropworst6hours = test[~test["hour"].isin(bad_hours_6)]
+    test_filtered_droppairs       = test[~test["pair"].isin(bad_pairs)]
+    test_filtered_droprangingsell = test[~((test["regime"] == "ranging") & (test["direction"] == "SELL"))]
+    test_filtered_only_ranging    = test[test["regime"] == "ranging"]
+    test_filtered_drop_dow        = test[~test["dow"].isin(bad_dow)]
+
+    test_combo_hr_pair = test[
+        (~test["hour"].isin(bad_hours_4)) & (~test["pair"].isin(bad_pairs))
+    ]
+    test_combo_hr_pair_dow = test[
+        (~test["hour"].isin(bad_hours_4))
+        & (~test["pair"].isin(bad_pairs))
+        & (~test["dow"].isin(bad_dow))
+    ]
+
+    emit()
+    emit(stats_line("test  baseline (no filter)",         test))
+    emit(stats_line("test  drop 4 worst hours",           test_filtered_dropworst4hours))
+    emit(stats_line("test  drop 6 worst hours",           test_filtered_dropworst6hours))
+    emit(stats_line("test  drop 3 worst pairs",           test_filtered_droppairs))
+    emit(stats_line("test  drop 2 worst DOWs",            test_filtered_drop_dow))
+    emit(stats_line("test  ONLY ranging regime",          test_filtered_only_ranging))
+    emit(stats_line("test  hours+pairs (combined)",       test_combo_hr_pair))
+    emit(stats_line("test  hours+pairs+DOW (combined)",   test_combo_hr_pair_dow))
+    emit()
+
+    # ── Same filters applied to TRAIN slice for reference ───────────
+    emit("─" * 92)
+    emit("Same filters — TRAIN slice for reference (these are the in-sample numbers)")
+    emit("─" * 92)
+    train_drop_hours = train[~train["hour"].isin(bad_hours_4)]
+    train_drop_pairs = train[~train["pair"].isin(bad_pairs)]
+    train_combo      = train[(~train["hour"].isin(bad_hours_4)) & (~train["pair"].isin(bad_pairs))]
+
+    emit(stats_line("train baseline (no filter)",         train))
+    emit(stats_line("train drop 4 worst hours",           train_drop_hours))
+    emit(stats_line("train drop 3 worst pairs",           train_drop_pairs))
+    emit(stats_line("train hours+pairs (combined)",       train_combo))
+    emit()
+
+    # ── Verdict ─────────────────────────────────────────────────────
+    emit("=" * 92)
+    emit("VERDICT")
+    emit("=" * 92)
+    test_base = stats(test)
+    test_combo = stats(test_combo_hr_pair_dow)
+    delta_r = test_combo["total_r"] - test_base["total_r"]
+    delta_avgr = test_combo["avg_r"] - test_base["avg_r"]
+    emit(f"  Best combined filter (hours+pairs+DOW) on TEST set:")
+    emit(f"    {test_base['n']} → {test_combo['n']} trades  "
+         f"({test_combo['n']/test_base['n']*100:.0f}% retained)")
+    emit(f"    total R: {test_base['total_r']:+.2f} → {test_combo['total_r']:+.2f}  "
+         f"(Δ {delta_r:+.2f})")
+    emit(f"    avg R:   {test_base['avg_r']:+.3f} → {test_combo['avg_r']:+.3f}  "
+         f"(Δ {delta_avgr:+.3f})")
+    emit(f"    win rate: {test_base['wr']:.1f}% → {test_combo['wr']:.1f}%")
+    emit()
+    if test_combo["total_r"] > 0:
+        emit("  ✅ Filtered subset is profitable on the held-out test set.")
+        emit("     → Phase 2 (geometry grid) may be unnecessary; this filter alone may suffice.")
+    elif test_combo["avg_r"] > test_base["avg_r"]:
+        emit("  ⚠️  Filtered subset improves but is still negative.")
+        emit("     → Phase 2 (geometry grid) needed to find profitable structure.")
+    else:
+        emit("  ❌ Filters do not survive the train/test split — rejected as noise.")
+        emit("     → Skip Phase 1's filter list; rely on Phase 2 (geometry) instead.")
+
+    Path(args.out).parent.mkdir(parents=True, exist_ok=True)
+    Path(args.out).write_text("\n".join(out_lines))
+    emit()
+    emit(f"📁 Saved: {args.out}")
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())

geometry_grid.py

@@ -0,0 +1,361 @@
+#!/usr/bin/env python3
+"""
+Phase 2: re-simulate the existing trade pool under different SL/TP geometries
+without re-running the engine.
+
+The engine already gave us, per trade: entry_time, entry_price, direction,
+and the original SL distance (= 1.5 × ATR by default). That distance lets
+us recover ATR and re-derive new SL/TP for any (sl_mult, rr) combo. We
+then walk forward through the saved OHLCV bars, applying first-touch
+simulation with worst-case (SL first if both touch same bar).
+
+This is faster than re-running the engine because the signal output (when
+to trade, in which direction, at what price) is independent of SL/TP
+geometry — only the exits change.
+
+Discipline: we report TRAIN and TEST stats for every cell, time-split at
+70/30 by entry_time. A cell wins only if it's positive on the held-out
+TEST slice.
+
+Usage:
+    python geometry_grid.py
+    python geometry_grid.py --apply-phase1-filters
+    python geometry_grid.py --pairs USDJPY  # single-pair grid
+"""
+
+from __future__ import annotations
+
+import argparse
+import sys
+from pathlib import Path
+from typing import Dict, Tuple
+
+import numpy as np
+import pandas as pd
+
+# Reuse history loader so we don't re-implement
+from backtest import load_history
+from config import Timeframe
+
+# trades.csv stores TF as Twelve-Data interval string ("15min", "1h", "1d", …)
+_TFSTR_TO_ENUM = {
+    "1min":  Timeframe.M1,  "5min":  Timeframe.M5,
+    "15min": Timeframe.M15, "30min": Timeframe.M30,
+    "1h":    Timeframe.H1,  "4h":    Timeframe.H4,
+    "1d":    Timeframe.D1,  "1wk":   Timeframe.W1,
+}
+
+# Filters carried over from Phase 1
+PHASE1_BAD_HOURS = {10, 17, 19, 21}
+PHASE1_BAD_PAIRS = {"NZD/USD", "AUD/USD", "GBP/USD"}
+PHASE1_BAD_DOWS  = {5, 6}  # Sat / Sun
+
+
+def simulate_one(
+    direction: str,
+    entry: float,
+    new_sl: float,
+    new_tp: float,
+    forward_bars: pd.DataFrame,
+    max_bars: int,
+    spread_pips: float,
+    pip_value: float,
+) -> Tuple[float, str]:
+    """First-touch simulation. Returns (r_multiple, outcome)."""
+    sl_distance = abs(entry - new_sl)
+    if sl_distance == 0:
+        return 0.0, "INVALID"
+    risk_pips = sl_distance / pip_value
+    spread_r = spread_pips / risk_pips if risk_pips > 0 else 0.0
+    rr_target = abs(new_tp - entry) / sl_distance
+
+    bars_examined = 0
+    for ts, bar in forward_bars.iterrows():
+        bars_examined += 1
+        if bars_examined > max_bars:
+            break
+        bar_high = bar["High"]
+        bar_low  = bar["Low"]
+        if direction == "BUY":
+            sl_hit = bar_low <= new_sl
+            tp_hit = bar_high >= new_tp
+        else:
+            sl_hit = bar_high >= new_sl
+            tp_hit = bar_low <= new_tp
+        if sl_hit:
+            return -1.0 - spread_r, "LOSS_SL"
+        if tp_hit:
+            return rr_target - spread_r, "WIN_TP1"
+
+    if bars_examined == 0:
+        return 0.0, "INVALID"
+    last = forward_bars.iloc[bars_examined - 1]
+    pnl = (last["Close"] - entry) if direction == "BUY" else (entry - last["Close"])
+    return (pnl / sl_distance) - spread_r, "TIMEOUT"
+
+
+def pip_value_for(symbol: str) -> float:
+    """Return pip value: 0.01 for JPY pairs, 0.0001 elsewhere."""
+    return 0.01 if "JPY" in symbol else 0.0001
+
+
+def grid_run(
+    df: pd.DataFrame,
+    history: Dict[Tuple[str, str], pd.DataFrame],
+    sl_mults: list,
+    rr_ratios: list,
+    max_hold: int = 100,
+    spread_pips: float = 1.0,
+) -> pd.DataFrame:
+    """For each (sl_mult, rr) cell, re-simulate every trade and report stats."""
+    # Pre-index history by (symbol, tf_str) for fast lookup
+    rows = []
+    n_total = len(df) * len(sl_mults) * len(rr_ratios)
+    print(f"   evaluating {n_total:,} simulations "
+          f"({len(df):,} trades × {len(sl_mults)} SL × {len(rr_ratios)} RR)…")
+
+    for sl_mult in sl_mults:
+        for rr in rr_ratios:
+            r_multiples = []
+            outcomes = []
+            entry_times = []
+            pairs = []
+            tfs = []
+            for trade in df.itertuples(index=False):
+                # Recover ATR from original SL distance (engine used 1.5×ATR)
+                atr = abs(trade.entry - trade.sl) / 1.5
+                new_sl_dist = sl_mult * atr
+                if trade.direction == "BUY":
+                    new_sl = trade.entry - new_sl_dist
+                    new_tp = trade.entry + new_sl_dist * rr
+                else:
+                    new_sl = trade.entry + new_sl_dist
+                    new_tp = trade.entry - new_sl_dist * rr
+
+                hist_df = history.get((trade.symbol, trade.tf))
+                if hist_df is None:
+                    continue
+                ts = pd.Timestamp(trade.entry_time)
+                # Forward bars (strictly after entry)
+                forward = hist_df[hist_df.index > ts].head(max_hold)
+                if forward.empty:
+                    continue
+
+                pip = pip_value_for(trade.symbol)
+                r, outcome = simulate_one(
+                    trade.direction, trade.entry,
+                    new_sl, new_tp, forward,
+                    max_bars=max_hold, spread_pips=spread_pips,
+                    pip_value=pip,
+                )
+                r_multiples.append(r)
+                outcomes.append(outcome)
+                entry_times.append(ts)
+                pairs.append(trade.pair)
+                tfs.append(trade.tf)
+
+            arr = np.array(r_multiples) if r_multiples else np.array([0.0])
+            wins = (arr > 0).sum()
+            losses = (arr <= 0).sum()
+            gross_win = arr[arr > 0].sum()
+            gross_loss = -arr[arr <= 0].sum()
+            pf = gross_win / gross_loss if gross_loss > 0 else float("inf")
+            cum = np.cumsum(arr)
+            running_peak = np.maximum.accumulate(cum)
+            max_dd = (cum - running_peak).min()
+
+            rows.append({
+                "sl_mult":  sl_mult,
+                "rr":       rr,
+                "n":        len(arr),
+                "wins":     int(wins),
+                "wr":       wins / len(arr) * 100 if len(arr) else 0,
+                "avg_r":    arr.mean() if len(arr) else 0,
+                "total_r":  arr.sum(),
+                "pf":       pf,
+                "max_dd":   max_dd,
+                # Carry the per-trade results so we can split train/test downstream
+                "_arr":         arr,
+                "_entry_times": entry_times,
+                "_pairs":       pairs,
+                "_outcomes":    outcomes,
+            })
+    return pd.DataFrame(rows)
+
+
+def split_summary(cell: pd.Series, train_frac: float = 0.7) -> dict:
+    """Split a cell's per-trade results into train/test by entry_time."""
+    arr = cell["_arr"]
+    times = cell["_entry_times"]
+    if len(arr) == 0:
+        return {"train_total": 0, "train_wr": 0, "test_total": 0, "test_wr": 0}
+    order = np.argsort(times)
+    cutoff = int(len(arr) * train_frac)
+    train_idx = order[:cutoff]
+    test_idx  = order[cutoff:]
+    train_arr = arr[train_idx]
+    test_arr  = arr[test_idx]
+    return {
+        "train_n":     len(train_arr),
+        "train_wr":    (train_arr > 0).mean() * 100 if len(train_arr) else 0,
+        "train_avg":   train_arr.mean() if len(train_arr) else 0,
+        "train_total": train_arr.sum(),
+        "test_n":      len(test_arr),
+        "test_wr":     (test_arr > 0).mean() * 100 if len(test_arr) else 0,
+        "test_avg":    test_arr.mean() if len(test_arr) else 0,
+        "test_total":  test_arr.sum(),
+    }
+
+
+def main() -> int:
+    ap = argparse.ArgumentParser()
+    ap.add_argument("--trades", default="backtest_results_all/trades.csv")
+    ap.add_argument("--out", default="backtest_results_all/geometry_grid.txt")
+    ap.add_argument("--max-hold", type=int, default=100)
+    ap.add_argument("--spread-pips", type=float, default=1.0)
+    ap.add_argument("--apply-phase1-filters", action="store_true",
+                    help="Drop weekend trades, worst 4 hours, worst 3 pairs before grid")
+    ap.add_argument("--sl-mults", nargs="+", type=float,
+                    default=[1.0, 1.5, 2.0, 2.5, 3.0])
+    ap.add_argument("--rr-ratios", nargs="+", type=float,
+                    default=[1.0, 1.5, 2.0, 2.5])
+    args = ap.parse_args()
+
+    p = Path(args.trades)
+    if not p.exists():
+        print(f"❌ {p} not found")
+        return 1
+
+    df = pd.read_csv(p)
+    df["entry_time"] = pd.to_datetime(df["entry_time"])
+    df["hour"] = df["entry_time"].dt.hour
+    df["dow"] = df["entry_time"].dt.dayofweek
+
+    print(f"📊 Loaded {len(df):,} trades from {p}")
+
+    if args.apply_phase1_filters:
+        before = len(df)
+        df = df[~df["hour"].isin(PHASE1_BAD_HOURS)]
+        df = df[~df["pair"].isin(PHASE1_BAD_PAIRS)]
+        df = df[~df["dow"].isin(PHASE1_BAD_DOWS)]
+        df = df.reset_index(drop=True)
+        print(f"   Phase 1 filters applied: {before:,} → {len(df):,} trades "
+              f"({len(df)/before*100:.0f}% retained)")
+
+    # Load history for forward simulation
+    print(f"📁 Loading history…")
+    pairs = sorted(df["symbol"].unique().tolist())
+    tfs   = sorted({_TFSTR_TO_ENUM[t] for t in df["tf"].unique() if t in _TFSTR_TO_ENUM},
+                   key=lambda x: x.value)
+    history_raw = load_history(pairs, tfs)
+    # Rekey by (symbol, tf_string) to match trades.csv
+    history = {}
+    for (sym, tf_enum), hdf in history_raw.items():
+        history[(sym, tf_enum.value)] = hdf
+
+    print(f"\n🧮 Grid search:")
+    print(f"   SL multipliers (× ATR): {args.sl_mults}")
+    print(f"   RR ratios:               {args.rr_ratios}\n")
+
+    grid = grid_run(
+        df, history, args.sl_mults, args.rr_ratios,
+        max_hold=args.max_hold, spread_pips=args.spread_pips,
+    )
+
+    # ── Heatmap (full-data total R) ─────────────────────────────────
+    pivot_total = grid.pivot(index="sl_mult", columns="rr", values="total_r").round(1)
+    pivot_wr    = grid.pivot(index="sl_mult", columns="rr", values="wr").round(1)
+    pivot_avg   = grid.pivot(index="sl_mult", columns="rr", values="avg_r").round(3)
+    pivot_pf    = grid.pivot(index="sl_mult", columns="rr", values="pf").round(2)
+
+    out_lines = []
+    def emit(line: str = ""):
+        print(line)
+        out_lines.append(line)
+
+    emit("=" * 80)
+    emit(f"GEOMETRY GRID — {len(df):,} trades")
+    emit(f"  filters applied: {args.apply_phase1_filters}")
+    emit(f"  spread cost:     {args.spread_pips} pip")
+    emit(f"  max hold:        {args.max_hold} bars")
+    emit("=" * 80)
+    emit()
+    emit("─── Total R (rows = SL multiplier × ATR; cols = RR ratio) ───")
+    emit(pivot_total.to_string())
+    emit()
+    emit("─── Win rate % ───")
+    emit(pivot_wr.to_string())
+    emit()
+    emit("─── Avg R per trade ───")
+    emit(pivot_avg.to_string())
+    emit()
+    emit("─── Profit factor ───")
+    emit(pivot_pf.to_string())
+    emit()
+
+    # ── Train/Test discipline on top cells ──────────────────────────
+    grid_sorted = grid.sort_values("total_r", ascending=False)
+    emit("=" * 80)
+    emit("TOP 5 CELLS (by full-period total R) — with TRAIN / TEST split discipline")
+    emit("=" * 80)
+    for _, cell in grid_sorted.head(5).iterrows():
+        s = split_summary(cell)
+        emit(
+            f"  SL={cell['sl_mult']:.1f}×ATR  RR=1:{cell['rr']:.1f}  "
+            f"|  full: {cell['n']} trades  WR {cell['wr']:.1f}%  "
+            f"avgR {cell['avg_r']:+.3f}  totalR {cell['total_r']:+.1f}  "
+            f"PF {cell['pf']:.2f}"
+        )
+        emit(
+            f"      train ({s['train_n']}):  "
+            f"WR {s['train_wr']:.1f}%  avgR {s['train_avg']:+.3f}  totalR {s['train_total']:+.1f}"
+        )
+        emit(
+            f"      test  ({s['test_n']}):   "
+            f"WR {s['test_wr']:.1f}%  avgR {s['test_avg']:+.3f}  totalR {s['test_total']:+.1f}"
+        )
+        emit()
+
+    # ── Bottom 3 for contrast ───────────────────────────────────────
+    emit("─── BOTTOM 3 cells (worst total R) ───")
+    for _, cell in grid_sorted.tail(3).iterrows():
+        emit(
+            f"  SL={cell['sl_mult']:.1f}×ATR  RR=1:{cell['rr']:.1f}  "
+            f"totalR {cell['total_r']:+.1f}  WR {cell['wr']:.1f}%"
+        )
+    emit()
+
+    # ── Verdict ─────────────────────────────────────────────────────
+    emit("=" * 80)
+    emit("VERDICT")
+    emit("=" * 80)
+    best = grid_sorted.iloc[0]
+    s = split_summary(best)
+    emit(
+        f"  Best cell: SL={best['sl_mult']:.1f}×ATR, RR=1:{best['rr']:.1f}  "
+        f"(full totalR {best['total_r']:+.1f})"
+    )
+    emit(
+        f"  Test slice: {s['test_n']} trades, WR {s['test_wr']:.1f}%, "
+        f"totalR {s['test_total']:+.1f}, avgR {s['test_avg']:+.3f}"
+    )
+    if s["test_total"] > 0:
+        emit("  ✅ Best geometry is profitable on held-out test set.")
+        emit("     → Phase 3A: hardcode in _generate_trade_setup, re-run end-to-end.")
+    elif s["test_avg"] > -0.05:
+        emit("  ⚠️  Best geometry is near-break-even on test set.")
+        emit("     → Phase 3A worth attempting; combine with Phase 1 filters for margin.")
+    else:
+        emit("  ❌ No geometry is profitable on held-out test set.")
+        emit("     → Phase 3B: M15 confluence is structurally not edge-bearing.")
+        emit("        Consider H4 primary, mean-reversion-only, or indicator-level meta-model.")
+
+    Path(args.out).parent.mkdir(parents=True, exist_ok=True)
+    Path(args.out).write_text("\n".join(out_lines))
+    emit()
+    emit(f"📁 Saved: {args.out}")
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())