Community Scripts

# ╔══════════════════════════════════════════════════════════════╗
# ║                  STRATEGY REQUEST LOG                       ║
# ╚══════════════════════════════════════════════════════════════╝
# Generated        : 2026-05-25 02:29:29
# Model            : XGBoost
# Feature Eng.     : buy when price closes below the lower Bollinger Band(20,2) and RSI(14) < 35, exit at the middle band + Auto-add features: ON
# Signal / Entry   : —
# Optimization     : —
# Risk Mgmt        : —
# Risk Filter      : —
# ══════════════════════════════════════════════════════════════

# ============================================================
# Bollinger Band Mean-Reversion + RSI Filter (XGBoost, Sharpe)
# ============================================================

# SECTION 0 — IMPORTS & CONSTANTS
import numpy as np
import pandas as pd

# ── Inlined strategy_utils ──
"""
strategy_utils.py — Standard utility functions for generated strategies.
Claude imports these instead of writing boilerplate from scratch.
This ensures consistent behavior across all generated strategies.
"""
import numpy as np
import pandas as pd
from sklearn.preprocessing import LabelEncoder


# Max backtest window per timeframe. A finer timeframe over a longer window
# blows up the results dict / parquet load / Modal train time (the 2026-05-12
# OOM was a 1-min × multi-year sweep) — and a 1-min strategy gains nothing from
# 2 years of 1-min bars. Enforced HERE because every training path (UI / API /
# Modal) funnels through run_strategy → load_ohlc. Env-overridable so a future
# "max plan" / dedicated-server tier can lift it.
_TF_MAX_DAYS = {
    "1min":  30,
    "5min":  90,
    "15min": 365,
    "1h":    730,
}


def _fetch_ohlc_from_internal(symbol: str, tf: str, start: str, end: str):
    """Phase 3.2: fetch parquet bytes from Server A's /internal/ohlc endpoint
    instead of reading a local file. Used inside Modal containers / Mac worker
    pool (Phase 3.4) so every train sees the same source of truth as the chart.

    Returns: pd.DataFrame (parquet decoded), or raises on any failure so the
    caller can fall back / surface a clear error in the job.
    """
    import hashlib as _hashlib, hmac as _hmac, io as _io, os as _os
    import urllib.request as _ur, urllib.parse as _urp

    base = (_os.environ.get("QM_INTERNAL_OHLC_BASE") or "").rstrip("/")
    secret = (_os.environ.get("INTERNAL_WS_SECRET") or "").strip()
    if not base:
        raise RuntimeError("QM_INTERNAL_OHLC_BASE not set")
    if not secret:
        raise RuntimeError("INTERNAL_WS_SECRET not set")

    msg = f"{symbol}|{tf}|{start}|{end}".encode("utf-8")
    sig = _hmac.new(secret.encode("utf-8"), msg, _hashlib.sha256).hexdigest()

    qs = _urp.urlencode({
        "symbol": symbol, "tf": tf,
        "start": start, "end": end, "sig": sig,
    })
    url = f"{base}/internal/ohlc?{qs}"

    req = _ur.Request(url, headers={"User-Agent": "qm-worker/1.0"})
    with _ur.urlopen(req, timeout=30) as resp:
        if resp.status != 200:
            raise RuntimeError(f"/internal/ohlc returned {resp.status}")
        payload = resp.read()
    print(f"[load_ohlc:internal] {symbol} {tf} fetched {len(payload)} bytes", flush=True)
    return pd.read_parquet(_io.BytesIO(payload))


def _parse_symbol_tf_from_path(data_path: str):
    """Pull SYMBOL + TF out of a path like .../EURUSD_1min.parquet."""
    import os as _os, re as _re
    base = _os.path.basename(str(data_path))
    m = _re.match(r"^([A-Z]{6})_(\d+min|\d+h)\.parquet$", base)
    if not m:
        return None, None
    return m.group(1), m.group(2)


def load_ohlc(data_path, start_date="", end_date=""):
    """Load OHLC parquet, sort index, filter dates. Always returns consistent format.

    The lower bound is clamped per timeframe (see _TF_MAX_DAYS) — a request for
    more history than the cap silently starts later.

    Phase 3.2: when env QM_USE_INTERNAL_OHLC=="1", fetch over HTTP from
    Server A's /internal/ohlc endpoint instead of pd.read_parquet on a local
    file (which on Modal is a stale Volume snapshot). The endpoint applies the
    same day-cap, so the local cap-check below is a defensive no-op in that
    path. Flag defaults to "0" → unchanged behavior.

    Returns: (df, close, open_, high, low)
    """
    import os as _os, re as _re

    _use_internal = _os.environ.get("QM_USE_INTERNAL_OHLC", "0") == "1"
    if _use_internal:
        _sym, _tf = _parse_symbol_tf_from_path(data_path)
        if not _sym or not _tf:
            raise RuntimeError(
                f"QM_USE_INTERNAL_OHLC=1 but DATA_PATH basename does not match "
                f"SYMBOL_TF.parquet: {data_path}"
            )
        df = _fetch_ohlc_from_internal(_sym, _tf, start_date or "", end_date or "")
    else:
        df = pd.read_parquet(data_path)
    df.index = pd.to_datetime(df.index)
    df = df.sort_index()

    # Per-timeframe window cap (timeframe inferred from the parquet filename).
    _m = _re.search(r"_(\d+min|\d+h)\.parquet$", _os.path.basename(str(data_path)))
    _tf = _m.group(1) if _m else None
    _max_days = _TF_MAX_DAYS.get(_tf)
    if _max_days and _max_days > 0 and len(df):
        _env_override = _os.environ.get(f"QM_MAX_DAYS_{_tf.upper()}")
        if _env_override and _env_override.isdigit():
            _max_days = int(_env_override)
        try:
            _eff_end = pd.Timestamp(end_date) if end_date else df.index.max()
            _eff_end = min(_eff_end, df.index.max())
            _floor = _eff_end - pd.Timedelta(days=_max_days)
            _req_start = pd.Timestamp(start_date) if start_date else df.index.min()
            if _req_start < _floor:
                print(f"[load_ohlc] {_tf} backtest window capped to {_max_days}d: "
                      f"start {_req_start.date()} -> {_floor.date()}", flush=True)
                start_date = _floor
        except Exception as _e:
            print(f"[load_ohlc] window-cap check skipped ({_e})", flush=True)

    if start_date:
        df = df[df.index >= start_date]
    if end_date:
        df = df[df.index <= end_date]

    return df, df["close"], df["open"], df["high"], df["low"]


def make_target(close, horizon=4):
    """Create target: direction N bars ahead. Default 4 bars = 1 hour on 15-min data.

    Returns: target (pd.Series of -1, 0, 1)
    """
    return np.sign(close.shift(-horizon) - close)


def split_data(df, target, feature_cols, train_split=0.7, validation_date=""):
    """Train/test split. Handles both ratio and date-based splits.
    Drops NaN from target before splitting. Encodes labels to [0,1,2].

    Returns: dict with keys:
        X_train, X_test, y_train, y_test,
        y_train_enc, y_test_enc, enc,
        close_train, close_test,
        split_idx, split_dt, n_train, n_test
    """
    # Drop NaN from target
    mask = target.notna()
    df = df[mask].copy()
    target = target[mask]

    close = df["close"]

    # Build feature matrix
    X = df[feature_cols].copy()
    X = X.bfill().ffill()
    X = X.replace([np.inf, -np.inf], np.nan).fillna(0.0)

    # Split
    if validation_date:
        split_idx = len(df[df.index <= validation_date])
    else:
        split_idx = int(len(df) * train_split)

    split_idx = max(1, min(split_idx, len(df) - 1))

    X_train = X.iloc[:split_idx]
    X_test = X.iloc[split_idx:]
    y_train = target.iloc[:split_idx]
    y_test = target.iloc[split_idx:]
    close_train = close.iloc[:split_idx]
    close_test = close.iloc[split_idx:]

    split_dt = str(df.index[split_idx])

    # Label encoding — always fit on [-1, 0, 1]
    enc = LabelEncoder()
    enc.fit([-1, 0, 1])
    y_train_enc = enc.transform(y_train)
    y_test_enc = enc.transform(y_test)

    return {
        "df": df, "X_train": X_train, "X_test": X_test,
        "y_train": y_train, "y_test": y_test,
        "y_train_enc": y_train_enc, "y_test_enc": y_test_enc,
        "enc": enc,
        "close": close, "close_train": close_train, "close_test": close_test,
        "split_idx": split_idx, "split_dt": split_dt,
        "n_train": len(X_train), "n_test": len(X_test),
    }


def compute_overlays(close, df_index):
    """Compute BB and MA overlays on full dataset. Always consistent.

    Returns: (bb_dict, ma_dict)
    """
    bb_mid = close.rolling(20).mean()
    bb_std = close.rolling(20).std()
    bb_upper = bb_mid + 2 * bb_std
    bb_lower = bb_mid - 2 * bb_std

    ma50 = close.rolling(50).mean()
    ma100 = close.rolling(100).mean()
    ma200 = close.rolling(200).mean()

    def _safe(s):
        s = s.reindex(df_index).bfill().ffill()
        return [float(x) if (x is not None and not np.isnan(x) and not np.isinf(x)) else None
                for x in s.values]

    bb = {"upper": _safe(bb_upper), "mid": _safe(bb_mid), "lower": _safe(bb_lower)}
    ma = {"ma50": _safe(ma50), "ma100": _safe(ma100), "ma200": _safe(ma200)}

    return bb, ma


def run_backtest(signal, close, capital=10000, cost=2e-5):
    """Run backtest with transaction costs.
    Uses price-based trade returns (same as webapp _compute_trades).
    Signal 0 = hold (keep current position), not close.

    Returns: dict with equity, trade_returns, long_returns, short_returns, bar_returns
    """
    sig_arr = signal.values
    price_arr = close.values
    idx = signal.index
    n = len(price_arr)

    # Trade returns — price-based (matches webapp _compute_trades exactly)
    trade_returns = []
    long_returns = []
    short_returns = []
    trade_log = []
    last_dir = None
    entry_price = None
    entry_bar = None

    for i in range(n):
        s = sig_arr[i]
        c = price_arr[i]
        if s != 0.0 and s != last_dir:
            # Direction change — close previous trade, open new
            if last_dir is not None and entry_price is not None and entry_price != 0:
                ret = float(last_dir * (c - entry_price) / entry_price - cost)
                trade_returns.append(ret)
                if last_dir == 1:
                    long_returns.append(ret)
                else:
                    short_returns.append(ret)
                trade_log.append({
                    "type": "Buy" if last_dir == 1 else "Sell",
                    "entry_time": str(idx[entry_bar]),
                    "exit_time": str(idx[i]),
                    "entry_price": round(entry_price, 5),
                    "exit_price": round(c, 5),
                    "pnl": round(last_dir * (c - entry_price), 5),
                    "pnl_pct": round(ret * 100, 3),
                    "exit_reason": "signal",
                })
            entry_price = c
            entry_bar = i
            last_dir = s

    # Close last open trade
    if last_dir is not None and entry_price is not None and n > 0 and entry_price != 0:
        c = price_arr[-1]
        ret = float(last_dir * (c - entry_price) / entry_price - cost)
        trade_returns.append(ret)
        if last_dir == 1:
            long_returns.append(ret)
        else:
            short_returns.append(ret)
        trade_log.append({
            "type": "Buy" if last_dir == 1 else "Sell",
            "entry_time": str(idx[entry_bar]),
            "exit_time": str(idx[-1]),
            "entry_price": round(entry_price, 5),
            "exit_price": round(c, 5),
            "pnl": round(last_dir * (c - entry_price), 5),
            "pnl_pct": round(ret * 100, 3),
            "exit_reason": "end",
        })

    # Equity curve from trade returns
    cumret = 1.0
    equity_vals = np.full(n, float(capital))
    trade_idx = 0
    in_trade = False
    t_entry_price = None
    t_dir = None

    for i in range(n):
        s = sig_arr[i]
        c = price_arr[i]
        if s != 0.0 and s != t_dir:
            if t_dir is not None and t_entry_price is not None and t_entry_price != 0:
                t_ret = t_dir * (c - t_entry_price) / t_entry_price - cost
                cumret *= (1 + t_ret)
            t_entry_price = c
            t_dir = s
        equity_vals[i] = capital * cumret

    # Bar returns for Sharpe
    bar_returns = np.zeros(n)
    for i in range(1, n):
        if price_arr[i - 1] != 0 and last_dir is not None:
            bar_returns[i] = sig_arr[i - 1] * (price_arr[i] - price_arr[i - 1]) / price_arr[i - 1] if sig_arr[i - 1] != 0 else 0.0

    return {
        "equity": pd.Series(equity_vals, index=close.index),
        "trade_returns": trade_returns,
        "long_returns": long_returns,
        "short_returns": short_returns,
        "bar_returns": bar_returns,
        "trade_log": trade_log,
    }


def compute_trade_stats(trades, capital=10000):
    """Single source of truth for trade statistics.
    Every display path reads from this — no recomputation anywhere.
    All values are rounded and JSON-safe (no inf/nan).
    """
    if not trades:
        return {"n": 0, "wins": 0, "losses": 0, "wr": 0, "avg": 0,
                "best": 0, "worst": 0, "ret": 0, "np": 0, "mdd": 0,
                "pf": 0, "rr": 0, "expect": 0}
    w = [r for r in trades if r > 0]
    l = [r for r in trades if r < 0]
    cumret = 1.0
    for r in trades:
        cumret *= (1 + r)
    net_p = capital * (cumret - 1)
    # Max drawdown
    eq = np.cumprod([1.0] + [1 + r for r in trades])
    peak = np.maximum.accumulate(eq)
    mdd = float(((eq - peak) / peak).min()) if len(eq) > 1 else 0.0
    # Profit Factor
    gross_w = sum(w) if w else 0
    gross_l = abs(sum(l)) if l else 0
    pf = gross_w / gross_l if gross_l > 0 else (9999.0 if gross_w > 0 else 0)
    # Risk:Reward
    avg_w = float(np.mean(w)) if w else 0
    avg_l = abs(float(np.mean(l))) if l else 0
    rr = avg_w / avg_l if avg_l > 0 else (9999.0 if avg_w > 0 else 0)
    # Expectancy
    expect = net_p / len(trades)
    return {
        "n": len(trades), "wins": len(w), "losses": len(l),
        "wr": round(len(w) / len(trades), 4),
        "avg": round(float(np.mean(trades)), 6),
        "best": round(max(w), 6) if w else 0,
        "worst": round(min(l), 6) if l else 0,
        "ret": round(cumret - 1, 6),
        "np": round(net_p, 2),
        "mdd": round(mdd, 6),
        "pf": round(pf, 2),
        "rr": round(rr, 2),
        "expect": round(expect, 2),
    }


def compute_metrics(bt_result, close_test, capital=10000):
    """Compute all standard metrics from backtest result.
    Uses trade-level compounding (same as webapp _trade_stats) for accuracy.

    Returns: dict with total_ret, bh_ret, sharpe_strat, sharpe_bh, mdd, n_trades
    """
    equity = bt_result["equity"]
    trade_returns = bt_result["trade_returns"]

    # Total return — trade-level compounding (matches webapp)
    if trade_returns:
        cumret = 1.0
        for r in trade_returns:
            cumret *= (1 + r)
        total_ret = cumret - 1
    else:
        total_ret = 0.0

    # Buy and hold
    bh_equity = capital * (close_test / close_test.iloc[0])
    bh_ret = (bh_equity.iloc[-1] - capital) / capital if capital != 0 else 0.0

    # Sharpe ratio — trade-level (matches webapp: sqrt(252*26) annualization)
    if len(trade_returns) >= 2 and float(np.std(trade_returns)) > 0:
        sharpe_strat = float(np.mean(trade_returns) / np.std(trade_returns) * np.sqrt(252 * 26))
    else:
        sharpe_strat = 0.0

    bh_rets = bh_equity.pct_change().dropna()
    if len(bh_rets) > 1 and bh_rets.std() != 0:
        sharpe_bh = float((bh_rets.mean() / bh_rets.std()) * np.sqrt(252 * 24 * 4))
    else:
        sharpe_bh = 0.0

    # Max drawdown — trade-level (matches webapp)
    if trade_returns:
        eq = np.cumprod([1.0] + [1 + r for r in trade_returns])
        peak = np.maximum.accumulate(eq)
        mdd = float(((eq - peak) / peak).min()) if len(eq) > 1 else 0.0
    else:
        mdd = 0.0

    return {
        "total_ret": float(total_ret),
        "bh_ret": float(bh_ret),
        "sharpe_strat": float(sharpe_strat) if not np.isnan(sharpe_strat) else 0.0,
        "sharpe_bh": float(sharpe_bh) if not np.isnan(sharpe_bh) else 0.0,
        "mdd": float(mdd),
        "n_trades": len(trade_returns),
    }


# Diagnostics line/histogram series (equity / drawdown / rolling_acc / conf_hist)
# only feed the small Diagnostics charts — they're never used by the price chart
# or scroll-back. On a 1-min model trained over the (2.2-capped) window these are
# still ~30k points each; downsample to a visually-identical resolution before the
# dict leaves the trainer so it doesn't carry that into Server-A RAM / Postgres.
_RESULTS_SERIES_MAX = 5000


def _downsample_idx(n, cap=_RESULTS_SERIES_MAX):
    """Evenly-spaced index list spanning [0, n-1] (first+last always kept), or
    None when no downsampling is needed (n <= cap)."""
    if n <= cap:
        return None
    return np.unique(np.linspace(0, n - 1, cap).astype(int)).tolist()


def _take(arr, idx):
    """Subset a list by an index list (idx may be None → return arr unchanged)."""
    if idx is None or not isinstance(arr, list):
        return arr
    return [arr[i] for i in idx]


# trade_log / train_trade_log are lists of per-trade dicts (display-only — the
# Trade Log tab). They scale with TRADE count, not bar count, so the bar-window
# cap (Phase 2.2) doesn't bound them — a degenerate near-every-bar model can put
# 10k+ trade dicts in the blob (>3 MB). Cap each (independently — a small-N model
# keeps every trade) to the most-recent N, recording `*_total` + `*_truncated`
# so the true count is still reported. Real strategies have far fewer than
# _TRADE_LOG_MAX trades, so this only ever bites pathological models.
_TRADE_LOG_MAX = 5000


def _cap_trade_log(tl):
    """Return (capped_list, original_len, was_truncated)."""
    if not isinstance(tl, list) or len(tl) <= _TRADE_LOG_MAX:
        return tl, (len(tl) if isinstance(tl, list) else 0), False
    return tl[-_TRADE_LOG_MAX:], len(tl), True


def build_return_dict(split_result, bt_result, metrics, model, feature_cols,
                      signal_full, p_pos_test, p_neg_test, custom_figs=None,
                      bt_train_result=None, pre_stats=None):
    """Assemble the complete return dict. Handles ALL serialization.
    Never returns Timestamps, numpy arrays, or non-JSON types.

    Returns: JSON-safe dict with all required keys
    """
    df = split_result["df"]
    close = split_result["close"]
    close_test = split_result["close_test"]
    X_test = split_result["X_test"]
    y_test = split_result["y_test"]
    equity = bt_result["equity"]
    bar_returns = bt_result["bar_returns"]

    # OHLC
    ohlc_dates = [str(x) for x in df.index.tolist()]

    def _safe_list(arr):
        return [float(x) if (x is not None and not np.isnan(x) and not np.isinf(x)) else None
                for x in arr]

    # Overlays
    bb, ma = compute_overlays(close, df.index)

    # Buy and hold equity
    capital = equity.iloc[0] if len(equity) > 0 else 10000
    bh_equity = capital * (close_test / close_test.iloc[0])

    # Confusion matrix
    from sklearn.metrics import confusion_matrix
    pred_test = model.predict(X_test)
    y_test_arr = np.asarray(y_test)
    cm = confusion_matrix(y_test_arr, pred_test, labels=[-1, 0, 1])

    # Rolling accuracy
    sig_arr = signal_full.reindex(close_test.index).values
    correct = pd.Series((pred_test == y_test_arr).astype(float), index=X_test.index)
    active_test = pd.Series(sig_arr != 0, index=close_test.index) if len(sig_arr) == len(close_test) else pd.Series(True, index=close_test.index)
    correct_active = correct.where(active_test, other=np.nan)
    rolling_acc = correct_active.rolling(30, min_periods=1).mean()

    # Feature importance
    importances = model.feature_importances_
    fi_pairs = sorted(zip(feature_cols, importances), key=lambda x: x[1])[-15:]

    # Drawdown
    rolling_max = equity.cummax()
    drawdown = (equity - rolling_max) / rolling_max.replace(0, np.nan)
    drawdown = drawdown.fillna(0.0)

    # ── Downsample the Diagnostics-only series (see _downsample_idx) ──────────
    _eq_dates = [str(x) for x in close_test.index.tolist()]
    _eq_strat = _safe_list(equity.values)
    _eq_bh    = _safe_list(bh_equity.values)
    _eq_idx   = _downsample_idx(len(_eq_dates))
    _eq_dates, _eq_strat, _eq_bh = _take(_eq_dates, _eq_idx), _take(_eq_strat, _eq_idx), _take(_eq_bh, _eq_idx)

    _ra_dates = [str(x) for x in rolling_acc.index.tolist()]
    _ra_vals  = [float(x) if (not np.isnan(x) and not np.isinf(x)) else None for x in rolling_acc.values]
    _ra_idx   = _downsample_idx(len(_ra_dates))
    _ra_dates, _ra_vals = _take(_ra_dates, _ra_idx), _take(_ra_vals, _ra_idx)

    _dd_dates = [str(x) for x in drawdown.index.tolist()]
    _dd_vals  = _safe_list(drawdown.values)
    _dd_idx   = _downsample_idx(len(_dd_dates))
    _dd_dates, _dd_vals = _take(_dd_dates, _dd_idx), _take(_dd_vals, _dd_idx)

    _cp_pos = [float(x) for x in (p_pos_test.tolist() if hasattr(p_pos_test, 'tolist') else list(p_pos_test))]
    _cp_neg = [float(x) for x in (p_neg_test.tolist() if hasattr(p_neg_test, 'tolist') else list(p_neg_test))]
    _cp_pos = _take(_cp_pos, _downsample_idx(len(_cp_pos)))
    _cp_neg = _take(_cp_neg, _downsample_idx(len(_cp_neg)))

    # ── Trade logs — display-only (Trade Log tab); cap to most-recent N with a
    #    `_total` field so the true count is still reported (see _cap_trade_log).
    #    NB: ret_dist arrays are left FULL — a downstream path in callbacks.py
    #    recomputes n_trades/win-rate from len(ret_dist), so a sample would skew
    #    the displayed counts; they're small anyway and gzip handles them.
    _tl_test, _tl_test_n, _tl_test_tr = _cap_trade_log(bt_result.get("trade_log", []))
    _tl_tr,   _tl_tr_n,   _tl_tr_tr   = _cap_trade_log(bt_train_result.get("trade_log", []) if bt_train_result else [])

    return {
        "ohlc": {
            "dates": ohlc_dates,
            "open": _safe_list(df["open"].values),
            "high": _safe_list(df["high"].values),
            "low": _safe_list(df["low"].values),
            "close": _safe_list(df["close"].values),
        },
        "signals": {
            "dates": [str(x) for x in signal_full.index.tolist()],
            "values": [float(x) for x in signal_full.values],
        },
        "bb": bb,
        "ma": ma,
        "equity": {
            "dates": _eq_dates,
            "strategy": _eq_strat,
            "bh": _eq_bh,
        },
        "feature_importance": {
            "names": [p[0] for p in fi_pairs],
            "values": [float(p[1]) for p in fi_pairs],
        },
        "conf_matrix": cm.tolist(),
        "conf_hist": {
            "p_pos": _cp_pos,
            "p_neg": _cp_neg,
        },
        "rolling_acc": {
            "dates": _ra_dates,
            "values": _ra_vals,
        },
        "drawdown": {
            "dates": _dd_dates,
            "values": _dd_vals,
        },
        "ret_dist": [float(x) for x in bt_result["trade_returns"]],
        "ret_dist_long": [float(x) for x in bt_result["long_returns"]],
        "ret_dist_short": [float(x) for x in bt_result["short_returns"]],
        "train_ret_dist": [float(x) for x in bt_train_result["trade_returns"]] if bt_train_result else [],
        "train_ret_dist_long": [float(x) for x in bt_train_result["long_returns"]] if bt_train_result else [],
        "train_ret_dist_short": [float(x) for x in bt_train_result["short_returns"]] if bt_train_result else [],
        "trade_log": _tl_test,
        "train_trade_log": _tl_tr,
        "trade_log_total": _tl_test_n,
        "train_trade_log_total": _tl_tr_n,
        "trade_log_truncated": _tl_test_tr,
        "train_trade_log_truncated": _tl_tr_tr,
        **(pre_stats or {}),
        "metrics": metrics,
        "split_dt": split_result["split_dt"],
        "split_idx": int(split_result["split_idx"]),
        "n_train": int(split_result["n_train"]),
        "n_test": int(split_result["n_test"]),
        "feature_cols": list(feature_cols),
        "custom_figs": custom_figs or [],
    }


# ════════════════════════════════════════════════════════════════════════════
# STRATEGY FRAMEWORK v2 — Config-driven architecture
# Claude writes feature_engineering() + strategy_config(). Framework does rest.
# ════════════════════════════════════════════════════════════════════════════

import importlib

_MODEL_REGISTRY = {
    "XGBClassifier":              ("xgboost", "XGBClassifier"),
    "RandomForestClassifier":     ("sklearn.ensemble", "RandomForestClassifier"),
    "GradientBoostingClassifier": ("sklearn.ensemble", "GradientBoostingClassifier"),
    "LogisticRegression":         ("sklearn.linear_model", "LogisticRegression"),
    "ExtraTreesClassifier":       ("sklearn.ensemble", "ExtraTreesClassifier"),
    "AdaBoostClassifier":         ("sklearn.ensemble", "AdaBoostClassifier"),
}


def _build_model_from_config(config, X_train, y_train_enc):
    """Build, fit, and wrap a model from strategy_config dict."""
    model_type = config.get("model_type", "RandomForestClassifier")
    model_params = dict(config.get("model_params", {}))

    if model_type not in _MODEL_REGISTRY:
        raise ValueError(f"Unknown model_type '{model_type}'. Valid: {list(_MODEL_REGISTRY.keys())}")

    module_path, class_name = _MODEL_REGISTRY[model_type]
    mod = importlib.import_module(module_path)
    cls = getattr(mod, class_name)

    # XGBoost defaults
    if class_name == "XGBClassifier":
        model_params.setdefault("use_label_encoder", False)
        model_params.setdefault("eval_metric", "mlogloss")
        model_params.setdefault("tree_method", "hist")
        # Determinism > speed (2026-05-25). XGBoost hist with n_jobs=-1 is
        # NON-reproducible even with random_state set — the parallel histogram
        # gradient-sum order varies across threads, so the SAME code + data
        # gives a slightly different model (and backtest) every run. Forcing
        # single-thread makes training bit-reproducible so: (a) a user who
        # copies a strategy and reruns it gets identical numbers, (b) the
        # community "Live" score matches a redeploy, (c) "same code, different
        # result" support reports go away. Cost: single-threaded XGB (a few
        # seconds slower on large windows; hist is fast so it's minor). FORCED
        # (not setdefault) so the guarantee can't be silently broken by a
        # strategy passing n_jobs. Exact reproducibility holds within the
        # platform (pinned versions / same Modal image); a user's own machine
        # with different xgboost/numpy/CPU can still differ in low-order bits.
        model_params["n_jobs"] = 1

    # Common defaults
    model_params.setdefault("random_state", 42)

    from model_wrapper import ModelWrapper

    clf = cls(**model_params)
    clf.fit(X_train, y_train_enc)

    enc = LabelEncoder()
    enc.fit([-1, 0, 1])
    return ModelWrapper(clf, original_classes=enc.classes_, n_features=X_train.shape[1])


def _generate_signals(model, X, threshold):
    """Framework-owned signal generation. Deterministic threshold logic."""
    proba = model.predict_proba(X)
    classes = list(model.classes_)

    idx_pos = classes.index(1) if 1 in classes else None
    idx_neg = classes.index(-1) if -1 in classes else None

    p_pos = proba[:, idx_pos] if idx_pos is not None else np.zeros(len(X))
    p_neg = proba[:, idx_neg] if idx_neg is not None else np.zeros(len(X))

    signal_vals = np.zeros(len(X))
    signal_vals = np.where(p_pos >= threshold, 1.0, signal_vals)
    signal_vals = np.where(p_neg >= threshold, -1.0, signal_vals)
    # Both exceed: pick stronger
    both = (p_pos >= threshold) & (p_neg >= threshold)
    signal_vals[both] = np.where(p_pos[both] >= p_neg[both], 1.0, -1.0)

    return pd.Series(signal_vals, index=X.index), p_pos, p_neg


# ── Filter functions (all no-ops when config value is None) ──────────────

def _apply_direction_filter(signal, direction):
    """Zero out signals that don't match allowed direction."""
    if direction is None or direction == "both":
        return signal
    s = signal.copy()
    if direction == "long":
        s[s < 0] = 0.0
    elif direction == "short":
        s[s > 0] = 0.0
    return s


def _apply_session_filter(signal, index, session_hours):
    """Zero out signals outside session hours [start, end] UTC."""
    if session_hours is None:
        return signal
    s = signal.copy()
    start_h, end_h = session_hours[0], session_hours[1]
    hours = index.hour
    if start_h <= end_h:
        mask = (hours >= start_h) & (hours < end_h)
    else:  # wrap around midnight, e.g. [22, 6]
        mask = (hours >= start_h) | (hours < end_h)
    s[~mask] = 0.0
    return s


def _apply_atr_filter(signal, close, high, low, min_atr):
    """Zero out signals when NATR(14) is below threshold."""
    if min_atr is None:
        return signal
    hl = high - low
    hc = (high - close.shift(1)).abs()
    lc = (low - close.shift(1)).abs()
    tr = pd.concat([hl, hc, lc], axis=1).max(axis=1)
    atr14 = tr.ewm(com=13, adjust=False).mean()
    natr = atr14 / close.replace(0, np.nan)
    s = signal.copy()
    s[natr < min_atr] = 0.0
    return s


def _apply_trend_filter(signal, close, trend_filter):
    """Only allow signals aligned with trend. e.g. 'sma_50': longs above SMA, shorts below."""
    if trend_filter is None:
        return signal
    # Parse: "sma_50" → SMA with period 50
    parts = trend_filter.lower().replace("-", "_").split("_")
    if len(parts) >= 2 and parts[0] in ("sma", "ema"):
        period = int(parts[1])
    else:
        return signal  # unknown filter, skip
    if parts[0] == "sma":
        trend_line = close.rolling(period).mean()
    else:
        trend_line = close.ewm(span=period, adjust=False).mean()
    s = signal.copy()
    # Longs only above trend, shorts only below
    s[(s > 0) & (close < trend_line)] = 0.0
    s[(s < 0) & (close > trend_line)] = 0.0
    return s


# ── run_backtest_v2: framework-owned SL/TP/cooldown/position management ──

def run_backtest_v2(signal, close, high, low, config, capital=10000, cost=2e-5):
    """Backtest with SL/TP/cooldown/direction handling built into the engine.
    Unlike run_backtest (v1), this function handles position exits internally.

    Returns: same dict shape as run_backtest()
    """
    stop_loss = config.get("stop_loss")
    take_profit = config.get("take_profit")
    cooldown = config.get("cooldown", 0)
    on_opposite = config.get("on_opposite", "reverse")

    sig_arr = signal.values
    close_arr = close.values
    high_arr = high.values
    low_arr = low.values
    idx = signal.index
    n = len(close_arr)

    trade_returns = []
    long_returns = []
    short_returns = []
    trade_log = []
    equity_vals = np.full(n, float(capital))
    cumret = 1.0

    position = 0.0       # current direction: 1.0, -1.0, or 0.0 (flat)
    entry_price = None
    entry_bar = None      # index into arrays for entry time
    cooldown_remaining = 0

    def _log_trade(exit_bar, exit_px, ret, reason):
        trade_log.append({
            "type": "Buy" if position == 1.0 else "Sell",
            "entry_time": str(idx[entry_bar]),
            "exit_time": str(idx[exit_bar]),
            "entry_price": round(entry_price, 5),
            "exit_price": round(exit_px, 5),
            "pnl": round(position * (exit_px - entry_price), 5),
            "pnl_pct": round(ret * 100, 3),
            "exit_reason": reason,
        })

    for i in range(n):
        c = close_arr[i]
        h = high_arr[i]
        lo = low_arr[i]
        s = sig_arr[i]

        # 1. Check SL/TP if in trade
        if position != 0.0 and entry_price is not None:
            hit_sl = False
            hit_tp = False
            exit_price = None

            if position == 1.0:  # long
                if stop_loss is not None and lo <= entry_price * (1 - stop_loss):
                    hit_sl = True
                    exit_price = entry_price * (1 - stop_loss)
                elif take_profit is not None and h >= entry_price * (1 + take_profit):
                    hit_tp = True
                    exit_price = entry_price * (1 + take_profit)
            else:  # short
                if stop_loss is not None and h >= entry_price * (1 + stop_loss):
                    hit_sl = True
                    exit_price = entry_price * (1 + stop_loss)
                elif take_profit is not None and lo <= entry_price * (1 - take_profit):
                    hit_tp = True
                    exit_price = entry_price * (1 - take_profit)

            if hit_sl or hit_tp:
                ret = float(position * (exit_price - entry_price) / entry_price - cost)
                trade_returns.append(ret)
                if position == 1.0:
                    long_returns.append(ret)
                else:
                    short_returns.append(ret)
                _log_trade(i, exit_price, ret, "SL" if hit_sl else "TP")
                cumret *= (1 + ret)
                position = 0.0
                entry_price = None
                entry_bar = None
                cooldown_remaining = cooldown
                equity_vals[i] = capital * cumret
                continue

        # 2. Cooldown
        if cooldown_remaining > 0:
            cooldown_remaining -= 1
            equity_vals[i] = capital * cumret
            continue

        # 3. Signal processing
        if s != 0.0:
            if position == 0.0:
                # Open new trade
                position = s
                entry_price = c
                entry_bar = i
            elif s != position:
                # Opposite signal
                if on_opposite == "reverse":
                    # Close current + open opposite
                    ret = float(position * (c - entry_price) / entry_price - cost)
                    trade_returns.append(ret)
                    if position == 1.0:
                        long_returns.append(ret)
                    else:
                        short_returns.append(ret)
                    _log_trade(i, c, ret, "signal")
                    cumret *= (1 + ret)
                    position = s
                    entry_price = c
                    entry_bar = i
                else:  # close_only
                    # Close current, go flat
                    ret = float(position * (c - entry_price) / entry_price - cost)
                    trade_returns.append(ret)
                    if position == 1.0:
                        long_returns.append(ret)
                    else:
                        short_returns.append(ret)
                    _log_trade(i, c, ret, "close_only")
                    cumret *= (1 + ret)
                    position = 0.0
                    entry_price = None
                    entry_bar = None
                    cooldown_remaining = cooldown

        equity_vals[i] = capital * cumret

    # Close last open trade at final close
    if position != 0.0 and entry_price is not None and n > 0 and entry_price != 0:
        c = close_arr[-1]
        ret = float(position * (c - entry_price) / entry_price - cost)
        trade_returns.append(ret)
        if position == 1.0:
            long_returns.append(ret)
        else:
            short_returns.append(ret)
        _log_trade(n - 1, c, ret, "end")
        cumret *= (1 + ret)
        equity_vals[-1] = capital * cumret

    # Bar returns for Sharpe (approximate)
    bar_returns = np.zeros(n)
    for i in range(1, n):
        if close_arr[i - 1] != 0 and sig_arr[i - 1] != 0:
            bar_returns[i] = sig_arr[i - 1] * (close_arr[i] - close_arr[i - 1]) / close_arr[i - 1]

    return {
        "equity": pd.Series(equity_vals, index=close.index),
        "trade_returns": trade_returns,
        "long_returns": long_returns,
        "short_returns": short_returns,
        "bar_returns": bar_returns,
        "trade_log": trade_log,
    }


# ── run_strategy: the v2 orchestrator ────────────────────────────────────

def run_strategy(feature_fn, config_fn, data_path, start_date="", end_date="",
                 validation_date="", train_split=0.7, register_model_fn=None):
    """Config-driven strategy execution. Claude writes feature_fn + config_fn,
    framework does everything else.

    Returns: results dict (same format as webapp expects)
    """
    config = config_fn()

    # Auto-correct SL/TP if Claude passed percentage instead of decimal
    for _key in ("stop_loss", "take_profit"):
        _val = config.get(_key)
        if _val is not None and _val > 0.1:  # >10% is almost certainly a percentage
            config[_key] = _val / 100.0
            print(f"[strategy] Auto-corrected {_key}: {_val} -> {config[_key]} (was percentage, converted to decimal)")

    # 1. Load data
    df, close, open_, high, low = load_ohlc(data_path, start_date, end_date)

    # 2. Feature engineering (Claude's function)
    df = feature_fn(df, close, open_, high, low)
    close = df["close"]
    open_ = df["open"]
    high = df["high"]
    low = df["low"]

    # 3. Warm-up detection: drop rows where features have NaN BEFORE any fill
    feature_cols = [c for c in df.columns if c not in ("open", "high", "low", "close")]
    raw_nans = df[feature_cols].isna().any(axis=1)
    valid_rows = ~raw_nans
    if valid_rows.any():
        first_valid = valid_rows.idxmax()
        if raw_nans.loc[:first_valid].any():
            df = df.loc[first_valid:].copy()
            close = df["close"]
            open_ = df["open"]
            high = df["high"]
            low = df["low"]

    # 4. Target
    horizon = config.get("target_horizon", 4)
    target = make_target(close, horizon=horizon)

    # 5. Split (ffill only within each partition — no bfill leak)
    mask = target.notna()
    df = df[mask].copy()
    target = target[mask]
    close = df["close"]
    high = df["high"]
    low = df["low"]

    X = df[feature_cols].copy()
    X = X.replace([np.inf, -np.inf], np.nan)

    if validation_date:
        split_idx = len(df[df.index <= validation_date])
    else:
        split_idx = int(len(df) * train_split)
    split_idx = max(1, min(split_idx, len(df) - 1))

    # ffill within train and test separately (no leak)
    X_train = X.iloc[:split_idx].ffill().fillna(0.0)
    X_test = X.iloc[split_idx:].ffill().fillna(0.0)
    X = pd.concat([X_train, X_test])

    y_train = target.iloc[:split_idx]
    y_test = target.iloc[split_idx:]
    close_train = close.iloc[:split_idx]
    close_test = close.iloc[split_idx:]
    high_test = high.iloc[split_idx:]
    low_test = low.iloc[split_idx:]

    enc = LabelEncoder()
    enc.fit([-1, 0, 1])
    y_train_enc = enc.transform(y_train)
    y_test_enc = enc.transform(y_test)

    split_dt = str(df.index[split_idx])

    sp = {
        "df": df, "X_train": X_train, "X_test": X_test,
        "y_train": y_train, "y_test": y_test,
        "y_train_enc": y_train_enc, "y_test_enc": y_test_enc,
        "enc": enc,
        "close": close, "close_train": close_train, "close_test": close_test,
        "split_idx": split_idx, "split_dt": split_dt,
        "n_train": len(X_train), "n_test": len(X_test),
    }

    # 6. Build model from config
    model = _build_model_from_config(config, X_train, y_train_enc)

    # 7. Generate signals
    threshold = config.get("signal_threshold", 0.55)
    signal_train, p_pos_train, p_neg_train = _generate_signals(model, X_train, threshold)
    signal_test, p_pos_test, p_neg_test = _generate_signals(model, X_test, threshold)

    # 8. Apply filters (order: direction → session → ATR → trend)
    direction = config.get("direction", "both")
    signal_test = _apply_direction_filter(signal_test, direction)
    signal_train = _apply_direction_filter(signal_train, direction)

    session_filter = config.get("session_filter")
    signal_test = _apply_session_filter(signal_test, signal_test.index, session_filter)
    signal_train = _apply_session_filter(signal_train, signal_train.index, session_filter)

    min_atr = config.get("min_atr")
    if min_atr is not None:
        signal_test = _apply_atr_filter(signal_test, close_test, high_test, low_test, min_atr)

    trend_filter = config.get("trend_filter")
    if trend_filter is not None:
        signal_test = _apply_trend_filter(signal_test, close_test, trend_filter)

    signal_full = pd.concat([signal_train, signal_test])

    # 9. Backtest with SL/TP/cooldown (test + train)
    high_train = high.iloc[:split_idx]
    low_train = low.iloc[:split_idx]
    has_risk = (config.get("stop_loss") is not None or
                config.get("take_profit") is not None or
                config.get("cooldown", 0) > 0 or
                config.get("on_opposite", "reverse") != "reverse")

    if has_risk:
        bt = run_backtest_v2(signal_test, close_test, high_test, low_test, config, capital=10000)
        bt_train = run_backtest_v2(signal_train, close_train, high_train, low_train, config, capital=10000)
    else:
        bt = run_backtest(signal_test, close_test, capital=10000)
        bt_train = run_backtest(signal_train, close_train, capital=10000)

    # 10. Metrics
    metrics = compute_metrics(bt, close_test, capital=10000)

    # 11. Pre-compute all trade stats (single source of truth)
    pre_stats = {
        "train_stats": compute_trade_stats(bt_train.get("trade_returns", []), capital=10000),
        "test_stats":  compute_trade_stats(bt.get("trade_returns", []), capital=10000),
        "long_stats":  compute_trade_stats(bt.get("long_returns", []), capital=10000),
        "short_stats": compute_trade_stats(bt.get("short_returns", []), capital=10000),
    }

    # 12. Register model
    if register_model_fn is not None:
        register_model_fn(model)

    # 13. Build return dict
    return build_return_dict(sp, bt, metrics, model, feature_cols,
                             signal_full, p_pos_test, p_neg_test, custom_figs=[],
                             bt_train_result=bt_train, pre_stats=pre_stats)

# ── End strategy_utils ──


DATA_PATH = '/root/Desktop/QuantifyMe/data/ohlc/AUDUSD_15min.parquet'
START_DATE = '2026-04-15'
END_DATE = '2026-05-25'
VALIDATION_DATE = ""
TRAIN_SPLIT = 0.7

# SECTION 1 — FEATURE ENGINEERING
def feature_engineering(df, close, open_, high, low):

    # ── Bollinger Bands (20, 2) ──────────────────────────────────────────────
    bb_period = 20
    bb_std    = 2.0
    bb_mid    = close.rolling(bb_period).mean()
    bb_sigma  = close.rolling(bb_period).std(ddof=0)
    bb_upper  = bb_mid + bb_std * bb_sigma
    bb_lower  = bb_mid - bb_std * bb_sigma

    df["bb_mid"]   = bb_mid
    df["bb_upper"] = bb_upper
    df["bb_lower"] = bb_lower

    # %B — position of close within the band (0 = lower, 1 = upper)
    bb_range       = bb_upper - bb_lower
    df["bb_pct_b"] = np.where(bb_range > 0, (close - bb_lower) / bb_range, 0.5)

    # Bandwidth — normalised band width (regime filter)
    df["bb_bandwidth"] = np.where(bb_mid > 0, bb_range / bb_mid, 0.0)

    # Distance from each band (signed, normalised by sigma)
    df["dist_lower"] = np.where(bb_sigma > 0, (close - bb_lower) / bb_sigma, 0.0)
    df["dist_upper"] = np.where(bb_sigma > 0, (bb_upper - close) / bb_sigma, 0.0)
    df["dist_mid"]   = np.where(bb_sigma > 0, (close - bb_mid)   / bb_sigma, 0.0)

    # Below lower band flag
    df["below_lower"] = np.where(close < bb_lower, 1, 0)
    # Above upper band flag
    df["above_upper"] = np.where(close > bb_upper, 1, 0)

    # ── RSI (14) ─────────────────────────────────────────────────────────────
    rsi_period = 14
    delta      = close.diff()
    gain       = delta.clip(lower=0)
    loss       = (-delta).clip(lower=0)
    avg_gain   = gain.ewm(com=rsi_period - 1, min_periods=rsi_period).mean()
    avg_loss   = loss.ewm(com=rsi_period - 1, min_periods=rsi_period).mean()
    rs         = np.where(avg_loss > 0, avg_gain / avg_loss, 100.0)
    rsi        = 100.0 - 100.0 / (1.0 + rs)
    df["rsi"]  = rsi

    # RSI-derived flags and distances
    df["rsi_oversold"]   = np.where(rsi < 35, 1, 0)
    df["rsi_overbought"] = np.where(rsi > 65, 1, 0)
    df["rsi_dist_35"]    = rsi - 35.0   # negative when oversold
    df["rsi_dist_65"]    = rsi - 65.0   # positive when overbought
    df["rsi_norm"]       = (rsi - 50.0) / 50.0  # centred, ±1 range

    # ── Core entry condition features ────────────────────────────────────────
    # Buy setup: close < lower BB AND RSI < 35
    df["long_setup"]  = np.where((close < bb_lower) & (rsi < 35), 1, 0)
    # Sell setup: close > upper BB AND RSI > 65
    df["short_setup"] = np.where((close > bb_upper) & (rsi > 65), 1, 0)

    # ── ATR (14) — volatility context ────────────────────────────────────────
    atr_period = 14
    hl         = high - low
    hc         = (high - close.shift(1)).abs()
    lc         = (low  - close.shift(1)).abs()
    tr         = pd.concat([hl, hc, lc], axis=1).max(axis=1)
    atr        = tr.ewm(com=atr_period - 1, min_periods=atr_period).mean()
    df["atr"]  = atr
    df["natr"] = np.where(close > 0, atr / close, 0.0)

    # ── Momentum / Rate-of-Change ─────────────────────────────────────────────
    for n in [1, 3, 5, 10]:
        df[f"roc_{n}"] = np.where(
            close.shift(n) > 0,
            (close - close.shift(n)) / close.shift(n),
            0.0
        )

    # ── EMA trend context (fast / slow) ──────────────────────────────────────
    ema_fast = close.ewm(span=9,  min_periods=9).mean()
    ema_slow = close.ewm(span=21, min_periods=21).mean()
    df["ema_fast"] = ema_fast
    df["ema_slow"] = ema_slow
    df["ema_diff"] = np.where(ema_slow > 0, (ema_fast - ema_slow) / ema_slow, 0.0)
    df["ema_bull"] = np.where(ema_fast > ema_slow, 1, 0)

    # SMA-50 trend filter helper (used by framework trend_filter)
    df["sma_50"] = close.rolling(50).mean()

    # ── Candle body & wick features ───────────────────────────────────────────
    body        = (close - open_).abs()
    candle_rng  = (high - low).replace(0, np.nan)
    df["body_ratio"]  = (body / candle_rng).fillna(0.0)
    df["upper_wick"]  = np.where(candle_rng.notna(), (high  - close.clip(lower=open_)) / candle_rng.fillna(1), 0.0)
    df["lower_wick"]  = np.where(candle_rng.notna(), (close.clip(upper=open_) - low)   / candle_rng.fillna(1), 0.0)
    df["bull_candle"]  = np.where(close > open_, 1, 0)

    # ── Volume-like proxy — true range z-score ────────────────────────────────
    tr_mean = tr.rolling(20).mean()
    tr_std  = tr.rolling(20).std(ddof=0).replace(0, np.nan)
    df["tr_zscore"] = ((tr - tr_mean) / tr_std).fillna(0.0)

    # ── Lagged RSI and %B (1, 2, 3 bars back) ────────────────────────────────
    for lag in [1, 2, 3]:
        df[f"rsi_lag{lag}"]    = df["rsi"].shift(lag)
        df[f"bb_pct_b_lag{lag}"] = df["bb_pct_b"].shift(lag)

    # ── RSI slope ────────────────────────────────────────────────────────────
    df["rsi_slope3"] = df["rsi"] - df["rsi"].shift(3)

    # ── Mean-reversion proximity: how far price is from middle band ───────────
    df["pct_to_mid"] = np.where(close > 0, (bb_mid - close) / close, 0.0)

    # ── Fill any NaNs from warm-up ────────────────────────────────────────────
    df = df.bfill().ffill()

    return df


# SECTION 2 — STRATEGY CONFIG
def strategy_config():
    return {
        "title":            "BB Mean-Reversion + RSI Oversold/Overbought (XGBoost)",
        "model_type":       "XGBClassifier",
        "model_params": {
            "n_estimators":      500,
            "max_depth":         4,
            "learning_rate":     0.03,
            "subsample":         0.75,
            "colsample_bytree":  0.70,
            "min_child_weight":  5,
            "gamma":             0.1,
            "reg_alpha":         0.05,
            "reg_lambda":        1.5,
            "objective":         "binary:logistic",
            "random_state":      42,
            "n_jobs":            -1,
        },
        "signal_threshold": 0.55,
        "direction":        "both",
        "stop_loss":        0.0010,
        "take_profit":      0.0020,
        "cooldown":         0,
        "max_positions":    1,
        "on_opposite":      "reverse",
        "session_filter":   [7, 17],
        "min_atr":          0.00005,
        "trend_filter":     None,
        "target_horizon":   4,
        "objective": (
            "Maximize Sharpe ratio by exploiting Bollinger Band mean-reversion "
            "with RSI confirmation. Entry conditions (close < lower BB, RSI < 35 "
            "for longs; close > upper BB, RSI > 65 for shorts) are encoded as "
            "features together with momentum, ATR volatility, candle structure, "
            "and lagged indicators. XGBoost with strong regularisation "
            "(reg_lambda=1.5, gamma=0.1, min_child_weight=5) and a low learning "
            "rate avoids overfitting on the 6-week window. Session filter "
            "[7,17] UTC targets liquid London/NY overlap, reducing noise. "
            "TP:SL ratio of 2:1 supports positive expected value even at "
            "moderate win rates, pushing Sharpe higher."
        ),
        "notes": (
            "Features: %B position, RSI (raw + flags + slope + lags), "
            "EMA cross, ATR/NATR, ROC(1/3/5/10), candle body/wick ratios, "
            "TR z-score, distance-to-midband, long/short setup flags. "
            "Round-trip cost ~2e-5 is implicitly absorbed by the 10-pip TP target. "
            "Cooldown=0 allows immediate re-entry after mean-reversion completes."
        ),
    }

# ── Framework v2: auto-generated wrapper ──
def train_and_backtest():
    _vd = VALIDATION_DATE if 'VALIDATION_DATE' in globals() else ''
    _ts = TRAIN_SPLIT if 'TRAIN_SPLIT' in globals() else 0.7
    return run_strategy(
        feature_engineering, strategy_config,
        DATA_PATH, START_DATE, END_DATE,
        _vd, _ts,
        register_model_fn=register_model
    )

QM_SWAP_SHORT_USD = 0.3
QM_SWAP_LONG_USD = -1.0
QM_COMMISSION_USD = 6.0
# ╔══════════════════════════════════════════════════════════════╗
# ║                  STRATEGY REQUEST LOG                       ║
# ╚══════════════════════════════════════════════════════════════╝
# Generated        : 2026-08-07 11:28:27
# Model            : XGBoost
# Feature Eng.     : Auto-add features: ON
# Signal / Entry   : —
# Optimization     : —
# Risk Mgmt        : —
# Risk Filter      : —
# ══════════════════════════════════════════════════════════════

# ============================================================
# QUANTIFY.ME TRADING STRATEGY — XGBoost RSI + Bollinger Bands
# ============================================================

import numpy as np
import pandas as pd

# ── Inlined strategy_utils ──
"""
strategy_utils.py — Standard utility functions for generated strategies.
Claude imports these instead of writing boilerplate from scratch.
This ensures consistent behavior across all generated strategies.
"""
import numpy as np
import pandas as pd
from sklearn.preprocessing import LabelEncoder


# ── Broker-accurate accounting (Factual Trading Product, Phase 1) ──────────
# These mirror v2/config.py but are duplicated HERE because strategy_utils is
# INLINED standalone into Modal training jobs (no `config` import available in
# the container). Keep in sync with config.py if those change.
LOT_UNITS      = 100_000      # units per standard lot
COMMISSION_USD = 6.0         # round-trip commission, USD per standard lot
SWAP_LONG_USD  = -1.00       # USD per lot per night held long  (cost)
SWAP_SHORT_USD =  0.30       # USD per lot per night held short (credit)
DEFAULT_FX_SPREAD_PIPS = 0.15  # default bid/ask spread (pips) for FX majors (#283)


def _usd_conv(symbol, rate):
    """Divisor to convert quote-currency P&L into the USD account currency.

    XXXUSD pairs (EUR/GBP/AUD/NZD-USD) are already quoted in USD → conv = 1.
    USDXXX pairs (USD-JPY/CHF/CAD) are quoted in the foreign ccy → divide the
    quote-ccy P&L by the pair's price (use the trade's exit price; brokers
    convert at close). Replaces the old "EURUSD-approximated for all" caveat.
    """
    s = (symbol or "").upper().replace("/", "")
    if len(s) >= 6 and s[:3] == "USD":
        return rate if rate else 1.0
    return 1.0


# Keep in sync with predictor._CRYPTO_PIP (self-contained; strategy_utils has no config import).
_CRYPTO_PIP = {"BTCUSD": 1.0}


def _pip_size(symbol):
    """Pip size: crypto uses its tick (BTCUSD=1.0), JPY pairs 0.01, other majors 0.0001."""
    s = (symbol or "").upper()
    if s in _CRYPTO_PIP:
        return _CRYPTO_PIP[s]
    return 0.01 if s.endswith("JPY") else 0.0001


def _default_spread_pips(symbol):
    """Default bid/ask spread (pips) when the caller doesn't specify (#283).
    FX majors ~0.15 pip (raw/ECN — pairs with the $6/lot commission). Crypto
    returns 0.0: its costs are modeled via HL maker/taker tiers, not a spread."""
    s = (symbol or "").upper()
    if s in _CRYPTO_PIP:
        return 0.0
    return DEFAULT_FX_SPREAD_PIPS


def _trade_pnl_usd(direction, entry_px, exit_px, lots, symbol, nights,
                   commission_usd=None, swap_long_usd=None, swap_short_usd=None):
    """Broker-accurate per-trade P&L in USD (account currency).

      units      = lots × LOT_UNITS
      pnl_quote  = units × (exit − entry) × dir        (dir +1 buy / −1 sell)
      pnl_usd    = pnl_quote / conv(symbol, exit_px)   (per-pair FX)
      pnl_usd   −= commission × lots                   (round-trip)
      pnl_usd   += swap(side) × lots × nights          (swap_long is negative)

    commission/swap default to the module constants when None — the Config
    tab / API values are threaded through run_strategy (#254: they used to be
    collected but never consumed, so tuning them silently did nothing).
    """
    _comm = COMMISSION_USD if commission_usd is None else float(commission_usd)
    _swl = SWAP_LONG_USD if swap_long_usd is None else float(swap_long_usd)
    _sws = SWAP_SHORT_USD if swap_short_usd is None else float(swap_short_usd)
    units = lots * LOT_UNITS
    pnl_quote = units * (exit_px - entry_px) * direction
    usd = pnl_quote / _usd_conv(symbol, exit_px)
    usd -= _comm * lots
    usd += (_swl if direction > 0 else _sws) * lots * nights
    return usd


def _pips(symbol, direction, entry_px, exit_px):
    """Signed pips for the trade (cTrader-parity Pips column)."""
    return direction * (exit_px - entry_px) / _pip_size(symbol)


def _nights_held(idx_dates, idx_dow, entry_bar, exit_bar):
    """Overnight rolls between entry and exit bars. Each calendar-day boundary
    the position is held across = one night; Wednesday→Thursday rolls ×3 (covers
    weekend settlement) — matches the legacy pipeline.py swap model."""
    if idx_dates is None or entry_bar is None or exit_bar is None:
        return 0.0
    nights = 0.0
    for j in range(int(entry_bar), int(exit_bar)):
        if idx_dates[j] != idx_dates[j + 1]:
            nights += 3.0 if idx_dow[j] == 2 else 1.0
    return nights


def _index_date_arrays(idx):
    """Return (dates, dayofweek) arrays for a DatetimeIndex, or (None, None)."""
    try:
        return idx.date, idx.dayofweek.values
    except Exception:
        try:
            return (np.array([t.date() for t in idx]),
                    np.array([t.dayofweek for t in idx]))
        except Exception:
            return None, None


# Max backtest window per timeframe. A finer timeframe over a longer window
# blows up the results dict / parquet load / Modal train time (the 2026-05-12
# OOM was a 1-min × multi-year sweep) — and a 1-min strategy gains nothing from
# 2 years of 1-min bars. Enforced HERE because every training path (UI / API /
# Modal) funnels through run_strategy → load_ohlc. Env-overridable so a future
# "max plan" / dedicated-server tier can lift it.
_TF_MAX_DAYS = {
    "1min":  30,
    "5min":  90,
    "15min": 365,
    "1h":    730,
}


def tf_max_days(tf):
    """Public accessor for the per-timeframe window cap, env override applied.

    The dash layer uses this to clamp cfg-start-date at entry (#364) so the UI
    agrees with the silent load_ohlc truncation below. None = no cap known.
    """
    import os as _os
    if not tf:
        return None
    _max = _TF_MAX_DAYS.get(str(tf).strip())
    if not _max:
        return None
    _env = _os.environ.get(f"QM_MAX_DAYS_{str(tf).strip().upper()}")
    if _env and _env.isdigit():
        _max = int(_env)
    return _max


def clamp_backtest_start(start_date, end_date, tf):
    """(start_out, was_clamped, max_days) — pure window-cap policy (#364).

    Accepts YYYY-MM-DD or DD-MM-YYYY strings; a clamped start comes back
    canonical YYYY-MM-DD. Unparseable input or an uncapped timeframe passes
    through untouched. Missing/unparseable end defaults to today.
    """
    _max = tf_max_days(tf)
    if not _max:
        return start_date, False, None

    def _parse(s):
        for _fmt in ("%Y-%m-%d", "%d-%m-%Y", "%d-%m-%y"):
            try:
                return pd.Timestamp(__import__("datetime").datetime.strptime(str(s).strip(), _fmt))
            except (ValueError, TypeError):
                continue
        return None

    _start = _parse(start_date)
    if _start is None:
        return start_date, False, _max
    _end = _parse(end_date) or pd.Timestamp.utcnow().normalize().tz_localize(None)
    _floor = _end - pd.Timedelta(days=_max)
    if _start < _floor:
        return _floor.strftime("%Y-%m-%d"), True, _max
    return start_date, False, _max


def normalize_bt_date(s):
    """'YYYY-MM-DD[ HH:MM[:SS]]' | 'DD-MM-YYYY' | 'DD-MM-YY' → 'YYYY-MM-DD', else None.

    #373: bt_config carries a MIX of exact datetimes (vline drags) and bare
    dates (typed input). Naive string compares between the two are prefix
    accidents ('2026-07-22 23:45:00' >= '2026-07-22' is True) — every date
    comparison must go through this date-part normalisation first.
    """
    if s is None:
        return None
    _s = str(s).strip()
    if not _s:
        return None
    # Drop a time component if present — the date part is what we compare.
    _date_part = _s.split(" ")[0].split("T")[0]
    import datetime as _dt
    for _fmt in ("%Y-%m-%d", "%d-%m-%Y", "%d-%m-%y"):
        try:
            return _dt.datetime.strptime(_date_part, _fmt).strftime("%Y-%m-%d")
        except ValueError:
            continue
    return None


def validate_date_order(start_date, end_date):
    """(ok, start_norm, end_norm) — reject start >= end at DATE granularity (#373).

    ok=False only when BOTH sides parse and start_norm >= end_norm; missing or
    unparseable input passes (never block on garbage — downstream defaults
    handle it, same stance as clamp_backtest_start).
    """
    _sn = normalize_bt_date(start_date)
    _en = normalize_bt_date(end_date)
    if _sn and _en and _sn >= _en:
        return False, _sn, _en
    return True, _sn, _en


def clamp_to_data_tail(start_date, end_date, tail_date):
    """(start_out, end_out, clamped) — snap date-parts past the data tail back (#373).

    tail_date is the last available candle's 'YYYY-MM-DD'. A start or end whose
    date-part lies beyond it comes back as tail_date, with 'start'/'end' added
    to the `clamped` set. Unparseable/empty values pass through untouched
    (the JS date poller already ignores garbage).
    """
    clamped = set()
    _tail = normalize_bt_date(tail_date)
    if not _tail:
        return start_date, end_date, clamped
    _sn = normalize_bt_date(start_date)
    _en = normalize_bt_date(end_date)
    if _sn and _sn > _tail:
        start_date, clamped = _tail, clamped | {"start"}
    if _en and _en > _tail:
        end_date, clamped = _tail, clamped | {"end"}
    return start_date, end_date, clamped


DEFAULT_BACKTEST_DAYS = 40


def resolve_backtest_window(start_date, end_date, tf):
    """(start, end, was_clamped, max_days) — the ONE default-window policy (#370).

    Every deploy surface resolves its window here so they can't drift: /landing
    and the /docs Quick Deploy widget (via /api/v1/one_shot) and the /docs
    generate→train→deploy composer (via /api/v1/train) all land on the same
    dates. Before this, the composer carried its own 28d/-5d window hardcoded in
    the page — a client-side mirror of a server policy, which is the same shape
    of bug as #367.

    Missing end defaults to today, so the stats cover up to the latest bar
    (user 2026-05-24: "I want it till Today!"). Missing start defaults to
    DEFAULT_BACKTEST_DAYS back. The result is then clamped to the timeframe's
    window cap, so the requested window is always the window that trains.
    """
    _end = str(end_date or "").strip()
    _start = str(start_date or "").strip()
    _today = pd.Timestamp.utcnow().normalize().tz_localize(None)
    if not _end:
        _end = _today.strftime("%Y-%m-%d")
    if not _start:
        _base = pd.Timestamp(_end) if _end else _today
        _start = (_base - pd.Timedelta(days=DEFAULT_BACKTEST_DAYS)).strftime("%Y-%m-%d")
    _start, _capped, _max = clamp_backtest_start(_start, _end, tf)
    return _start, _end, _capped, _max


def _fetch_ohlc_from_internal(symbol: str, tf: str, start: str, end: str):
    """Phase 3.2: fetch parquet bytes from Server A's /internal/ohlc endpoint
    instead of reading a local file. Used inside Modal containers / Mac worker
    pool (Phase 3.4) so every train sees the same source of truth as the chart.

    Returns: pd.DataFrame (parquet decoded), or raises on any failure so the
    caller can fall back / surface a clear error in the job.
    """
    import hashlib as _hashlib, hmac as _hmac, io as _io, os as _os
    import urllib.request as _ur, urllib.parse as _urp

    base = (_os.environ.get("QM_INTERNAL_OHLC_BASE") or "").rstrip("/")
    secret = (_os.environ.get("INTERNAL_WS_SECRET") or "").strip()
    if not base:
        raise RuntimeError("QM_INTERNAL_OHLC_BASE not set")
    if not secret:
        raise RuntimeError("INTERNAL_WS_SECRET not set")

    msg = f"{symbol}|{tf}|{start}|{end}".encode("utf-8")
    sig = _hmac.new(secret.encode("utf-8"), msg, _hashlib.sha256).hexdigest()

    qs = _urp.urlencode({
        "symbol": symbol, "tf": tf,
        "start": start, "end": end, "sig": sig,
    })
    url = f"{base}/internal/ohlc?{qs}"

    req = _ur.Request(url, headers={"User-Agent": "qm-worker/1.0"})
    with _ur.urlopen(req, timeout=30) as resp:
        if resp.status != 200:
            raise RuntimeError(f"/internal/ohlc returned {resp.status}")
        payload = resp.read()
    print(f"[load_ohlc:internal] {symbol} {tf} fetched {len(payload)} bytes", flush=True)
    return pd.read_parquet(_io.BytesIO(payload))


def _parse_symbol_tf_from_path(data_path: str):
    """Pull SYMBOL + TF out of a path like .../EURUSD_1min.parquet."""
    import os as _os, re as _re
    base = _os.path.basename(str(data_path))
    m = _re.match(r"^([A-Z]{6})_(\d+min|\d+h)\.parquet$", base)
    if not m:
        return None, None
    return m.group(1), m.group(2)


def load_ohlc(data_path, start_date="", end_date=""):
    """Load OHLC parquet, sort index, filter dates. Always returns consistent format.

    The lower bound is clamped per timeframe (see _TF_MAX_DAYS) — a request for
    more history than the cap silently starts later.

    Phase 3.2: when env QM_USE_INTERNAL_OHLC=="1", fetch over HTTP from
    Server A's /internal/ohlc endpoint instead of pd.read_parquet on a local
    file (which on Modal is a stale Volume snapshot). The endpoint applies the
    same day-cap, so the local cap-check below is a defensive no-op in that
    path. Flag defaults to "0" → unchanged behavior.

    Returns: (df, close, open_, high, low)
    """
    import os as _os, re as _re

    _use_internal = _os.environ.get("QM_USE_INTERNAL_OHLC", "0") == "1"
    if _use_internal:
        _sym, _tf = _parse_symbol_tf_from_path(data_path)
        if not _sym or not _tf:
            raise RuntimeError(
                f"QM_USE_INTERNAL_OHLC=1 but DATA_PATH basename does not match "
                f"SYMBOL_TF.parquet: {data_path}"
            )
        df = _fetch_ohlc_from_internal(_sym, _tf, start_date or "", end_date or "")
    else:
        df = pd.read_parquet(data_path)
    df.index = pd.to_datetime(df.index)
    df = df.sort_index()

    # Per-timeframe window cap (timeframe inferred from the parquet filename).
    _m = _re.search(r"_(\d+min|\d+h)\.parquet$", _os.path.basename(str(data_path)))
    _tf = _m.group(1) if _m else None
    _max_days = tf_max_days(_tf)
    if _max_days and _max_days > 0 and len(df):
        try:
            _eff_end = pd.Timestamp(end_date) if end_date else df.index.max()
            _eff_end = min(_eff_end, df.index.max())
            _floor = _eff_end - pd.Timedelta(days=_max_days)
            _req_start = pd.Timestamp(start_date) if start_date else df.index.min()
            if _req_start < _floor:
                print(f"[load_ohlc] {_tf} backtest window capped to {_max_days}d: "
                      f"start {_req_start.date()} -> {_floor.date()}", flush=True)
                start_date = _floor
        except Exception as _e:
            print(f"[load_ohlc] window-cap check skipped ({_e})", flush=True)

    if start_date:
        df = df[df.index >= start_date]
    if end_date:
        df = df[df.index <= end_date]

    return df, df["close"], df["open"], df["high"], df["low"]


TF_MINUTES = {"1min": 1, "5min": 5, "15min": 15, "1h": 60}
MAX_AUX_FEEDS = 3
AUX_MIN_COVERAGE = 0.90  # completed aux bar must exist for >=90% of primary
                         # rows *after* the feed has warmed up (see note below)


def normalize_aux_specs(raw, primary_symbol, primary_tf):
    """#387/#388: validate/normalize aux feed specs. Returns (specs, warnings).

    Honest-drop discipline (#377): every removal produces a user-facing
    warning that is true in every branch - never silently shrink."""
    from data_paths import SUPPORTED_SYMBOLS, SUPPORTED_TIMEFRAMES
    specs, warnings, seen = [], [], set()
    for item in (raw or []):
        if not isinstance(item, dict):
            warnings.append("an aux feed entry was not an object - ignored.")
            continue
        sym = str(item.get("symbol") or primary_symbol).upper().strip()
        tf = str(item.get("timeframe") or primary_tf).strip()
        if sym not in SUPPORTED_SYMBOLS:
            warnings.append(f"aux feed {sym} {tf}: unsupported symbol - dropped.")
            continue
        if tf not in SUPPORTED_TIMEFRAMES:
            warnings.append(f"aux feed {sym} {tf}: unsupported timeframe - dropped.")
            continue
        if sym == str(primary_symbol).upper() and tf == primary_tf:
            warnings.append(
                f"aux feed {sym} {tf} equals the primary feed - dropped.")
            continue
        if (sym, tf) in seen:
            continue
        if len(specs) >= MAX_AUX_FEEDS:
            warnings.append(
                f"aux feed {sym} {tf}: over the {MAX_AUX_FEEDS}-feed cap - dropped.")
            continue
        seen.add((sym, tf))
        specs.append({"symbol": sym, "timeframe": tf})
    return specs, warnings


def _parse_aux_feeds_regex_fallback(code):
    """Line-anchored regex fallback for `parse_aux_feeds_from_code`, used
    only when `ast.parse` itself fails (pasted code mid-edit and genuinely
    unparsable). Deliberately limited to a single physical line ending in
    `]` - it cannot see multi-line literals or trailing comments, but on
    unparsable input there is no AST to walk anyway, and a still-legal
    single-line stamp is worth recovering rather than giving up entirely.
    Still takes the LAST match, never the first."""
    import ast as _ast
    import re as _re
    matches = _re.findall(r"^AUX_FEEDS\s*=\s*(\[.*\])\s*$", code or "",
                          _re.MULTILINE)
    if not matches:
        return []
    try:
        parsed = _ast.literal_eval(matches[-1])
    except Exception:
        return []
    return parsed if isinstance(parsed, list) else []


def parse_aux_feeds_from_code(code):
    """Task 9 (#387/#388): read the AUX_FEEDS the deployed strategy code
    actually carries, for the deploy bundle / live predictor.

    MUST take the LAST module-scope match, never the first. Task 7's
    builder guarantees AUX_FEEDS by APPENDING its authoritative assignment
    as the last module-scope statement (module execution is last-write-
    wins) - the model's own earlier AUX_FEEDS line can legitimately survive
    above it, because the builder's cleanliness strip deliberately declines
    to remove statements it does not fully own. A first-match reader (e.g.
    re.search) would read the MODEL's value - the exact exploit Task 7
    spent five review rounds closing, reintroduced here.

    Parses via `ast` rather than a line-anchored regex: walks the module's
    top-level statements for `Assign` nodes with a single, bare
    `Name("AUX_FEEDS")` target, keeps the LAST one, and `literal_eval`s its
    value. This handles a multi-line literal and a trailing same-line
    comment correctly (both legal Python the BACKTEST already handles fine
    by executing the module - only this static reader used to choke on
    them), and it structurally cannot mistake an "AUX_FEEDS = [...]" string
    that only appears inside a docstring for a real assignment (a
    docstring's TEXT is never re-parsed as further statements).

    Falls back to a line-anchored regex only when `ast.parse` itself fails
    (pasted code can be genuinely mid-edit / unparsable) - the regex path
    still recovers a still-legal single-line stamp elsewhere in text that
    doesn't compile, rather than giving up outright.

    Returns [] on any failure (missing, malformed, or not a list) - never
    raises."""
    import ast as _ast
    code = code or ""
    try:
        tree = _ast.parse(code)
    except Exception:
        return _parse_aux_feeds_regex_fallback(code)
    last_value_node = None
    for node in tree.body:
        if (isinstance(node, _ast.Assign)
                and len(node.targets) == 1
                and isinstance(node.targets[0], _ast.Name)
                and node.targets[0].id == "AUX_FEEDS"):
            last_value_node = node.value
    if last_value_node is None:
        return []
    try:
        parsed = _ast.literal_eval(last_value_node)
    except Exception:
        return []
    return parsed if isinstance(parsed, list) else []


def _default_aux_loader(symbol, tf, start=None, end=None):
    """#403: aux data must travel the SAME road as the primary feed.

    load_ohlc() switches to /internal/ohlc when QM_USE_INTERNAL_OHLC=1, which
    is exactly what Modal containers set - and since M22c removed the Volume
    there is no OHLC on the container filesystem at all. A loader that reads
    a local parquet can therefore only ever raise there, and merge_aux_feeds
    turns that into a silent "dropped" - the model trains with no aux columns
    and still reports success. One transport decision, made in one place.
    """
    import os as _os
    if _os.environ.get("QM_USE_INTERNAL_OHLC", "0") == "1":
        return _fetch_ohlc_from_internal(symbol, tf, start or "", end or "")
    from data_paths import data_path_for
    aux = pd.read_parquet(data_path_for(symbol, tf))
    # #406: honour the window in BOTH branches. The HTTP branch gets it for
    # free (the endpoint filters), so leaving the local branch unfiltered
    # made the same call mean different things depending on an env var - and
    # the finer-than-primary merge below then ran three groupby.apply passes
    # over the ENTIRE 1min history instead of the few days actually being
    # merged. Measured on Server A: a 1min feed into a 5min primary took
    # 128.78s unfiltered vs 0.04s for an equal-timeframe feed.
    if start or end:
        aux = aux.copy()
        aux.index = pd.to_datetime(aux.index)
        if start:
            aux = aux[aux.index >= pd.Timestamp(start)]
        if end:
            aux = aux[aux.index <= pd.Timestamp(end)]
    return aux


def merge_aux_feeds(df, aux_specs, primary_symbol, primary_tf, loader=None):
    """#387/#388: merge aux (symbol, timeframe) feeds into the primary df as
    aux_* columns, BEFORE feature_engineering. The single audited home of the
    no-lookahead rule - backtest (run_strategy) and live (predictor) both call
    this, so train/serve skew is structurally impossible.

    Info-time of primary row t is t + primary_tf (the bar's own close is known
    then - identical to how the existing pipeline treats the bar's own OHLC).
    An aux bar labeled s (timeframe atf) becomes visible at s + atf.

    Loader contract (#403): loader(symbol, timeframe, start, end) -> DataFrame.
    start/end are ISO strings bracketing the window this merge needs; a live
    loader that always serves the latest candles may ignore them.
    """
    loader = loader or _default_aux_loader
    report = {"merged": [], "dropped": []}
    # CC-2 (final whole-branch review): TF_MINUTES is a SECOND source of
    # truth for supported timeframes, separate from SUPPORTED_TIMEFRAMES /
    # normalize_aux_specs' own validation - they can drift apart (e.g. a
    # future SUPPORTED_TIMEFRAMES addition this dict never learns about).
    # run_strategy calls this function for EVERY strategy, aux or not, so a
    # primary_tf missing here must never raise: the NO-AUX path (the vast
    # majority of strategies) must return the df untouched exactly like the
    # `not aux_specs` early-out below, and the WITH-AUX path must drop every
    # requested feed with an honest reason naming the unsupported primary
    # timeframe - never a bare KeyError either way.
    ptf_min = TF_MINUTES.get(primary_tf)
    if ptf_min is None:
        if not aux_specs:
            return df, report
        for spec in aux_specs:
            report["dropped"].append({
                "symbol": spec.get("symbol"), "timeframe": spec.get("timeframe"),
                "reason": f"primary timeframe {primary_tf!r} is not supported "
                          f"for aux merging"})
        return df, report
    if not aux_specs:
        return df, report
    # pd.merge_asof silently produces WRONG results (not an error) when its
    # LEFT frame isn't sorted by the join key. `aux` is sorted defensively
    # below but df's own index wasn't - sort it too.
    df = df.sort_index()
    info_time = df.index + pd.Timedelta(minutes=ptf_min)
    for spec in aux_specs:
        sym, atf = spec["symbol"], spec["timeframe"]
        atf_min = TF_MINUTES[atf]
        prefix = f"aux_{sym.lower()}_{atf}"
        # #403: ask for the window we actually need. /internal/ohlc day-caps
        # an unbounded request to the most RECENT window, so a historical
        # backtest that asked for "everything" would get aux data that does
        # not overlap the primary at all - and the overlap failure is silent
        # (all-NaN columns, then a coverage drop). The pad covers an FX
        # weekend plus one aux bar so row 0 already has a completed aux bar.
        _pad = pd.Timedelta(days=3) + pd.Timedelta(minutes=atf_min)
        _start = (df.index.min() - _pad).strftime("%Y-%m-%d %H:%M:%S")
        _end = (df.index.max() + pd.Timedelta(minutes=ptf_min)).strftime(
            "%Y-%m-%d %H:%M:%S")
        try:
            aux = loader(sym, atf, _start, _end)
        except Exception as exc:
            report["dropped"].append({"symbol": sym, "timeframe": atf,
                                      "reason": f"load failed: {exc}"})
            continue
        if aux is None or len(aux) == 0:
            report["dropped"].append({"symbol": sym, "timeframe": atf,
                                      "reason": "no data"})
            continue
        aux = aux.sort_index()
        aux.index = pd.to_datetime(aux.index)
        cols = []
        if atf_min >= ptf_min:
            # Completed-bar view: asof on COMPLETION time (label + atf).
            right = aux[["open", "high", "low", "close"]].copy()
            right.index = right.index + pd.Timedelta(minutes=atf_min)
            right.columns = [f"{prefix}_{c}" for c in right.columns]
            merged = pd.merge_asof(
                pd.DataFrame(index=info_time), right,
                left_index=True, right_index=True, direction="backward")
            merged.index = df.index  # back to primary labels
            # Coverage is measured only from the first row where a completed
            # aux bar is even possible - every coarser/equal feed has an
            # unavoidable, correct, all-NaN warm-up before its first bar
            # completes (see test_rows_before_first_aux_bar_are_nan); that
            # is not a data-quality problem and must not count against the
            # feed. Gaps AFTER warm-up (missing/short aux history, holes in
            # the source data) still count and can still trip the floor.
            close_col = merged[f"{prefix}_close"]
            first_valid = close_col.first_valid_index()
            if first_valid is None:
                coverage = 0.0
            else:
                coverage = close_col.loc[first_valid:].notna().mean()
            if coverage < AUX_MIN_COVERAGE:
                report["dropped"].append({
                    "symbol": sym, "timeframe": atf,
                    "reason": f"completed-bar coverage {coverage:.0%} < "
                              f"{AUX_MIN_COVERAGE:.0%} of primary rows"})
                continue
            for c in merged.columns:
                df[c] = merged[c]
                cols.append(c)

            if atf_min > ptf_min:
                # Forming-bar view (Malcolm's "the 1h bar up till 11:15"):
                # a running OHLC aggregate over the aux symbol's PRIMARY-tf
                # bars inside the CURRENT coarse window, truncated at each
                # row's own info-time. Leak-free by construction - only bars
                # that have already closed enter the running aggregate.
                # Same-symbol feeds reuse df's own OHLC (no extra load);
                # cross-symbol feeds load the aux symbol at the primary tf.
                try:
                    fine = (df if sym == str(primary_symbol).upper()
                            else loader(sym, primary_tf))
                    fine = fine.sort_index()
                    fine.index = pd.to_datetime(fine.index)
                    win = fine.index.floor(f"{atf_min}min")
                    g = fine.groupby(win)
                    forming = pd.DataFrame({
                        f"{prefix}_live_open":  g["open"].transform("first"),
                        f"{prefix}_live_high":  g["high"].cummax(),
                        f"{prefix}_live_low":   g["low"].cummin(),
                        f"{prefix}_live_close": fine["close"],
                        "_window": win,
                    }, index=fine.index)
                    # visible once THIS fine bar itself completes
                    forming.index = forming.index + pd.Timedelta(minutes=ptf_min)
                    fm = pd.merge_asof(
                        pd.DataFrame(index=info_time), forming,
                        left_index=True, right_index=True, direction="backward")
                    fm.index = df.index
                    # A forming value carried over from a PREVIOUS coarse
                    # window (e.g. a gap in the fine data) must not leak into
                    # a new window before that window's own first fine bar
                    # has closed - mask those rows back to NaN.
                    row_window = df.index.floor(f"{atf_min}min")
                    live_cols = [c for c in fm.columns if c != "_window"]
                    stale = fm["_window"].values != row_window.values
                    fm.loc[stale, live_cols] = np.nan
                    for c in live_cols:
                        df[c] = fm[c]
                        cols.append(c)
                except Exception as exc:
                    print(f"[aux_feeds] forming-bar columns for {sym} {atf} "
                          f"skipped ({exc})", flush=True)
        else:
            # Finer-than-primary: raw fine OHLC aggregates straight back into
            # the primary bar (adds nothing at bar-close granularity), so
            # provide within-bar PATH features instead. Fine bars labeled in
            # [t, t+ptf) all complete by row t's info-time (t+ptf) - no shift
            # needed, unlike the coarser branch above. `floor` groups each
            # fine bar into the primary bar that CONTAINS it (same alignment
            # assumption the forming-bar branch already makes) - a fine bar
            # from the NEXT primary bar must never land in this group, or
            # that's lookahead (see test_finer_uses_only_own_bar).
            window = aux.index.floor(f"{ptf_min}min")
            g = aux.groupby(window)
            # pct_change/cummax must reset at each group's own start (a
            # global pct_change would leak the previous bar's last close
            # into this bar's first fine-bar return).
            #
            # #407: these were three groupby.apply calls carrying Python
            # lambdas - one interpreter round-trip PER PRIMARY BAR. Cost
            # tracked group COUNT, not row count: 200k fine bars took 14.63s
            # against a 5min primary (40k groups) but 4.55s against a 15min
            # one (13.3k groups) - same data, 3x the groups, 3.2x the time.
            # The groupby transforms below reset per group exactly as the
            # lambdas did, so the numbers are unchanged (test_407 pins them
            # against the original implementation as an oracle).
            _close = aux["close"]
            _cg = _close.groupby(window)
            rvol = _cg.pct_change().groupby(window).std()
            _cummax = _cg.cummax()
            maxdd = ((_cummax - _close) / _cummax).groupby(window).max()
            upratio = (aux["close"] > aux["open"]).groupby(window).mean()
            feats = pd.DataFrame({
                f"{prefix}_rvol": rvol,
                f"{prefix}_upratio": upratio,
                f"{prefix}_maxdd": maxdd,
            })
            feats = feats.reindex(df.index)
            # Coverage measures "did we have ANY fine data for this primary
            # bar" via group non-emptiness (g.size(), reindexed) - NOT via
            # the `_rvol` column. `rvol` needs >=3 fine bars in a group to be
            # non-NaN (2 to get a single pct-change, 2 valid values to get a
            # std) while `upratio`/`maxdd` only need 1 - keying coverage to
            # `rvol` would drop an otherwise-good sparse feed (e.g. exactly
            # 2 fine bars/primary bar) entirely, even though 2 of 3 columns
            # are fully populated. As with the completed-bar branch above,
            # measured from the first row a group is even possible onward -
            # an unavoidable leading warm-up before the fine feed's first
            # group is not a data-quality problem and must not count
            # against the feed.
            group_sizes = g.size().reindex(df.index)
            first_valid = group_sizes.first_valid_index()
            if first_valid is None:
                coverage = 0.0
            else:
                coverage = group_sizes.loc[first_valid:].notna().mean()
            if coverage < AUX_MIN_COVERAGE:
                report["dropped"].append({
                    "symbol": sym, "timeframe": atf,
                    "reason": f"fine-bar coverage {coverage:.0%} < "
                              f"{AUX_MIN_COVERAGE:.0%} of primary rows"})
                continue
            for c in feats.columns:
                df[c] = feats[c]
                cols.append(c)
        report["merged"].append({"symbol": sym, "timeframe": atf,
                                 "columns": cols})
    return df, report


# Suffix -> meaning, checked longest-first so "aux_x_1h_live_close" resolves
# to the live_* meaning instead of the plain "close" one (both end in
# "_close" - order is what disambiguates them, not a smarter match).
_AUX_SUFFIX_ORDER = (
    "live_open", "live_high", "live_low", "live_close",
    "open", "high", "low", "close",
    "rvol", "upratio", "maxdd",
)


def _aux_column_meaning(col, sym, tf):
    """#387/#388: one-line, per-suffix meaning for an aux_* column, used to
    teach codegen what already exists on df so it never recomputes it."""
    for suf in _AUX_SUFFIX_ORDER:
        if col.endswith("_" + suf):
            if suf.startswith("live_"):
                return (f"the FORMING {tf} bar of {sym}, truncated at the "
                        f"current bar - no lookahead")
            if suf in ("open", "high", "low", "close"):
                return f"the last COMPLETED {tf} bar of {sym} ({suf})"
            if suf == "rvol":
                return f"std of {tf} returns within the current bar"
            if suf == "upratio":
                return f"fraction of up {tf} bars within the current bar"
            if suf == "maxdd":
                return "worst drawdown within the current bar"
    # Defensive fallback - every column merge_aux_feeds produces matches one
    # of the suffixes above; this only fires if that ever drifts.
    return f"an aux feed column from {sym} {tf}"


def aux_columns_prompt(report):
    """#387/#388: turn a merge_aux_feeds() report into a codegen prompt block
    teaching the LLM which aux_* columns already sit on df, so it never
    recomputes them and never assigns to them. "" when nothing merged - the
    no-aux generation path must gain ZERO content from this function."""
    merged = (report or {}).get("merged") or []
    if not merged:
        return ""
    lines = [
        "",
        "",
        "AUX FEED COLUMNS (already merged onto df before feature_engineering runs):",
    ]
    for entry in merged:
        sym = entry.get("symbol", "")
        tf = entry.get("timeframe", "")
        for col in entry.get("columns") or []:
            lines.append(f"  - {col}: {_aux_column_meaning(col, sym, tf)}")
    lines.append(
        "These columns already exist on df - never recompute them and never "
        "read any other data source.")
    lines.append("Never assign to any aux_* column.")
    return "\n".join(lines)


def make_target(close, horizon=4):
    """Create target: direction N bars ahead. Default 4 bars = 1 hour on 15-min data.

    Returns: target (pd.Series of -1, 0, 1)
    """
    return np.sign(close.shift(-horizon) - close)


def split_bounds(n_rows, split_idx, purge_bars=0, embargo_bars=0):
    """#411: the ONE chronological splitter. Train slice is [:train_end],
    test slice is [test_start:].

    purge_bars: rows removed from the END of train. make_target looks
    `target_horizon` bars forward, so without purging the last `horizon`
    train rows carry labels computed from the first test rows - leakage
    across the boundary. purge_bars=target_horizon closes it.
    embargo_bars: extra gap before test starts (serial-correlation guard).

    Defaults (0, 0) reproduce the legacy split EXACTLY - the headline 70/30
    path must stay bit-identical (see TestHeadlineUnchanged).
    """
    split_idx = max(1, min(int(split_idx), n_rows - 1))
    train_end = max(1, split_idx - int(purge_bars))
    test_start = min(n_rows - 1, split_idx + int(embargo_bars))
    if train_end > test_start:
        train_end = test_start
    return train_end, test_start


def split_data(df, target, feature_cols, train_split=0.7, validation_date=""):
    """Train/test split. Handles both ratio and date-based splits.
    Drops NaN from target before splitting. Encodes labels to [0,1,2].

    Returns: dict with keys:
        X_train, X_test, y_train, y_test,
        y_train_enc, y_test_enc, enc,
        close_train, close_test,
        split_idx, split_dt, n_train, n_test
    """
    # Drop NaN from target
    mask = target.notna()
    df = df[mask].copy()
    target = target[mask]

    close = df["close"]

    # Build feature matrix
    X = df[feature_cols].copy()
    X = X.bfill().ffill()
    X = X.replace([np.inf, -np.inf], np.nan).fillna(0.0)

    # Split
    if validation_date:
        split_idx = len(df[df.index <= validation_date])
    else:
        split_idx = int(len(df) * train_split)

    # #411 - route through the single chronological splitter (default
    # purge/embargo = 0, so this is bit-identical to the old inline math).
    split_idx, _ = split_bounds(len(df), split_idx)

    X_train = X.iloc[:split_idx]
    X_test = X.iloc[split_idx:]
    y_train = target.iloc[:split_idx]
    y_test = target.iloc[split_idx:]
    close_train = close.iloc[:split_idx]
    close_test = close.iloc[split_idx:]

    split_dt = str(df.index[split_idx])

    # Label encoding — always fit on [-1, 0, 1]
    enc = LabelEncoder()
    enc.fit([-1, 0, 1])
    y_train_enc = enc.transform(y_train)
    y_test_enc = enc.transform(y_test)

    return {
        "df": df, "X_train": X_train, "X_test": X_test,
        "y_train": y_train, "y_test": y_test,
        "y_train_enc": y_train_enc, "y_test_enc": y_test_enc,
        "enc": enc,
        "close": close, "close_train": close_train, "close_test": close_test,
        "split_idx": split_idx, "split_dt": split_dt,
        "n_train": len(X_train), "n_test": len(X_test),
    }


def compute_overlays(close, df_index):
    """Compute BB and MA overlays on full dataset. Always consistent.

    Returns: (bb_dict, ma_dict)
    """
    bb_mid = close.rolling(20).mean()
    bb_std = close.rolling(20).std()
    bb_upper = bb_mid + 2 * bb_std
    bb_lower = bb_mid - 2 * bb_std

    ma50 = close.rolling(50).mean()
    ma100 = close.rolling(100).mean()
    ma200 = close.rolling(200).mean()

    def _safe(s):
        s = s.reindex(df_index).bfill().ffill()
        return [float(x) if (x is not None and not np.isnan(x) and not np.isinf(x)) else None
                for x in s.values]

    bb = {"upper": _safe(bb_upper), "mid": _safe(bb_mid), "lower": _safe(bb_lower)}
    ma = {"ma50": _safe(ma50), "ma100": _safe(ma100), "ma200": _safe(ma200)}

    return bb, ma


def run_backtest(signal, close, capital=10000, cost=2e-5, leverage=1.0,
                 lot_size=None, balance=None, symbol="",
                 commission_usd=None, swap_long_usd=None, swap_short_usd=None,
                 spread_pips=0.0):
    """Run backtest with transaction costs.
    Uses price-based trade returns (same as webapp _compute_trades).
    Signal 0 = hold (keep current position), not close.

    Two accounting models:
      • `lot_size is None`  → LEGACY: per-trade return = leverage × fractional
        return-on-equity (clamp −100%); `pnl` is the raw price delta; no `$`.
        Direct callers (e.g. live_track / Server G) stay on this path.
      • `lot_size` set       → BROKER-ACCURATE (Phase 1): per-trade `pnl_usd` via
        position size × per-pair FX − commission/swap; `pnl_pct` becomes
        return-on-equity = pnl_usd / balance (leverage no longer multiplies).

    Returns: dict with equity, trade_returns, long_returns, short_returns, bar_returns
    """
    def _lev(r):
        return max(r * leverage, -1.0)

    broker = lot_size is not None
    bal = float(balance) if balance is not None else float(capital)
    if broker and bal <= 0:
        bal = float(capital) or 10000.0

    sig_arr = signal.values
    price_arr = close.values
    idx = signal.index
    n = len(price_arr)
    _idx_dates, _idx_dow = _index_date_arrays(idx) if broker else (None, None)
    _spread_px = float(spread_pips or 0.0) * _pip_size(symbol)  # bid/ask cost (#283)

    def _close_ret(direction, e_px, x_px, e_bar, x_bar):
        """Return (ret_for_equity, pnl_usd_or_None) for a closed trade."""
        if broker:
            nights = _nights_held(_idx_dates, _idx_dow, e_bar, x_bar)
            usd = _trade_pnl_usd(direction, e_px, x_px, lot_size, symbol, nights,
                                 commission_usd=commission_usd,
                                 swap_long_usd=swap_long_usd,
                                 swap_short_usd=swap_short_usd)
            # one full round-trip spread (buy@ask + sell@bid), in USD (#283)
            usd -= _spread_px * (lot_size or 0.0) * LOT_UNITS / _usd_conv(symbol, x_px)
            return max(usd / bal, -1.0), usd
        # legacy leverage mode: subtract spread as a fraction of entry price (#283)
        return _lev(float(direction * (x_px - e_px) / e_px - cost - _spread_px / e_px)), None

    # Trade returns — price-based (matches webapp _compute_trades exactly)
    trade_returns = []
    long_returns = []
    short_returns = []
    trade_log = []
    last_dir = None
    entry_price = None
    entry_bar = None

    for i in range(n):
        s = sig_arr[i]
        c = price_arr[i]
        if s != 0.0 and s != last_dir:
            # Direction change — close previous trade, open new
            if last_dir is not None and entry_price is not None and entry_price != 0:
                ret, _usd = _close_ret(last_dir, entry_price, c, entry_bar, i)
                trade_returns.append(ret)
                if last_dir == 1:
                    long_returns.append(ret)
                else:
                    short_returns.append(ret)
                trade_log.append({
                    "type": "Buy" if last_dir == 1 else "Sell",
                    "entry_time": str(idx[entry_bar]),
                    "exit_time": str(idx[i]),
                    "entry_price": round(entry_price, 5),
                    "exit_price": round(c, 5),
                    "pnl": round(last_dir * (c - entry_price), 5),
                    "pnl_usd": round(_usd, 2) if _usd is not None else None,
                    "pips": round(_pips(symbol, last_dir, entry_price, c), 1) if broker else None,
                    "pnl_pct": round(ret * 100, 3),
                    "exit_reason": "signal",
                })
            entry_price = c
            entry_bar = i
            last_dir = s

    # Close last open trade
    if last_dir is not None and entry_price is not None and n > 0 and entry_price != 0:
        c = price_arr[-1]
        ret, _usd = _close_ret(last_dir, entry_price, c, entry_bar, n - 1)
        trade_returns.append(ret)
        if last_dir == 1:
            long_returns.append(ret)
        else:
            short_returns.append(ret)
        trade_log.append({
            "type": "Buy" if last_dir == 1 else "Sell",
            "entry_time": str(idx[entry_bar]),
            "exit_time": str(idx[-1]),
            "entry_price": round(entry_price, 5),
            "exit_price": round(c, 5),
            "pnl": round(last_dir * (c - entry_price), 5),
            "pnl_usd": round(_usd, 2) if _usd is not None else None,
            "pips": round(_pips(symbol, last_dir, entry_price, c), 1) if broker else None,
            "pnl_pct": round(ret * 100, 3),
            "exit_reason": "end",
        })

    # Equity curve from trade returns
    cumret = 1.0
    equity_vals = np.full(n, bal if broker else float(capital))
    trade_idx = 0
    in_trade = False
    t_entry_price = None
    t_entry_bar = None
    t_dir = None

    for i in range(n):
        s = sig_arr[i]
        c = price_arr[i]
        if s != 0.0 and s != t_dir:
            if t_dir is not None and t_entry_price is not None and t_entry_price != 0:
                t_ret, _ = _close_ret(t_dir, t_entry_price, c, t_entry_bar, i)
                cumret *= (1 + t_ret)
            t_entry_price = c
            t_entry_bar = i
            t_dir = s
        equity_vals[i] = (bal if broker else capital) * cumret

    # Bar returns for Sharpe
    bar_returns = np.zeros(n)
    for i in range(1, n):
        if price_arr[i - 1] != 0 and last_dir is not None:
            bar_returns[i] = sig_arr[i - 1] * (price_arr[i] - price_arr[i - 1]) / price_arr[i - 1] if sig_arr[i - 1] != 0 else 0.0

    return {
        "equity": pd.Series(equity_vals, index=close.index),
        "trade_returns": trade_returns,
        "long_returns": long_returns,
        "short_returns": short_returns,
        "bar_returns": bar_returns,
        "trade_log": trade_log,
    }


def compute_trade_stats(trades, capital=10000):
    """Single source of truth for trade statistics.
    Every display path reads from this — no recomputation anywhere.
    All values are rounded and JSON-safe (no inf/nan).
    """
    if not trades:
        return {"n": 0, "wins": 0, "losses": 0, "wr": 0, "avg": 0,
                "best": 0, "worst": 0, "ret": 0, "np": 0, "mdd": 0,
                "pf": 0, "rr": 0, "expect": 0}
    w = [r for r in trades if r > 0]
    l = [r for r in trades if r < 0]
    cumret = 1.0
    for r in trades:
        cumret *= (1 + r)
    net_p = capital * (cumret - 1)
    # Max drawdown
    eq = np.cumprod([1.0] + [1 + r for r in trades])
    peak = np.maximum.accumulate(eq)
    mdd = float(((eq - peak) / peak).min()) if len(eq) > 1 else 0.0
    # Profit Factor
    gross_w = sum(w) if w else 0
    gross_l = abs(sum(l)) if l else 0
    pf = gross_w / gross_l if gross_l > 0 else (9999.0 if gross_w > 0 else 0)
    # Risk:Reward
    avg_w = float(np.mean(w)) if w else 0
    avg_l = abs(float(np.mean(l))) if l else 0
    rr = avg_w / avg_l if avg_l > 0 else (9999.0 if avg_w > 0 else 0)
    # Expectancy
    expect = net_p / len(trades)
    return {
        "n": len(trades), "wins": len(w), "losses": len(l),
        "wr": round(len(w) / len(trades), 4),
        "avg": round(float(np.mean(trades)), 6),
        "best": round(max(w), 6) if w else 0,
        "worst": round(min(l), 6) if l else 0,
        "ret": round(cumret - 1, 6),
        "np": round(net_p, 2),
        "mdd": round(mdd, 6),
        "pf": round(pf, 2),
        "rr": round(rr, 2),
        "expect": round(expect, 2),
    }


def compute_metrics(bt_result, close_test, capital=10000):
    """Compute all standard metrics from backtest result.
    Uses trade-level compounding (same as webapp _trade_stats) for accuracy.

    Returns: dict with total_ret, bh_ret, sharpe_strat, sharpe_bh, mdd, n_trades
    """
    equity = bt_result["equity"]
    trade_returns = bt_result["trade_returns"]

    # Total return — trade-level compounding (matches webapp)
    if trade_returns:
        cumret = 1.0
        for r in trade_returns:
            cumret *= (1 + r)
        total_ret = cumret - 1
    else:
        total_ret = 0.0

    # Buy and hold
    bh_equity = capital * (close_test / close_test.iloc[0])
    bh_ret = (bh_equity.iloc[-1] - capital) / capital if capital != 0 else 0.0

    # Sharpe ratio — trade-level (matches webapp: sqrt(252*26) annualization)
    if len(trade_returns) >= 2 and float(np.std(trade_returns)) > 0:
        sharpe_strat = float(np.mean(trade_returns) / np.std(trade_returns) * np.sqrt(252 * 26))
    else:
        sharpe_strat = 0.0

    bh_rets = bh_equity.pct_change().dropna()
    if len(bh_rets) > 1 and bh_rets.std() != 0:
        sharpe_bh = float((bh_rets.mean() / bh_rets.std()) * np.sqrt(252 * 24 * 4))
    else:
        sharpe_bh = 0.0

    # Max drawdown — trade-level (matches webapp)
    if trade_returns:
        eq = np.cumprod([1.0] + [1 + r for r in trade_returns])
        peak = np.maximum.accumulate(eq)
        mdd = float(((eq - peak) / peak).min()) if len(eq) > 1 else 0.0
    else:
        mdd = 0.0

    return {
        "total_ret": float(total_ret),
        "bh_ret": float(bh_ret),
        "sharpe_strat": float(sharpe_strat) if not np.isnan(sharpe_strat) else 0.0,
        "sharpe_bh": float(sharpe_bh) if not np.isnan(sharpe_bh) else 0.0,
        "mdd": float(mdd),
        "n_trades": len(trade_returns),
    }


# Diagnostics line/histogram series (equity / drawdown / rolling_acc / conf_hist)
# only feed the small Diagnostics charts — they're never used by the price chart
# or scroll-back. On a 1-min model trained over the (2.2-capped) window these are
# still ~30k points each; downsample to a visually-identical resolution before the
# dict leaves the trainer so it doesn't carry that into Server-A RAM / Postgres.
_RESULTS_SERIES_MAX = 5000


def _downsample_idx(n, cap=_RESULTS_SERIES_MAX):
    """Evenly-spaced index list spanning [0, n-1] (first+last always kept), or
    None when no downsampling is needed (n <= cap)."""
    if n <= cap:
        return None
    return np.unique(np.linspace(0, n - 1, cap).astype(int)).tolist()


def _take(arr, idx):
    """Subset a list by an index list (idx may be None → return arr unchanged)."""
    if idx is None or not isinstance(arr, list):
        return arr
    return [arr[i] for i in idx]


# trade_log / train_trade_log are lists of per-trade dicts (display-only — the
# Trade Log tab). They scale with TRADE count, not bar count, so the bar-window
# cap (Phase 2.2) doesn't bound them — a degenerate near-every-bar model can put
# 10k+ trade dicts in the blob (>3 MB). Cap each (independently — a small-N model
# keeps every trade) to the most-recent N, recording `*_total` + `*_truncated`
# so the true count is still reported. Real strategies have far fewer than
# _TRADE_LOG_MAX trades, so this only ever bites pathological models.
_TRADE_LOG_MAX = 5000


def _cap_trade_log(tl):
    """Return (capped_list, original_len, was_truncated)."""
    if not isinstance(tl, list) or len(tl) <= _TRADE_LOG_MAX:
        return tl, (len(tl) if isinstance(tl, list) else 0), False
    return tl[-_TRADE_LOG_MAX:], len(tl), True


def _safe_fi_pairs(model, feature_cols, top=15):
    """Feature-importance pairs that never crash (FACT-F2 / #3).

    A non-tree model wrapped by a stale/diverged ModelWrapper can hand back
    ``None`` (LogReg has ``coef_``, not ``feature_importances_``). Guard here so
    build_return_dict is safe regardless of which wrapper produced the PKL: fall
    back to ``coef_``, then to zero-weight pairs aligned to ``feature_cols``.
    """
    importances = getattr(model, "feature_importances_", None)
    if importances is None:
        coef = getattr(model, "coef_", None)
        if coef is not None:
            importances = np.abs(np.asarray(coef)).mean(axis=0)
    if importances is None or len(importances) != len(feature_cols):
        # length mismatch or nothing usable -> zero weights, still renders
        return [(f, 0.0) for f in feature_cols][-top:]
    return sorted(zip(feature_cols, importances), key=lambda x: x[1])[-top:]


def build_return_dict(split_result, bt_result, metrics, model, feature_cols,
                      signal_full, p_pos_test, p_neg_test, custom_figs=None,
                      bt_train_result=None, pre_stats=None):
    """Assemble the complete return dict. Handles ALL serialization.
    Never returns Timestamps, numpy arrays, or non-JSON types.

    Returns: JSON-safe dict with all required keys
    """
    df = split_result["df"]
    close = split_result["close"]
    close_test = split_result["close_test"]
    X_test = split_result["X_test"]
    y_test = split_result["y_test"]
    equity = bt_result["equity"]
    bar_returns = bt_result["bar_returns"]

    # OHLC
    ohlc_dates = [str(x) for x in df.index.tolist()]

    def _safe_list(arr):
        return [float(x) if (x is not None and not np.isnan(x) and not np.isinf(x)) else None
                for x in arr]

    # Overlays
    bb, ma = compute_overlays(close, df.index)

    # Buy and hold equity
    capital = equity.iloc[0] if len(equity) > 0 else 10000
    bh_equity = capital * (close_test / close_test.iloc[0])

    # Confusion matrix
    from sklearn.metrics import confusion_matrix
    pred_test = model.predict(X_test)
    y_test_arr = np.asarray(y_test)
    cm = confusion_matrix(y_test_arr, pred_test, labels=[-1, 0, 1])

    # Rolling accuracy
    sig_arr = signal_full.reindex(close_test.index).values
    correct = pd.Series((pred_test == y_test_arr).astype(float), index=X_test.index)
    active_test = pd.Series(sig_arr != 0, index=close_test.index) if len(sig_arr) == len(close_test) else pd.Series(True, index=close_test.index)
    correct_active = correct.where(active_test, other=np.nan)
    rolling_acc = correct_active.rolling(30, min_periods=1).mean()

    # Feature importance (FACT-F2 guard: never crash on None/short importances)
    fi_pairs = _safe_fi_pairs(model, feature_cols)

    # Drawdown
    rolling_max = equity.cummax()
    drawdown = (equity - rolling_max) / rolling_max.replace(0, np.nan)
    drawdown = drawdown.fillna(0.0)

    # ── Downsample the Diagnostics-only series (see _downsample_idx) ──────────
    _eq_dates = [str(x) for x in close_test.index.tolist()]
    _eq_strat = _safe_list(equity.values)
    _eq_bh    = _safe_list(bh_equity.values)
    _eq_idx   = _downsample_idx(len(_eq_dates))
    _eq_dates, _eq_strat, _eq_bh = _take(_eq_dates, _eq_idx), _take(_eq_strat, _eq_idx), _take(_eq_bh, _eq_idx)

    _ra_dates = [str(x) for x in rolling_acc.index.tolist()]
    _ra_vals  = [float(x) if (not np.isnan(x) and not np.isinf(x)) else None for x in rolling_acc.values]
    _ra_idx   = _downsample_idx(len(_ra_dates))
    _ra_dates, _ra_vals = _take(_ra_dates, _ra_idx), _take(_ra_vals, _ra_idx)

    _dd_dates = [str(x) for x in drawdown.index.tolist()]
    _dd_vals  = _safe_list(drawdown.values)
    _dd_idx   = _downsample_idx(len(_dd_dates))
    _dd_dates, _dd_vals = _take(_dd_dates, _dd_idx), _take(_dd_vals, _dd_idx)

    _cp_pos = [float(x) for x in (p_pos_test.tolist() if hasattr(p_pos_test, 'tolist') else list(p_pos_test))]
    _cp_neg = [float(x) for x in (p_neg_test.tolist() if hasattr(p_neg_test, 'tolist') else list(p_neg_test))]
    _cp_pos = _take(_cp_pos, _downsample_idx(len(_cp_pos)))
    _cp_neg = _take(_cp_neg, _downsample_idx(len(_cp_neg)))

    # ── Trade logs — display-only (Trade Log tab); cap to most-recent N with a
    #    `_total` field so the true count is still reported (see _cap_trade_log).
    #    NB: ret_dist arrays are left FULL — a downstream path in callbacks.py
    #    recomputes n_trades/win-rate from len(ret_dist), so a sample would skew
    #    the displayed counts; they're small anyway and gzip handles them.
    _tl_test, _tl_test_n, _tl_test_tr = _cap_trade_log(bt_result.get("trade_log", []))
    _tl_tr,   _tl_tr_n,   _tl_tr_tr   = _cap_trade_log(bt_train_result.get("trade_log", []) if bt_train_result else [])

    return {
        "ohlc": {
            "dates": ohlc_dates,
            "open": _safe_list(df["open"].values),
            "high": _safe_list(df["high"].values),
            "low": _safe_list(df["low"].values),
            "close": _safe_list(df["close"].values),
        },
        "signals": {
            "dates": [str(x) for x in signal_full.index.tolist()],
            "values": [float(x) for x in signal_full.values],
        },
        "bb": bb,
        "ma": ma,
        "equity": {
            "dates": _eq_dates,
            "strategy": _eq_strat,
            "bh": _eq_bh,
        },
        "feature_importance": {
            "names": [p[0] for p in fi_pairs],
            "values": [float(p[1]) for p in fi_pairs],
        },
        "conf_matrix": cm.tolist(),
        "conf_hist": {
            "p_pos": _cp_pos,
            "p_neg": _cp_neg,
        },
        "rolling_acc": {
            "dates": _ra_dates,
            "values": _ra_vals,
        },
        "drawdown": {
            "dates": _dd_dates,
            "values": _dd_vals,
        },
        "ret_dist": [float(x) for x in bt_result["trade_returns"]],
        "ret_dist_long": [float(x) for x in bt_result["long_returns"]],
        "ret_dist_short": [float(x) for x in bt_result["short_returns"]],
        "train_ret_dist": [float(x) for x in bt_train_result["trade_returns"]] if bt_train_result else [],
        "train_ret_dist_long": [float(x) for x in bt_train_result["long_returns"]] if bt_train_result else [],
        "train_ret_dist_short": [float(x) for x in bt_train_result["short_returns"]] if bt_train_result else [],
        "trade_log": _tl_test,
        "train_trade_log": _tl_tr,
        "trade_log_total": _tl_test_n,
        "train_trade_log_total": _tl_tr_n,
        "trade_log_truncated": _tl_test_tr,
        "train_trade_log_truncated": _tl_tr_tr,
        **(pre_stats or {}),
        "metrics": metrics,
        "split_dt": split_result["split_dt"],
        "split_idx": int(split_result["split_idx"]),
        "n_train": int(split_result["n_train"]),
        "n_test": int(split_result["n_test"]),
        "feature_cols": list(feature_cols),
        "custom_figs": custom_figs or [],
    }


# ════════════════════════════════════════════════════════════════════════════
# STRATEGY FRAMEWORK v2 — Config-driven architecture
# Claude writes feature_engineering() + strategy_config(). Framework does rest.
# ════════════════════════════════════════════════════════════════════════════

import importlib

_MODEL_REGISTRY = {
    "XGBClassifier":              ("xgboost", "XGBClassifier"),
    "RandomForestClassifier":     ("sklearn.ensemble", "RandomForestClassifier"),
    "GradientBoostingClassifier": ("sklearn.ensemble", "GradientBoostingClassifier"),
    "LogisticRegression":         ("sklearn.linear_model", "LogisticRegression"),
    "ExtraTreesClassifier":       ("sklearn.ensemble", "ExtraTreesClassifier"),
    "AdaBoostClassifier":         ("sklearn.ensemble", "AdaBoostClassifier"),
}


RULE_MODEL_TYPE = "RuleBased"


def _build_model_from_config(config, X_train, y_train_enc):
    """Build, fit, and wrap a model from strategy_config dict."""
    model_type = config.get("model_type", "RandomForestClassifier")
    model_params = dict(config.get("model_params", {}))

    # #374 — rule-based mode: a fixed human-written rule, NOT a fitted model.
    # Nothing is trained; y_train_enc is deliberately ignored. predict_proba
    # returns hard 0/1 so signal_threshold cannot alter a decision and reruns
    # are bit-identical. Wrapped in ModelWrapper so the PKL/deploy/predictor
    # path is byte-for-byte the same shape as a trained model's.
    if model_type == RULE_MODEL_TYPE:
        from model_wrapper import ModelWrapper, RuleModel
        rules = config.get("rules") or []
        if not rules:
            raise ValueError(
                "model_type 'RuleBased' requires a non-empty 'rules' list in "
                "strategy_config() — e.g. "
                "[{'signal': 1, 'when': [{'feature': 'rsi14', 'op': '<', 'value': 30}]}]"
            )
        rm = RuleModel(rules, n_features=X_train.shape[1])
        rm.rule_signal(X_train.head(1))   # fail fast on a bad feature/op
        return ModelWrapper(rm, original_classes=np.array([-1, 0, 1]),
                            n_features=X_train.shape[1])

    if model_type not in _MODEL_REGISTRY:
        raise ValueError(f"Unknown model_type '{model_type}'. "
                         f"Valid: {list(_MODEL_REGISTRY.keys()) + [RULE_MODEL_TYPE]}")

    module_path, class_name = _MODEL_REGISTRY[model_type]
    mod = importlib.import_module(module_path)
    cls = getattr(mod, class_name)

    # XGBoost defaults
    if class_name == "XGBClassifier":
        model_params.setdefault("use_label_encoder", False)
        model_params.setdefault("eval_metric", "mlogloss")
        model_params.setdefault("tree_method", "hist")
        # Determinism > speed (2026-05-25). XGBoost hist with n_jobs=-1 is
        # NON-reproducible even with random_state set — the parallel histogram
        # gradient-sum order varies across threads, so the SAME code + data
        # gives a slightly different model (and backtest) every run. Forcing
        # single-thread makes training bit-reproducible so: (a) a user who
        # copies a strategy and reruns it gets identical numbers, (b) the
        # community "Live" score matches a redeploy, (c) "same code, different
        # result" support reports go away. Cost: single-threaded XGB (a few
        # seconds slower on large windows; hist is fast so it's minor). FORCED
        # (not setdefault) so the guarantee can't be silently broken by a
        # strategy passing n_jobs. Exact reproducibility holds within the
        # platform (pinned versions / same Modal image); a user's own machine
        # with different xgboost/numpy/CPU can still differ in low-order bits.
        model_params["n_jobs"] = 1

    # Common defaults
    model_params.setdefault("random_state", 42)

    from model_wrapper import ModelWrapper

    clf = cls(**model_params)

    # #405: label codes are deliberately FIXED at 0/1/2 for -1/0/1 so a code
    # means the same thing in every model, run and deployment. The cost is
    # that a window missing a class hands the fit a non-contiguous set - a
    # real prod case, GBPUSD 1h with no flat bars gives {0, 2}. XGBoost
    # rejects that outright ("Expected: [0 1], got [0 2]"); sklearn accepts it
    # but then returns one proba column per PRESENT class, which downstream
    # code - indexing by global class position - misreads without noticing.
    # So: fit on a dense remap, and tell the wrapper what each column means.
    _present = np.unique(np.asarray(y_train_enc))
    if len(_present) < 2:
        raise ValueError(
            f"The training window contains only one label class "
            f"({_present.tolist()}), so there is nothing for a classifier to "
            f"separate. Widen the training window, or lower target_horizon / "
            f"the labelling threshold so both up and down moves appear.")
    _dense = {int(c): i for i, c in enumerate(_present)}
    _y_fit = np.asarray([_dense[int(v)] for v in np.asarray(y_train_enc)])
    clf.fit(X_train, _y_fit)

    enc = LabelEncoder()
    enc.fit([-1, 0, 1])
    return ModelWrapper(clf, original_classes=enc.classes_,
                        n_features=X_train.shape[1],
                        class_positions=_present)


def _generate_signals(model, X, threshold):
    """Framework-owned signal generation. Deterministic threshold logic."""
    proba = model.predict_proba(X)
    classes = list(model.classes_)

    idx_pos = classes.index(1) if 1 in classes else None
    idx_neg = classes.index(-1) if -1 in classes else None

    p_pos = proba[:, idx_pos] if idx_pos is not None else np.zeros(len(X))
    p_neg = proba[:, idx_neg] if idx_neg is not None else np.zeros(len(X))

    signal_vals = np.zeros(len(X))
    signal_vals = np.where(p_pos >= threshold, 1.0, signal_vals)
    signal_vals = np.where(p_neg >= threshold, -1.0, signal_vals)
    # Both exceed: pick stronger
    both = (p_pos >= threshold) & (p_neg >= threshold)
    signal_vals[both] = np.where(p_pos[both] >= p_neg[both], 1.0, -1.0)

    return pd.Series(signal_vals, index=X.index), p_pos, p_neg


# ── Filter functions (all no-ops when config value is None) ──────────────

def _apply_direction_filter(signal, direction):
    """Zero out signals that don't match allowed direction."""
    if direction is None or direction == "both":
        return signal
    s = signal.copy()
    if direction == "long":
        s[s < 0] = 0.0
    elif direction == "short":
        s[s > 0] = 0.0
    return s


def _apply_session_filter(signal, index, session_hours):
    """Zero out signals outside session hours [start, end] UTC."""
    if session_hours is None:
        return signal
    s = signal.copy()
    start_h, end_h = session_hours[0], session_hours[1]
    hours = index.hour
    if start_h <= end_h:
        mask = (hours >= start_h) & (hours < end_h)
    else:  # wrap around midnight, e.g. [22, 6]
        mask = (hours >= start_h) | (hours < end_h)
    s[~mask] = 0.0
    return s


def _apply_atr_filter(signal, close, high, low, min_atr):
    """Zero out signals when NATR(14) is below threshold."""
    if min_atr is None:
        return signal
    hl = high - low
    hc = (high - close.shift(1)).abs()
    lc = (low - close.shift(1)).abs()
    tr = pd.concat([hl, hc, lc], axis=1).max(axis=1)
    atr14 = tr.ewm(com=13, adjust=False).mean()
    natr = atr14 / close.replace(0, np.nan)
    s = signal.copy()
    s[natr < min_atr] = 0.0
    return s


def _apply_trend_filter(signal, close, trend_filter):
    """Only allow signals aligned with trend. e.g. 'sma_50': longs above SMA, shorts below."""
    if trend_filter is None:
        return signal
    # Parse: "sma_50" → SMA with period 50
    parts = trend_filter.lower().replace("-", "_").split("_")
    if len(parts) >= 2 and parts[0] in ("sma", "ema"):
        period = int(parts[1])
    else:
        return signal  # unknown filter, skip
    if parts[0] == "sma":
        trend_line = close.rolling(period).mean()
    else:
        trend_line = close.ewm(span=period, adjust=False).mean()
    s = signal.copy()
    # Longs only above trend, shorts only below
    s[(s > 0) & (close < trend_line)] = 0.0
    s[(s < 0) & (close > trend_line)] = 0.0
    return s


# ── run_backtest_v2: framework-owned SL/TP/cooldown/position management ──

def run_backtest_v2(signal, close, high, low, config, capital=10000, cost=2e-5, leverage=1.0,
                    lot_size=None, balance=None, symbol="",
                    commission_usd=None, swap_long_usd=None, swap_short_usd=None,
                    spread_pips=None):
    """Backtest with SL/TP/cooldown/direction handling built into the engine.
    Unlike run_backtest (v1), this function handles position exits internally.

    Two accounting models (see run_backtest for the full contract):
      • `lot_size is None`  → LEGACY leverage×fractional return-on-equity; `pnl`
        is the raw price delta. Direct callers (live_track / Server G) stay here.
      • `lot_size` set       → BROKER-ACCURATE: per-trade `pnl_usd` (position size
        × per-pair FX − commission/swap); `pnl_pct` = pnl_usd / balance.

    Returns: same dict shape as run_backtest()
    """
    def _lev(r):
        # Clamp at -100% so a single trade can't drive equity negative / flip
        # the cumprod sign (SL'd trades are bounded; only un-SL'd adverse
        # closes could approach the clamp).
        return max(r * leverage, -1.0)

    stop_loss = config.get("stop_loss")
    take_profit = config.get("take_profit")
    cooldown = config.get("cooldown", 0)
    on_opposite = config.get("on_opposite", "reverse")

    # FACT9 Phase 2 — SL/TP unit. 'pct' (legacy/default) = fraction of price
    # (entry × level). 'pips' = pip distance (level × pip_size). 'usd' = the
    # price move whose pre-cost USD P&L equals `level` dollars at this lot. The
    # pips/usd modes need a position size → only meaningful in broker mode
    # (lot_size set); 'pct' reproduces the original entry × (1 ± level) exactly.
    risk_unit = (config.get("risk_unit") or "pct").lower()
    _pip_sz = _pip_size(symbol)
    _risk_units = (lot_size or 0.0) * LOT_UNITS

    # Bid/ask spread cost (#283). Brokers fill buys at ask / sells at bid, so a
    # round-trip pays one full spread; the engine otherwise fills at the raw
    # close (mid), silently overstating P&L. `spread_pips` = full round-trip
    # spread in pips, charged once per closed trade. Default 0.0 → no change to
    # existing backtests. Realistic EURUSD: ~0.1 pip raw/ECN, ~1.0 pip retail.
    # Precedence: explicit spread_pips arg > config["spread_pips"] > 0.0.
    _sp = spread_pips if spread_pips is not None else config.get("spread_pips")
    _spread_px = float(_sp or 0.0) * _pip_sz

    def _risk_dist(level, e_px):
        """SL/TP price distance for `level` at entry `e_px`, per risk_unit."""
        if level is None:
            return None
        if risk_unit == "pips":
            return level * _pip_sz
        if risk_unit == "usd":
            return (level * _usd_conv(symbol, e_px) / _risk_units) if _risk_units > 0 else e_px * level
        return e_px * level  # 'pct' (legacy)

    sig_arr = signal.values
    close_arr = close.values
    high_arr = high.values
    low_arr = low.values
    idx = signal.index
    n = len(close_arr)

    broker = lot_size is not None
    bal = float(balance) if balance is not None else float(capital)
    if broker and bal <= 0:
        bal = float(capital) or 10000.0
    _idx_dates, _idx_dow = _index_date_arrays(idx) if broker else (None, None)
    _base = bal if broker else float(capital)

    trade_returns = []
    long_returns = []
    short_returns = []
    trade_log = []
    equity_vals = np.full(n, _base)
    cumret = 1.0

    position = 0.0       # current direction: 1.0, -1.0, or 0.0 (flat)
    entry_price = None
    entry_bar = None      # index into arrays for entry time
    cooldown_remaining = 0

    def _close_ret(direction, e_px, x_px, e_bar, x_bar):
        """Return (ret_for_equity, pnl_usd_or_None) for a closed trade."""
        if broker:
            nights = _nights_held(_idx_dates, _idx_dow, e_bar, x_bar)
            usd = _trade_pnl_usd(direction, e_px, x_px, lot_size, symbol, nights,
                                 commission_usd=commission_usd,
                                 swap_long_usd=swap_long_usd,
                                 swap_short_usd=swap_short_usd)
            # one full round-trip spread (buy@ask + sell@bid), in USD (#283)
            usd -= _spread_px * _risk_units / _usd_conv(symbol, x_px)
            return max(usd / bal, -1.0), usd
        # legacy leverage mode: subtract spread as a fraction of entry price (#283)
        return _lev(float(direction * (x_px - e_px) / e_px - cost - _spread_px / e_px)), None

    def _log_trade(exit_bar, exit_px, ret, reason, usd=None):
        trade_log.append({
            "type": "Buy" if position == 1.0 else "Sell",
            "entry_time": str(idx[entry_bar]),
            "exit_time": str(idx[exit_bar]),
            "entry_price": round(entry_price, 5),
            "exit_price": round(exit_px, 5),
            "pnl": round(position * (exit_px - entry_price), 5),
            "pnl_usd": round(usd, 2) if usd is not None else None,
            "pips": round(_pips(symbol, position, entry_price, exit_px), 1) if broker else None,
            "pnl_pct": round(ret * 100, 3),
            "exit_reason": reason,
        })

    for i in range(n):
        c = close_arr[i]
        h = high_arr[i]
        lo = low_arr[i]
        s = sig_arr[i]

        # 1. Check SL/TP if in trade
        if position != 0.0 and entry_price is not None:
            hit_sl = False
            hit_tp = False
            exit_price = None

            d_sl = _risk_dist(stop_loss, entry_price)
            d_tp = _risk_dist(take_profit, entry_price)
            if position == 1.0:  # long
                if d_sl is not None and lo <= entry_price - d_sl:
                    hit_sl = True
                    exit_price = entry_price - d_sl
                elif d_tp is not None and h >= entry_price + d_tp:
                    hit_tp = True
                    exit_price = entry_price + d_tp
            else:  # short
                if d_sl is not None and h >= entry_price + d_sl:
                    hit_sl = True
                    exit_price = entry_price + d_sl
                elif d_tp is not None and lo <= entry_price - d_tp:
                    hit_tp = True
                    exit_price = entry_price - d_tp

            if hit_sl or hit_tp:
                ret, _usd = _close_ret(position, entry_price, exit_price, entry_bar, i)
                trade_returns.append(ret)
                if position == 1.0:
                    long_returns.append(ret)
                else:
                    short_returns.append(ret)
                _log_trade(i, exit_price, ret, "SL" if hit_sl else "TP", usd=_usd)
                cumret *= (1 + ret)
                position = 0.0
                entry_price = None
                entry_bar = None
                cooldown_remaining = cooldown
                equity_vals[i] = _base * cumret
                continue

        # 2. Cooldown
        if cooldown_remaining > 0:
            cooldown_remaining -= 1
            equity_vals[i] = _base * cumret
            continue

        # 3. Signal processing
        if s != 0.0:
            if position == 0.0:
                # Open new trade
                position = s
                entry_price = c
                entry_bar = i
            elif s != position:
                # Opposite signal
                if on_opposite == "reverse":
                    # Close current + open opposite
                    ret, _usd = _close_ret(position, entry_price, c, entry_bar, i)
                    trade_returns.append(ret)
                    if position == 1.0:
                        long_returns.append(ret)
                    else:
                        short_returns.append(ret)
                    _log_trade(i, c, ret, "signal", usd=_usd)
                    cumret *= (1 + ret)
                    position = s
                    entry_price = c
                    entry_bar = i
                else:  # close_only
                    # Close current, go flat
                    ret, _usd = _close_ret(position, entry_price, c, entry_bar, i)
                    trade_returns.append(ret)
                    if position == 1.0:
                        long_returns.append(ret)
                    else:
                        short_returns.append(ret)
                    _log_trade(i, c, ret, "close_only", usd=_usd)
                    cumret *= (1 + ret)
                    position = 0.0
                    entry_price = None
                    entry_bar = None
                    cooldown_remaining = cooldown

        equity_vals[i] = _base * cumret

    # Close last open trade at final close
    if position != 0.0 and entry_price is not None and n > 0 and entry_price != 0:
        c = close_arr[-1]
        ret, _usd = _close_ret(position, entry_price, c, entry_bar, n - 1)
        trade_returns.append(ret)
        if position == 1.0:
            long_returns.append(ret)
        else:
            short_returns.append(ret)
        _log_trade(n - 1, c, ret, "end", usd=_usd)
        cumret *= (1 + ret)
        equity_vals[-1] = _base * cumret

    # Bar returns for Sharpe (approximate)
    bar_returns = np.zeros(n)
    for i in range(1, n):
        if close_arr[i - 1] != 0 and sig_arr[i - 1] != 0:
            bar_returns[i] = sig_arr[i - 1] * (close_arr[i] - close_arr[i - 1]) / close_arr[i - 1]

    return {
        "equity": pd.Series(equity_vals, index=close.index),
        "trade_returns": trade_returns,
        "long_returns": long_returns,
        "short_returns": short_returns,
        "bar_returns": bar_returns,
        "trade_log": trade_log,
    }


# ── run_strategy: the v2 orchestrator ────────────────────────────────────

def run_strategy(feature_fn, config_fn, data_path, start_date="", end_date="",
                 validation_date="", train_split=0.7, register_model_fn=None,
                 leverage=30.0, lot_size=None, balance=None, symbol="",
                 stop_loss=None, take_profit=None, risk_unit=None,
                 signal_threshold=None,
                 commission_usd=None, swap_long_usd=None, swap_short_usd=None,
                 spread_pips=None, aux_feeds=None,
                 purge_bars=0, embargo_bars=0):
    """Config-driven strategy execution. Claude writes feature_fn + config_fn,
    framework does everything else.

    `lot_size`/`balance`/`symbol` drive the broker-accurate $ accounting (Phase 1
    Factual Trading Product). `lot_size is None` keeps the LEGACY leverage path —
    direct callers that omit it (e.g. live_track / Server G) are unchanged. The
    synthesized train wrapper passes lot_size=1.0 by default → broker mode.

    `aux_feeds` (#387/#388): explicit list of {"symbol","timeframe"} dicts, or
    None to fall back to config_fn()["aux_feeds"]. An explicit [] means "no
    aux feeds" and does NOT fall through to config. Normalized against the
    RUN-TIME primary (symbol/data_path), which is how a pasted strategy whose
    aux feed collides with a dropdown-selected primary gets that feed dropped.

    `purge_bars`/`embargo_bars` (#411): passed straight through to
    split_bounds() at the train/test split below. Defaults (0, 0) reproduce
    the legacy split exactly - the headline 70/30 backtest is unaffected.

    Returns: results dict (same format as webapp expects)
    """
    config = config_fn()

    # #244 — an explicitly-passed confidence/signal threshold (API
    # confidence_threshold or Config tab) is the source of truth: override the
    # generated strategy_config() value. None = strategy's own value stands.
    if signal_threshold is not None:
        config["signal_threshold"] = float(signal_threshold)

    # FACT9 Phase 2 — Config-tab SL/TP are the source of truth: override the
    # generated strategy_config() when the caller (synthesized wrapper) provides
    # them. risk_unit selects interpretation downstream (pct/pips/usd). In
    # pips/usd mode Config OWNS the levels — any pct-scale value strategy_config()
    # left behind is dropped (a 0.005 ratio must never be read as 0.005 pips).
    if risk_unit is not None:
        _ru = str(risk_unit).lower()
        config["risk_unit"] = _ru
        if _ru in ("pips", "usd"):
            config["stop_loss"] = stop_loss
            config["take_profit"] = take_profit
        else:
            if stop_loss is not None:
                config["stop_loss"] = stop_loss
            if take_profit is not None:
                config["take_profit"] = take_profit
    else:
        if stop_loss is not None:
            config["stop_loss"] = stop_loss
        if take_profit is not None:
            config["take_profit"] = take_profit

    # Broker-accounting params. symbol falls back to the SYMBOL parsed from the
    # data path (e.g. .../USDJPY_15min.parquet) so per-pair FX conversion works
    # even when the caller doesn't pass it explicitly.
    _broker = lot_size is not None
    _sym = symbol or config.get("symbol") or ""
    if not _sym:
        _sym, _ = _parse_symbol_tf_from_path(data_path)
        _sym = _sym or ""
    _bal = float(balance) if balance is not None else 10000.0
    _cap = _bal if _broker else 10000

    # Auto-correct SL/TP ONLY in legacy price-% mode (Claude passing 50 for 0.5).
    # pips/usd units are absolute magnitudes (25 pips, $250) → never divide them.
    if (config.get("risk_unit") or "pct").lower() == "pct":
        for _key in ("stop_loss", "take_profit"):
            _val = config.get(_key)
            if _val is not None and _val > 0.1:  # >10% is almost certainly a percentage
                config[_key] = _val / 100.0
                print(f"[strategy] Auto-corrected {_key}: {_val} -> {config[_key]} (was percentage, converted to decimal)")

    # 1. Load data
    df, close, open_, high, low = load_ohlc(data_path, start_date, end_date)

    # 1b. #387/#388 - merge aux feeds BEFORE feature_engineering. Run-time
    # normalization (against _sym/_tf, the RUN-TIME primary) is where a
    # pasted strategy whose aux feed now equals the dropdown-selected
    # primary gets that feed dropped - one rule, no special case at paste
    # time. Precedence: explicit aux_feeds kwarg (incl. []) wins; None falls
    # back to config_fn()["aux_feeds"].
    _aux_raw = aux_feeds if aux_feeds is not None else config.get("aux_feeds")
    _tf = _parse_symbol_tf_from_path(data_path)[1] or "15min"
    _aux_specs, _aux_warns = normalize_aux_specs(_aux_raw or [], _sym, _tf)
    df, _aux_report = merge_aux_feeds(df, _aux_specs, _sym, _tf)
    for _w in _aux_warns:
        _aux_report["dropped"].append({"reason": _w})
    # #403: say it out loud. A dropped feed is caught into the report and the
    # run continues (correct - a missing aux feed must not kill a train), but
    # with nothing printed, an aux feature that never loaded on Modal looked
    # exactly like a successful run for a full day. The log line is the only
    # thing standing between "degraded" and "undetectably broken".
    if _aux_report["dropped"]:
        for _d in _aux_report["dropped"]:
            print(f"[strategy] aux feed DROPPED: {_d}")
    if _aux_report["merged"]:
        print(f"[strategy] aux feeds merged: {_aux_report['merged']}")
    # merge_aux_feeds returns a (possibly new, e.g. via sort_index) df object
    # when aux feeds were merged - refresh close/open_/high/low so they read
    # from the SAME frame handed to feature_fn rather than stale references
    # into the pre-merge object (values are unaffected either way - the
    # merge never touches primary OHLC or drops rows - but identity must
    # match, not just values, for feature_fn to see a self-consistent df).
    close = df["close"]
    open_ = df["open"]
    high = df["high"]
    low = df["low"]

    # 2. Feature engineering (Claude's function)
    df = feature_fn(df, close, open_, high, low)
    close = df["close"]
    open_ = df["open"]
    high = df["high"]
    low = df["low"]

    # 3. Warm-up detection: drop rows where features have NaN BEFORE any fill
    feature_cols = [c for c in df.columns if c not in ("open", "high", "low", "close")]
    raw_nans = df[feature_cols].isna().any(axis=1)
    valid_rows = ~raw_nans
    if valid_rows.any():
        first_valid = valid_rows.idxmax()
        if raw_nans.loc[:first_valid].any():
            df = df.loc[first_valid:].copy()
            close = df["close"]
            open_ = df["open"]
            high = df["high"]
            low = df["low"]

    # 4. Target
    horizon = config.get("target_horizon", 4)
    target = make_target(close, horizon=horizon)

    # 5. Split (ffill only within each partition — no bfill leak)
    mask = target.notna()
    df = df[mask].copy()
    target = target[mask]
    close = df["close"]
    high = df["high"]
    low = df["low"]

    X = df[feature_cols].copy()
    X = X.replace([np.inf, -np.inf], np.nan)

    if validation_date:
        split_idx = len(df[df.index <= validation_date])
    else:
        split_idx = int(len(df) * train_split)
    # #411 - the ONE chronological splitter. purge_bars trims the tail of
    # train so no row's forward-looking label (make_target) reaches into
    # test; embargo_bars pushes test_start out further. Defaults (0, 0)
    # collapse train_end == test_start == split_idx, i.e. bit-identical to
    # the legacy split.
    train_end, test_start = split_bounds(len(df), split_idx,
                                         purge_bars, embargo_bars)
    split_idx = test_start  # split_dt / n_test keep their existing meaning

    # ffill within train and test separately (no leak)
    X_train = X.iloc[:train_end].ffill().fillna(0.0)
    X_test = X.iloc[test_start:].ffill().fillna(0.0)
    X = pd.concat([X_train, X_test])

    y_train = target.iloc[:train_end]
    y_test = target.iloc[test_start:]
    close_train = close.iloc[:train_end]
    close_test = close.iloc[test_start:]
    high_test = high.iloc[test_start:]
    low_test = low.iloc[test_start:]

    enc = LabelEncoder()
    enc.fit([-1, 0, 1])
    y_train_enc = enc.transform(y_train)
    y_test_enc = enc.transform(y_test)

    split_dt = str(df.index[split_idx])

    sp = {
        "df": df, "X_train": X_train, "X_test": X_test,
        "y_train": y_train, "y_test": y_test,
        "y_train_enc": y_train_enc, "y_test_enc": y_test_enc,
        "enc": enc,
        "close": close, "close_train": close_train, "close_test": close_test,
        "split_idx": split_idx, "split_dt": split_dt,
        "n_train": len(X_train), "n_test": len(X_test),
    }

    # 6. Build model from config
    model = _build_model_from_config(config, X_train, y_train_enc)

    # 7. Generate signals
    threshold = config.get("signal_threshold", 0.55)
    signal_train, p_pos_train, p_neg_train = _generate_signals(model, X_train, threshold)
    signal_test, p_pos_test, p_neg_test = _generate_signals(model, X_test, threshold)

    # 8. Apply filters (order: direction → session → ATR → trend)
    direction = config.get("direction", "both")
    signal_test = _apply_direction_filter(signal_test, direction)
    signal_train = _apply_direction_filter(signal_train, direction)

    session_filter = config.get("session_filter")
    signal_test = _apply_session_filter(signal_test, signal_test.index, session_filter)
    signal_train = _apply_session_filter(signal_train, signal_train.index, session_filter)

    min_atr = config.get("min_atr")
    if min_atr is not None:
        signal_test = _apply_atr_filter(signal_test, close_test, high_test, low_test, min_atr)

    trend_filter = config.get("trend_filter")
    if trend_filter is not None:
        signal_test = _apply_trend_filter(signal_test, close_test, trend_filter)

    signal_full = pd.concat([signal_train, signal_test])

    # 9. Backtest with SL/TP/cooldown (test + train)
    high_train = high.iloc[:train_end]
    low_train = low.iloc[:train_end]
    has_risk = (config.get("stop_loss") is not None or
                config.get("take_profit") is not None or
                config.get("cooldown", 0) > 0 or
                config.get("on_opposite", "reverse") != "reverse")

    # Bid/ask spread (#283): caller value wins, else a per-symbol default so
    # every training applies an honest spread (FX majors ~0.15 pip; crypto 0.0).
    _spread = float(spread_pips) if spread_pips is not None else _default_spread_pips(_sym)
    _costs = dict(commission_usd=commission_usd, swap_long_usd=swap_long_usd,
                  swap_short_usd=swap_short_usd, spread_pips=_spread)
    if has_risk:
        bt = run_backtest_v2(signal_test, close_test, high_test, low_test, config, capital=_cap, leverage=leverage, lot_size=lot_size, balance=_bal, symbol=_sym, **_costs)
        bt_train = run_backtest_v2(signal_train, close_train, high_train, low_train, config, capital=_cap, leverage=leverage, lot_size=lot_size, balance=_bal, symbol=_sym, **_costs)
    else:
        bt = run_backtest(signal_test, close_test, capital=_cap, leverage=leverage, lot_size=lot_size, balance=_bal, symbol=_sym, **_costs)
        bt_train = run_backtest(signal_train, close_train, capital=_cap, leverage=leverage, lot_size=lot_size, balance=_bal, symbol=_sym, **_costs)

    # 10. Metrics
    metrics = compute_metrics(bt, close_test, capital=_cap)

    # 11. Pre-compute all trade stats (single source of truth)
    pre_stats = {
        "train_stats": compute_trade_stats(bt_train.get("trade_returns", []), capital=_cap),
        "test_stats":  compute_trade_stats(bt.get("trade_returns", []), capital=_cap),
        "long_stats":  compute_trade_stats(bt.get("long_returns", []), capital=_cap),
        "short_stats": compute_trade_stats(bt.get("short_returns", []), capital=_cap),
    }

    # 12. Register model
    if register_model_fn is not None:
        register_model_fn(model)

    # 13. Build return dict
    _results = build_return_dict(sp, bt, metrics, model, feature_cols,
                                 signal_full, p_pos_test, p_neg_test, custom_figs=[],
                                 bt_train_result=bt_train, pre_stats=pre_stats)
    # #387/#388 - aux feed merge/drop report (incl. normalize warnings).
    _results["aux_feeds_report"] = _aux_report
    # Stash the leverage actually used so display/predictor/API can read it.
    _results["leverage"] = leverage
    # Broker-accounting params (Phase 1) — None lot_size = legacy path.
    if _broker:
        _results["lot_size"] = lot_size
        _results["balance"] = _bal
        _results["symbol"] = _sym
    # FACT9 Phase 2 — risk model (unit + levels) + a FIXED pip-equivalent for the
    # live predictor (Server B is HTTP-only → can't convert $ live; the backtest
    # already applied the exact per-trade $, live uses this constant pip stop).
    _ru2 = (config.get("risk_unit") or "pct").lower()
    _results["risk_unit"] = _ru2
    _results["stop_loss"] = config.get("stop_loss")
    _results["take_profit"] = config.get("take_profit")
    if _ru2 in ("pips", "usd"):
        try:
            _refpx = float(close.iloc[-1])
            _ps2 = _pip_size(_sym)
            _u2 = (lot_size or 1.0) * LOT_UNITS
            def _lvl_to_pips(_lv):
                if _lv is None:
                    return None
                if _ru2 == "pips":
                    return float(_lv)
                return float(_lv) * _usd_conv(_sym, _refpx) / (_ps2 * _u2)  # usd→pips
            _results["sl_pips"] = _lvl_to_pips(config.get("stop_loss"))
            _results["tp_pips"] = _lvl_to_pips(config.get("take_profit"))
        except Exception:
            pass
    return _results

# ── End strategy_utils ──


DATA_PATH = '/root/Desktop/QuantifyMe/data/ohlc/EURUSD_15min.parquet'
START_DATE = '2025-07-31'
END_DATE = '2026-02-03'
VALIDATION_DATE = ""
TRAIN_SPLIT = 0.6999526739233317
LEVERAGE = 30.0
LOTS = 1.0
BALANCE = 10000.0
RISK_UNIT = 'pips'
STOP_LOSS = 25.0
TAKE_PROFIT = 50.0


def feature_engineering(df, close, open_, high, low):
    """
    Feature Engineering: Standard momentum & volatility indicators
    - Multi-horizon returns: 1, 3, 5, 10, 20 bars
    - RSI 14: momentum oscillator
    - Bollinger Bands 20-period, 2 std: volatility & mean reversion
    - Moving averages 50, 100, 200: trend structure
    
    No lookahead bias. NaN filled via bfill/ffill at end.
    """
    
    # ============ Multi-horizon Returns ============
    df['ret_1'] = close.pct_change(1)
    df['ret_3'] = close.pct_change(3)
    df['ret_5'] = close.pct_change(5)
    df['ret_10'] = close.pct_change(10)
    df['ret_20'] = close.pct_change(20)
    
    # ============ RSI 14 ============
    delta = close.diff()
    gain = np.where(delta > 0, delta, 0)
    loss = np.where(delta < 0, -delta, 0)
    
    gain_ema = pd.Series(gain, index=close.index).ewm(span=14, adjust=False).mean()
    loss_ema = pd.Series(loss, index=close.index).ewm(span=14, adjust=False).mean()
    
    rs = gain_ema / (loss_ema + 1e-10)
    df['rsi_14'] = 100 - (100 / (1 + rs))
    
    # ============ Bollinger Bands 20, 2 ============
    sma_20 = close.rolling(20).mean()
    std_20 = close.rolling(20).std()
    
    df['bb_upper'] = sma_20 + 2 * std_20
    df['bb_lower'] = sma_20 - 2 * std_20
    df['bb_mid'] = sma_20
    df['bb_width'] = df['bb_upper'] - df['bb_lower']
    df['bb_pct'] = (close - df['bb_lower']) / (df['bb_width'] + 1e-10)
    
    # ============ Moving Averages (Trend Filter) ============
    df['sma_50'] = close.rolling(50).mean()
    df['sma_100'] = close.rolling(100).mean()
    df['sma_200'] = close.rolling(200).mean()
    
    # ============ ATR for Volatility (used in framework filters) ============
    tr = np.maximum.reduce([
        high - low,
        np.abs(high - close.shift(1)),
        np.abs(low - close.shift(1))
    ])
    df['atr_14'] = pd.Series(tr, index=close.index).rolling(14).mean()
    df['natr_14'] = df['atr_14'] / close * 100
    
    # ============ Fill NaN from indicator warm-up ============
    df = df.bfill().ffill()
    
    return df


def strategy_config():
    """
    XGBoost classifier optimized for Sharpe ratio on EUR/USD 15-min data.
    
    Strategy:
    - Model: XGBClassifier (binary classification: up/down next 4 bars)
    - Signal: confidence (predict_proba) > 0.55
    - Direction: both (long and short)
    - Position: 1 concurrent trade, close_only on opposite signal
    - Risk: no hard SL/TP at config level; framework applies cooldown & position mgmt
    
    Hyperparameters tuned for balance between:
      * Sharpe ratio (high-quality signals, low noise)
      * Drawdown control (regularization + shallow trees)
      * Transaction cost tolerance (at 2e-5 round-trip)
    """
    
    return {
        "title": "EUR/USD XGBoost RSI + Bollinger Bands Scalper",
        
        "model_type": "XGBClassifier",
        
        "model_params": {
            "n_estimators": 300,
            "max_depth": 5,
            "learning_rate": 0.05,
            "subsample": 0.8,
            "colsample_bytree": 0.8,
            "gamma": 1.0,
            "min_child_weight": 2,
            "random_state": 42,
        },
        
        "signal_threshold": 0.55,
        
        "direction": "both",
        
        "stop_loss": None,
        "take_profit": None,
        
        "cooldown": 0,
        
        "max_positions": 1,
        
        "on_opposite": "close_only",
        
        "session_filter": None,
        
        "min_atr": None,
        
        "trend_filter": None,
        
        "target_horizon": 4,
        
        "objective": (
            "Maximize Sharpe ratio on EUR/USD 15-min data via XGBClassifier. "
            "Features: returns (1, 3, 5, 10, 20 bars), RSI 14, Bollinger Bands 20/2, "
            "SMAs (50, 100, 200), ATR 14. Model trained on 70% historical data, "
            "tested on 30% holdout (2025-07-31 to 2026-07-31)."
        ),
        
        "notes": (
            "Regularization (gamma=1.0, min_child_weight=2, max_depth=5) keeps tree complexity low "
            "to improve generalization and reduce overfitting. Learning rate 0.05 with 300 trees "
            "provides stable gradient descent. Subsample & colsample at 0.8 add robustness. "
            "Signal threshold 0.55 filters weak predictions. No hard SL/TP; framework handles "
            "risk via position management. Target horizon 4 bars (~1 hour on 15-min) balances "
            "scalper speed with signal clarity."
        ),
    }

# ---------------------------------------------------------------
# AUX_FEEDS - written by the strategy BUILDER, not by the model.
# Last module-level statement on purpose: module execution is
# last-write-wins, so this is the value run_strategy() sees no matter
# what appears above. Do not move it, and add nothing after it.
# ---------------------------------------------------------------
AUX_FEEDS = []

# ── Framework v2: auto-generated wrapper ──
def train_and_backtest():
    _vd = VALIDATION_DATE if 'VALIDATION_DATE' in globals() else ''
    _ts = TRAIN_SPLIT if 'TRAIN_SPLIT' in globals() else 0.7
    _aux = AUX_FEEDS if 'AUX_FEEDS' in globals() else None
    _lev = LEVERAGE if 'LEVERAGE' in globals() else 30.0
    _lots = LOTS if 'LOTS' in globals() else 1.0
    _bal = BALANCE if 'BALANCE' in globals() else 10000.0
    _ru = RISK_UNIT if 'RISK_UNIT' in globals() else None
    _sl = STOP_LOSS if 'STOP_LOSS' in globals() else None
    _tp = TAKE_PROFIT if 'TAKE_PROFIT' in globals() else None
    _sth = SIGNAL_THRESHOLD if 'SIGNAL_THRESHOLD' in globals() else None
    _comm = QM_COMMISSION_USD if 'QM_COMMISSION_USD' in globals() else None
    _swl = QM_SWAP_LONG_USD if 'QM_SWAP_LONG_USD' in globals() else None
    _sws = QM_SWAP_SHORT_USD if 'QM_SWAP_SHORT_USD' in globals() else None
    return run_strategy(
        feature_engineering, strategy_config,
        DATA_PATH, START_DATE, END_DATE,
        _vd, _ts,
        register_model_fn=register_model,
        leverage=_lev, lot_size=_lots, balance=_bal,
        stop_loss=_sl, take_profit=_tp, risk_unit=_ru,
        signal_threshold=_sth,
        commission_usd=_comm, swap_long_usd=_swl, swap_short_usd=_sws,
        aux_feeds=_aux
    )

# ╔══════════════════════════════════════════════════════════════╗
# ║                  STRATEGY REQUEST LOG                       ║
# ╚══════════════════════════════════════════════════════════════╝
# Generated        : 2026-05-25 02:27:36
# Model            : XGBoost
# Feature Eng.     : go long when EMA(9) crosses above EMA(21), exit when it crosses back below + Auto-add features: ON
# Signal / Entry   : —
# Optimization     : —
# Risk Mgmt        : —
# Risk Filter      : —
# ══════════════════════════════════════════════════════════════

# ============================================================
# SECTION 0 — IMPORTS & CONSTANTS
import numpy as np
import pandas as pd

# ── Inlined strategy_utils ──
"""
strategy_utils.py — Standard utility functions for generated strategies.
Claude imports these instead of writing boilerplate from scratch.
This ensures consistent behavior across all generated strategies.
"""
import numpy as np
import pandas as pd
from sklearn.preprocessing import LabelEncoder


# Max backtest window per timeframe. A finer timeframe over a longer window
# blows up the results dict / parquet load / Modal train time (the 2026-05-12
# OOM was a 1-min × multi-year sweep) — and a 1-min strategy gains nothing from
# 2 years of 1-min bars. Enforced HERE because every training path (UI / API /
# Modal) funnels through run_strategy → load_ohlc. Env-overridable so a future
# "max plan" / dedicated-server tier can lift it.
_TF_MAX_DAYS = {
    "1min":  30,
    "5min":  90,
    "15min": 365,
    "1h":    730,
}


def _fetch_ohlc_from_internal(symbol: str, tf: str, start: str, end: str):
    """Phase 3.2: fetch parquet bytes from Server A's /internal/ohlc endpoint
    instead of reading a local file. Used inside Modal containers / Mac worker
    pool (Phase 3.4) so every train sees the same source of truth as the chart.

    Returns: pd.DataFrame (parquet decoded), or raises on any failure so the
    caller can fall back / surface a clear error in the job.
    """
    import hashlib as _hashlib, hmac as _hmac, io as _io, os as _os
    import urllib.request as _ur, urllib.parse as _urp

    base = (_os.environ.get("QM_INTERNAL_OHLC_BASE") or "").rstrip("/")
    secret = (_os.environ.get("INTERNAL_WS_SECRET") or "").strip()
    if not base:
        raise RuntimeError("QM_INTERNAL_OHLC_BASE not set")
    if not secret:
        raise RuntimeError("INTERNAL_WS_SECRET not set")

    msg = f"{symbol}|{tf}|{start}|{end}".encode("utf-8")
    sig = _hmac.new(secret.encode("utf-8"), msg, _hashlib.sha256).hexdigest()

    qs = _urp.urlencode({
        "symbol": symbol, "tf": tf,
        "start": start, "end": end, "sig": sig,
    })
    url = f"{base}/internal/ohlc?{qs}"

    req = _ur.Request(url, headers={"User-Agent": "qm-worker/1.0"})
    with _ur.urlopen(req, timeout=30) as resp:
        if resp.status != 200:
            raise RuntimeError(f"/internal/ohlc returned {resp.status}")
        payload = resp.read()
    print(f"[load_ohlc:internal] {symbol} {tf} fetched {len(payload)} bytes", flush=True)
    return pd.read_parquet(_io.BytesIO(payload))


def _parse_symbol_tf_from_path(data_path: str):
    """Pull SYMBOL + TF out of a path like .../EURUSD_1min.parquet."""
    import os as _os, re as _re
    base = _os.path.basename(str(data_path))
    m = _re.match(r"^([A-Z]{6})_(\d+min|\d+h)\.parquet$", base)
    if not m:
        return None, None
    return m.group(1), m.group(2)


def load_ohlc(data_path, start_date="", end_date=""):
    """Load OHLC parquet, sort index, filter dates. Always returns consistent format.

    The lower bound is clamped per timeframe (see _TF_MAX_DAYS) — a request for
    more history than the cap silently starts later.

    Phase 3.2: when env QM_USE_INTERNAL_OHLC=="1", fetch over HTTP from
    Server A's /internal/ohlc endpoint instead of pd.read_parquet on a local
    file (which on Modal is a stale Volume snapshot). The endpoint applies the
    same day-cap, so the local cap-check below is a defensive no-op in that
    path. Flag defaults to "0" → unchanged behavior.

    Returns: (df, close, open_, high, low)
    """
    import os as _os, re as _re

    _use_internal = _os.environ.get("QM_USE_INTERNAL_OHLC", "0") == "1"
    if _use_internal:
        _sym, _tf = _parse_symbol_tf_from_path(data_path)
        if not _sym or not _tf:
            raise RuntimeError(
                f"QM_USE_INTERNAL_OHLC=1 but DATA_PATH basename does not match "
                f"SYMBOL_TF.parquet: {data_path}"
            )
        df = _fetch_ohlc_from_internal(_sym, _tf, start_date or "", end_date or "")
    else:
        df = pd.read_parquet(data_path)
    df.index = pd.to_datetime(df.index)
    df = df.sort_index()

    # Per-timeframe window cap (timeframe inferred from the parquet filename).
    _m = _re.search(r"_(\d+min|\d+h)\.parquet$", _os.path.basename(str(data_path)))
    _tf = _m.group(1) if _m else None
    _max_days = _TF_MAX_DAYS.get(_tf)
    if _max_days and _max_days > 0 and len(df):
        _env_override = _os.environ.get(f"QM_MAX_DAYS_{_tf.upper()}")
        if _env_override and _env_override.isdigit():
            _max_days = int(_env_override)
        try:
            _eff_end = pd.Timestamp(end_date) if end_date else df.index.max()
            _eff_end = min(_eff_end, df.index.max())
            _floor = _eff_end - pd.Timedelta(days=_max_days)
            _req_start = pd.Timestamp(start_date) if start_date else df.index.min()
            if _req_start < _floor:
                print(f"[load_ohlc] {_tf} backtest window capped to {_max_days}d: "
                      f"start {_req_start.date()} -> {_floor.date()}", flush=True)
                start_date = _floor
        except Exception as _e:
            print(f"[load_ohlc] window-cap check skipped ({_e})", flush=True)

    if start_date:
        df = df[df.index >= start_date]
    if end_date:
        df = df[df.index <= end_date]

    return df, df["close"], df["open"], df["high"], df["low"]


def make_target(close, horizon=4):
    """Create target: direction N bars ahead. Default 4 bars = 1 hour on 15-min data.

    Returns: target (pd.Series of -1, 0, 1)
    """
    return np.sign(close.shift(-horizon) - close)


def split_data(df, target, feature_cols, train_split=0.7, validation_date=""):
    """Train/test split. Handles both ratio and date-based splits.
    Drops NaN from target before splitting. Encodes labels to [0,1,2].

    Returns: dict with keys:
        X_train, X_test, y_train, y_test,
        y_train_enc, y_test_enc, enc,
        close_train, close_test,
        split_idx, split_dt, n_train, n_test
    """
    # Drop NaN from target
    mask = target.notna()
    df = df[mask].copy()
    target = target[mask]

    close = df["close"]

    # Build feature matrix
    X = df[feature_cols].copy()
    X = X.bfill().ffill()
    X = X.replace([np.inf, -np.inf], np.nan).fillna(0.0)

    # Split
    if validation_date:
        split_idx = len(df[df.index <= validation_date])
    else:
        split_idx = int(len(df) * train_split)

    split_idx = max(1, min(split_idx, len(df) - 1))

    X_train = X.iloc[:split_idx]
    X_test = X.iloc[split_idx:]
    y_train = target.iloc[:split_idx]
    y_test = target.iloc[split_idx:]
    close_train = close.iloc[:split_idx]
    close_test = close.iloc[split_idx:]

    split_dt = str(df.index[split_idx])

    # Label encoding — always fit on [-1, 0, 1]
    enc = LabelEncoder()
    enc.fit([-1, 0, 1])
    y_train_enc = enc.transform(y_train)
    y_test_enc = enc.transform(y_test)

    return {
        "df": df, "X_train": X_train, "X_test": X_test,
        "y_train": y_train, "y_test": y_test,
        "y_train_enc": y_train_enc, "y_test_enc": y_test_enc,
        "enc": enc,
        "close": close, "close_train": close_train, "close_test": close_test,
        "split_idx": split_idx, "split_dt": split_dt,
        "n_train": len(X_train), "n_test": len(X_test),
    }


def compute_overlays(close, df_index):
    """Compute BB and MA overlays on full dataset. Always consistent.

    Returns: (bb_dict, ma_dict)
    """
    bb_mid = close.rolling(20).mean()
    bb_std = close.rolling(20).std()
    bb_upper = bb_mid + 2 * bb_std
    bb_lower = bb_mid - 2 * bb_std

    ma50 = close.rolling(50).mean()
    ma100 = close.rolling(100).mean()
    ma200 = close.rolling(200).mean()

    def _safe(s):
        s = s.reindex(df_index).bfill().ffill()
        return [float(x) if (x is not None and not np.isnan(x) and not np.isinf(x)) else None
                for x in s.values]

    bb = {"upper": _safe(bb_upper), "mid": _safe(bb_mid), "lower": _safe(bb_lower)}
    ma = {"ma50": _safe(ma50), "ma100": _safe(ma100), "ma200": _safe(ma200)}

    return bb, ma


def run_backtest(signal, close, capital=10000, cost=2e-5):
    """Run backtest with transaction costs.
    Uses price-based trade returns (same as webapp _compute_trades).
    Signal 0 = hold (keep current position), not close.

    Returns: dict with equity, trade_returns, long_returns, short_returns, bar_returns
    """
    sig_arr = signal.values
    price_arr = close.values
    idx = signal.index
    n = len(price_arr)

    # Trade returns — price-based (matches webapp _compute_trades exactly)
    trade_returns = []
    long_returns = []
    short_returns = []
    trade_log = []
    last_dir = None
    entry_price = None
    entry_bar = None

    for i in range(n):
        s = sig_arr[i]
        c = price_arr[i]
        if s != 0.0 and s != last_dir:
            # Direction change — close previous trade, open new
            if last_dir is not None and entry_price is not None and entry_price != 0:
                ret = float(last_dir * (c - entry_price) / entry_price - cost)
                trade_returns.append(ret)
                if last_dir == 1:
                    long_returns.append(ret)
                else:
                    short_returns.append(ret)
                trade_log.append({
                    "type": "Buy" if last_dir == 1 else "Sell",
                    "entry_time": str(idx[entry_bar]),
                    "exit_time": str(idx[i]),
                    "entry_price": round(entry_price, 5),
                    "exit_price": round(c, 5),
                    "pnl": round(last_dir * (c - entry_price), 5),
                    "pnl_pct": round(ret * 100, 3),
                    "exit_reason": "signal",
                })
            entry_price = c
            entry_bar = i
            last_dir = s

    # Close last open trade
    if last_dir is not None and entry_price is not None and n > 0 and entry_price != 0:
        c = price_arr[-1]
        ret = float(last_dir * (c - entry_price) / entry_price - cost)
        trade_returns.append(ret)
        if last_dir == 1:
            long_returns.append(ret)
        else:
            short_returns.append(ret)
        trade_log.append({
            "type": "Buy" if last_dir == 1 else "Sell",
            "entry_time": str(idx[entry_bar]),
            "exit_time": str(idx[-1]),
            "entry_price": round(entry_price, 5),
            "exit_price": round(c, 5),
            "pnl": round(last_dir * (c - entry_price), 5),
            "pnl_pct": round(ret * 100, 3),
            "exit_reason": "end",
        })

    # Equity curve from trade returns
    cumret = 1.0
    equity_vals = np.full(n, float(capital))
    trade_idx = 0
    in_trade = False
    t_entry_price = None
    t_dir = None

    for i in range(n):
        s = sig_arr[i]
        c = price_arr[i]
        if s != 0.0 and s != t_dir:
            if t_dir is not None and t_entry_price is not None and t_entry_price != 0:
                t_ret = t_dir * (c - t_entry_price) / t_entry_price - cost
                cumret *= (1 + t_ret)
            t_entry_price = c
            t_dir = s
        equity_vals[i] = capital * cumret

    # Bar returns for Sharpe
    bar_returns = np.zeros(n)
    for i in range(1, n):
        if price_arr[i - 1] != 0 and last_dir is not None:
            bar_returns[i] = sig_arr[i - 1] * (price_arr[i] - price_arr[i - 1]) / price_arr[i - 1] if sig_arr[i - 1] != 0 else 0.0

    return {
        "equity": pd.Series(equity_vals, index=close.index),
        "trade_returns": trade_returns,
        "long_returns": long_returns,
        "short_returns": short_returns,
        "bar_returns": bar_returns,
        "trade_log": trade_log,
    }


def compute_trade_stats(trades, capital=10000):
    """Single source of truth for trade statistics.
    Every display path reads from this — no recomputation anywhere.
    All values are rounded and JSON-safe (no inf/nan).
    """
    if not trades:
        return {"n": 0, "wins": 0, "losses": 0, "wr": 0, "avg": 0,
                "best": 0, "worst": 0, "ret": 0, "np": 0, "mdd": 0,
                "pf": 0, "rr": 0, "expect": 0}
    w = [r for r in trades if r > 0]
    l = [r for r in trades if r < 0]
    cumret = 1.0
    for r in trades:
        cumret *= (1 + r)
    net_p = capital * (cumret - 1)
    # Max drawdown
    eq = np.cumprod([1.0] + [1 + r for r in trades])
    peak = np.maximum.accumulate(eq)
    mdd = float(((eq - peak) / peak).min()) if len(eq) > 1 else 0.0
    # Profit Factor
    gross_w = sum(w) if w else 0
    gross_l = abs(sum(l)) if l else 0
    pf = gross_w / gross_l if gross_l > 0 else (9999.0 if gross_w > 0 else 0)
    # Risk:Reward
    avg_w = float(np.mean(w)) if w else 0
    avg_l = abs(float(np.mean(l))) if l else 0
    rr = avg_w / avg_l if avg_l > 0 else (9999.0 if avg_w > 0 else 0)
    # Expectancy
    expect = net_p / len(trades)
    return {
        "n": len(trades), "wins": len(w), "losses": len(l),
        "wr": round(len(w) / len(trades), 4),
        "avg": round(float(np.mean(trades)), 6),
        "best": round(max(w), 6) if w else 0,
        "worst": round(min(l), 6) if l else 0,
        "ret": round(cumret - 1, 6),
        "np": round(net_p, 2),
        "mdd": round(mdd, 6),
        "pf": round(pf, 2),
        "rr": round(rr, 2),
        "expect": round(expect, 2),
    }


def compute_metrics(bt_result, close_test, capital=10000):
    """Compute all standard metrics from backtest result.
    Uses trade-level compounding (same as webapp _trade_stats) for accuracy.

    Returns: dict with total_ret, bh_ret, sharpe_strat, sharpe_bh, mdd, n_trades
    """
    equity = bt_result["equity"]
    trade_returns = bt_result["trade_returns"]

    # Total return — trade-level compounding (matches webapp)
    if trade_returns:
        cumret = 1.0
        for r in trade_returns:
            cumret *= (1 + r)
        total_ret = cumret - 1
    else:
        total_ret = 0.0

    # Buy and hold
    bh_equity = capital * (close_test / close_test.iloc[0])
    bh_ret = (bh_equity.iloc[-1] - capital) / capital if capital != 0 else 0.0

    # Sharpe ratio — trade-level (matches webapp: sqrt(252*26) annualization)
    if len(trade_returns) >= 2 and float(np.std(trade_returns)) > 0:
        sharpe_strat = float(np.mean(trade_returns) / np.std(trade_returns) * np.sqrt(252 * 26))
    else:
        sharpe_strat = 0.0

    bh_rets = bh_equity.pct_change().dropna()
    if len(bh_rets) > 1 and bh_rets.std() != 0:
        sharpe_bh = float((bh_rets.mean() / bh_rets.std()) * np.sqrt(252 * 24 * 4))
    else:
        sharpe_bh = 0.0

    # Max drawdown — trade-level (matches webapp)
    if trade_returns:
        eq = np.cumprod([1.0] + [1 + r for r in trade_returns])
        peak = np.maximum.accumulate(eq)
        mdd = float(((eq - peak) / peak).min()) if len(eq) > 1 else 0.0
    else:
        mdd = 0.0

    return {
        "total_ret": float(total_ret),
        "bh_ret": float(bh_ret),
        "sharpe_strat": float(sharpe_strat) if not np.isnan(sharpe_strat) else 0.0,
        "sharpe_bh": float(sharpe_bh) if not np.isnan(sharpe_bh) else 0.0,
        "mdd": float(mdd),
        "n_trades": len(trade_returns),
    }


# Diagnostics line/histogram series (equity / drawdown / rolling_acc / conf_hist)
# only feed the small Diagnostics charts — they're never used by the price chart
# or scroll-back. On a 1-min model trained over the (2.2-capped) window these are
# still ~30k points each; downsample to a visually-identical resolution before the
# dict leaves the trainer so it doesn't carry that into Server-A RAM / Postgres.
_RESULTS_SERIES_MAX = 5000


def _downsample_idx(n, cap=_RESULTS_SERIES_MAX):
    """Evenly-spaced index list spanning [0, n-1] (first+last always kept), or
    None when no downsampling is needed (n <= cap)."""
    if n <= cap:
        return None
    return np.unique(np.linspace(0, n - 1, cap).astype(int)).tolist()


def _take(arr, idx):
    """Subset a list by an index list (idx may be None → return arr unchanged)."""
    if idx is None or not isinstance(arr, list):
        return arr
    return [arr[i] for i in idx]


# trade_log / train_trade_log are lists of per-trade dicts (display-only — the
# Trade Log tab). They scale with TRADE count, not bar count, so the bar-window
# cap (Phase 2.2) doesn't bound them — a degenerate near-every-bar model can put
# 10k+ trade dicts in the blob (>3 MB). Cap each (independently — a small-N model
# keeps every trade) to the most-recent N, recording `*_total` + `*_truncated`
# so the true count is still reported. Real strategies have far fewer than
# _TRADE_LOG_MAX trades, so this only ever bites pathological models.
_TRADE_LOG_MAX = 5000


def _cap_trade_log(tl):
    """Return (capped_list, original_len, was_truncated)."""
    if not isinstance(tl, list) or len(tl) <= _TRADE_LOG_MAX:
        return tl, (len(tl) if isinstance(tl, list) else 0), False
    return tl[-_TRADE_LOG_MAX:], len(tl), True


def build_return_dict(split_result, bt_result, metrics, model, feature_cols,
                      signal_full, p_pos_test, p_neg_test, custom_figs=None,
                      bt_train_result=None, pre_stats=None):
    """Assemble the complete return dict. Handles ALL serialization.
    Never returns Timestamps, numpy arrays, or non-JSON types.

    Returns: JSON-safe dict with all required keys
    """
    df = split_result["df"]
    close = split_result["close"]
    close_test = split_result["close_test"]
    X_test = split_result["X_test"]
    y_test = split_result["y_test"]
    equity = bt_result["equity"]
    bar_returns = bt_result["bar_returns"]

    # OHLC
    ohlc_dates = [str(x) for x in df.index.tolist()]

    def _safe_list(arr):
        return [float(x) if (x is not None and not np.isnan(x) and not np.isinf(x)) else None
                for x in arr]

    # Overlays
    bb, ma = compute_overlays(close, df.index)

    # Buy and hold equity
    capital = equity.iloc[0] if len(equity) > 0 else 10000
    bh_equity = capital * (close_test / close_test.iloc[0])

    # Confusion matrix
    from sklearn.metrics import confusion_matrix
    pred_test = model.predict(X_test)
    y_test_arr = np.asarray(y_test)
    cm = confusion_matrix(y_test_arr, pred_test, labels=[-1, 0, 1])

    # Rolling accuracy
    sig_arr = signal_full.reindex(close_test.index).values
    correct = pd.Series((pred_test == y_test_arr).astype(float), index=X_test.index)
    active_test = pd.Series(sig_arr != 0, index=close_test.index) if len(sig_arr) == len(close_test) else pd.Series(True, index=close_test.index)
    correct_active = correct.where(active_test, other=np.nan)
    rolling_acc = correct_active.rolling(30, min_periods=1).mean()

    # Feature importance
    importances = model.feature_importances_
    fi_pairs = sorted(zip(feature_cols, importances), key=lambda x: x[1])[-15:]

    # Drawdown
    rolling_max = equity.cummax()
    drawdown = (equity - rolling_max) / rolling_max.replace(0, np.nan)
    drawdown = drawdown.fillna(0.0)

    # ── Downsample the Diagnostics-only series (see _downsample_idx) ──────────
    _eq_dates = [str(x) for x in close_test.index.tolist()]
    _eq_strat = _safe_list(equity.values)
    _eq_bh    = _safe_list(bh_equity.values)
    _eq_idx   = _downsample_idx(len(_eq_dates))
    _eq_dates, _eq_strat, _eq_bh = _take(_eq_dates, _eq_idx), _take(_eq_strat, _eq_idx), _take(_eq_bh, _eq_idx)

    _ra_dates = [str(x) for x in rolling_acc.index.tolist()]
    _ra_vals  = [float(x) if (not np.isnan(x) and not np.isinf(x)) else None for x in rolling_acc.values]
    _ra_idx   = _downsample_idx(len(_ra_dates))
    _ra_dates, _ra_vals = _take(_ra_dates, _ra_idx), _take(_ra_vals, _ra_idx)

    _dd_dates = [str(x) for x in drawdown.index.tolist()]
    _dd_vals  = _safe_list(drawdown.values)
    _dd_idx   = _downsample_idx(len(_dd_dates))
    _dd_dates, _dd_vals = _take(_dd_dates, _dd_idx), _take(_dd_vals, _dd_idx)

    _cp_pos = [float(x) for x in (p_pos_test.tolist() if hasattr(p_pos_test, 'tolist') else list(p_pos_test))]
    _cp_neg = [float(x) for x in (p_neg_test.tolist() if hasattr(p_neg_test, 'tolist') else list(p_neg_test))]
    _cp_pos = _take(_cp_pos, _downsample_idx(len(_cp_pos)))
    _cp_neg = _take(_cp_neg, _downsample_idx(len(_cp_neg)))

    # ── Trade logs — display-only (Trade Log tab); cap to most-recent N with a
    #    `_total` field so the true count is still reported (see _cap_trade_log).
    #    NB: ret_dist arrays are left FULL — a downstream path in callbacks.py
    #    recomputes n_trades/win-rate from len(ret_dist), so a sample would skew
    #    the displayed counts; they're small anyway and gzip handles them.
    _tl_test, _tl_test_n, _tl_test_tr = _cap_trade_log(bt_result.get("trade_log", []))
    _tl_tr,   _tl_tr_n,   _tl_tr_tr   = _cap_trade_log(bt_train_result.get("trade_log", []) if bt_train_result else [])

    return {
        "ohlc": {
            "dates": ohlc_dates,
            "open": _safe_list(df["open"].values),
            "high": _safe_list(df["high"].values),
            "low": _safe_list(df["low"].values),
            "close": _safe_list(df["close"].values),
        },
        "signals": {
            "dates": [str(x) for x in signal_full.index.tolist()],
            "values": [float(x) for x in signal_full.values],
        },
        "bb": bb,
        "ma": ma,
        "equity": {
            "dates": _eq_dates,
            "strategy": _eq_strat,
            "bh": _eq_bh,
        },
        "feature_importance": {
            "names": [p[0] for p in fi_pairs],
            "values": [float(p[1]) for p in fi_pairs],
        },
        "conf_matrix": cm.tolist(),
        "conf_hist": {
            "p_pos": _cp_pos,
            "p_neg": _cp_neg,
        },
        "rolling_acc": {
            "dates": _ra_dates,
            "values": _ra_vals,
        },
        "drawdown": {
            "dates": _dd_dates,
            "values": _dd_vals,
        },
        "ret_dist": [float(x) for x in bt_result["trade_returns"]],
        "ret_dist_long": [float(x) for x in bt_result["long_returns"]],
        "ret_dist_short": [float(x) for x in bt_result["short_returns"]],
        "train_ret_dist": [float(x) for x in bt_train_result["trade_returns"]] if bt_train_result else [],
        "train_ret_dist_long": [float(x) for x in bt_train_result["long_returns"]] if bt_train_result else [],
        "train_ret_dist_short": [float(x) for x in bt_train_result["short_returns"]] if bt_train_result else [],
        "trade_log": _tl_test,
        "train_trade_log": _tl_tr,
        "trade_log_total": _tl_test_n,
        "train_trade_log_total": _tl_tr_n,
        "trade_log_truncated": _tl_test_tr,
        "train_trade_log_truncated": _tl_tr_tr,
        **(pre_stats or {}),
        "metrics": metrics,
        "split_dt": split_result["split_dt"],
        "split_idx": int(split_result["split_idx"]),
        "n_train": int(split_result["n_train"]),
        "n_test": int(split_result["n_test"]),
        "feature_cols": list(feature_cols),
        "custom_figs": custom_figs or [],
    }


# ════════════════════════════════════════════════════════════════════════════
# STRATEGY FRAMEWORK v2 — Config-driven architecture
# Claude writes feature_engineering() + strategy_config(). Framework does rest.
# ════════════════════════════════════════════════════════════════════════════

import importlib

_MODEL_REGISTRY = {
    "XGBClassifier":              ("xgboost", "XGBClassifier"),
    "RandomForestClassifier":     ("sklearn.ensemble", "RandomForestClassifier"),
    "GradientBoostingClassifier": ("sklearn.ensemble", "GradientBoostingClassifier"),
    "LogisticRegression":         ("sklearn.linear_model", "LogisticRegression"),
    "ExtraTreesClassifier":       ("sklearn.ensemble", "ExtraTreesClassifier"),
    "AdaBoostClassifier":         ("sklearn.ensemble", "AdaBoostClassifier"),
}


def _build_model_from_config(config, X_train, y_train_enc):
    """Build, fit, and wrap a model from strategy_config dict."""
    model_type = config.get("model_type", "RandomForestClassifier")
    model_params = dict(config.get("model_params", {}))

    if model_type not in _MODEL_REGISTRY:
        raise ValueError(f"Unknown model_type '{model_type}'. Valid: {list(_MODEL_REGISTRY.keys())}")

    module_path, class_name = _MODEL_REGISTRY[model_type]
    mod = importlib.import_module(module_path)
    cls = getattr(mod, class_name)

    # XGBoost defaults
    if class_name == "XGBClassifier":
        model_params.setdefault("use_label_encoder", False)
        model_params.setdefault("eval_metric", "mlogloss")
        model_params.setdefault("tree_method", "hist")
        # Determinism > speed (2026-05-25). XGBoost hist with n_jobs=-1 is
        # NON-reproducible even with random_state set — the parallel histogram
        # gradient-sum order varies across threads, so the SAME code + data
        # gives a slightly different model (and backtest) every run. Forcing
        # single-thread makes training bit-reproducible so: (a) a user who
        # copies a strategy and reruns it gets identical numbers, (b) the
        # community "Live" score matches a redeploy, (c) "same code, different
        # result" support reports go away. Cost: single-threaded XGB (a few
        # seconds slower on large windows; hist is fast so it's minor). FORCED
        # (not setdefault) so the guarantee can't be silently broken by a
        # strategy passing n_jobs. Exact reproducibility holds within the
        # platform (pinned versions / same Modal image); a user's own machine
        # with different xgboost/numpy/CPU can still differ in low-order bits.
        model_params["n_jobs"] = 1

    # Common defaults
    model_params.setdefault("random_state", 42)

    from model_wrapper import ModelWrapper

    clf = cls(**model_params)
    clf.fit(X_train, y_train_enc)

    enc = LabelEncoder()
    enc.fit([-1, 0, 1])
    return ModelWrapper(clf, original_classes=enc.classes_, n_features=X_train.shape[1])


def _generate_signals(model, X, threshold):
    """Framework-owned signal generation. Deterministic threshold logic."""
    proba = model.predict_proba(X)
    classes = list(model.classes_)

    idx_pos = classes.index(1) if 1 in classes else None
    idx_neg = classes.index(-1) if -1 in classes else None

    p_pos = proba[:, idx_pos] if idx_pos is not None else np.zeros(len(X))
    p_neg = proba[:, idx_neg] if idx_neg is not None else np.zeros(len(X))

    signal_vals = np.zeros(len(X))
    signal_vals = np.where(p_pos >= threshold, 1.0, signal_vals)
    signal_vals = np.where(p_neg >= threshold, -1.0, signal_vals)
    # Both exceed: pick stronger
    both = (p_pos >= threshold) & (p_neg >= threshold)
    signal_vals[both] = np.where(p_pos[both] >= p_neg[both], 1.0, -1.0)

    return pd.Series(signal_vals, index=X.index), p_pos, p_neg


# ── Filter functions (all no-ops when config value is None) ──────────────

def _apply_direction_filter(signal, direction):
    """Zero out signals that don't match allowed direction."""
    if direction is None or direction == "both":
        return signal
    s = signal.copy()
    if direction == "long":
        s[s < 0] = 0.0
    elif direction == "short":
        s[s > 0] = 0.0
    return s


def _apply_session_filter(signal, index, session_hours):
    """Zero out signals outside session hours [start, end] UTC."""
    if session_hours is None:
        return signal
    s = signal.copy()
    start_h, end_h = session_hours[0], session_hours[1]
    hours = index.hour
    if start_h <= end_h:
        mask = (hours >= start_h) & (hours < end_h)
    else:  # wrap around midnight, e.g. [22, 6]
        mask = (hours >= start_h) | (hours < end_h)
    s[~mask] = 0.0
    return s


def _apply_atr_filter(signal, close, high, low, min_atr):
    """Zero out signals when NATR(14) is below threshold."""
    if min_atr is None:
        return signal
    hl = high - low
    hc = (high - close.shift(1)).abs()
    lc = (low - close.shift(1)).abs()
    tr = pd.concat([hl, hc, lc], axis=1).max(axis=1)
    atr14 = tr.ewm(com=13, adjust=False).mean()
    natr = atr14 / close.replace(0, np.nan)
    s = signal.copy()
    s[natr < min_atr] = 0.0
    return s


def _apply_trend_filter(signal, close, trend_filter):
    """Only allow signals aligned with trend. e.g. 'sma_50': longs above SMA, shorts below."""
    if trend_filter is None:
        return signal
    # Parse: "sma_50" → SMA with period 50
    parts = trend_filter.lower().replace("-", "_").split("_")
    if len(parts) >= 2 and parts[0] in ("sma", "ema"):
        period = int(parts[1])
    else:
        return signal  # unknown filter, skip
    if parts[0] == "sma":
        trend_line = close.rolling(period).mean()
    else:
        trend_line = close.ewm(span=period, adjust=False).mean()
    s = signal.copy()
    # Longs only above trend, shorts only below
    s[(s > 0) & (close < trend_line)] = 0.0
    s[(s < 0) & (close > trend_line)] = 0.0
    return s


# ── run_backtest_v2: framework-owned SL/TP/cooldown/position management ──

def run_backtest_v2(signal, close, high, low, config, capital=10000, cost=2e-5):
    """Backtest with SL/TP/cooldown/direction handling built into the engine.
    Unlike run_backtest (v1), this function handles position exits internally.

    Returns: same dict shape as run_backtest()
    """
    stop_loss = config.get("stop_loss")
    take_profit = config.get("take_profit")
    cooldown = config.get("cooldown", 0)
    on_opposite = config.get("on_opposite", "reverse")

    sig_arr = signal.values
    close_arr = close.values
    high_arr = high.values
    low_arr = low.values
    idx = signal.index
    n = len(close_arr)

    trade_returns = []
    long_returns = []
    short_returns = []
    trade_log = []
    equity_vals = np.full(n, float(capital))
    cumret = 1.0

    position = 0.0       # current direction: 1.0, -1.0, or 0.0 (flat)
    entry_price = None
    entry_bar = None      # index into arrays for entry time
    cooldown_remaining = 0

    def _log_trade(exit_bar, exit_px, ret, reason):
        trade_log.append({
            "type": "Buy" if position == 1.0 else "Sell",
            "entry_time": str(idx[entry_bar]),
            "exit_time": str(idx[exit_bar]),
            "entry_price": round(entry_price, 5),
            "exit_price": round(exit_px, 5),
            "pnl": round(position * (exit_px - entry_price), 5),
            "pnl_pct": round(ret * 100, 3),
            "exit_reason": reason,
        })

    for i in range(n):
        c = close_arr[i]
        h = high_arr[i]
        lo = low_arr[i]
        s = sig_arr[i]

        # 1. Check SL/TP if in trade
        if position != 0.0 and entry_price is not None:
            hit_sl = False
            hit_tp = False
            exit_price = None

            if position == 1.0:  # long
                if stop_loss is not None and lo <= entry_price * (1 - stop_loss):
                    hit_sl = True
                    exit_price = entry_price * (1 - stop_loss)
                elif take_profit is not None and h >= entry_price * (1 + take_profit):
                    hit_tp = True
                    exit_price = entry_price * (1 + take_profit)
            else:  # short
                if stop_loss is not None and h >= entry_price * (1 + stop_loss):
                    hit_sl = True
                    exit_price = entry_price * (1 + stop_loss)
                elif take_profit is not None and lo <= entry_price * (1 - take_profit):
                    hit_tp = True
                    exit_price = entry_price * (1 - take_profit)

            if hit_sl or hit_tp:
                ret = float(position * (exit_price - entry_price) / entry_price - cost)
                trade_returns.append(ret)
                if position == 1.0:
                    long_returns.append(ret)
                else:
                    short_returns.append(ret)
                _log_trade(i, exit_price, ret, "SL" if hit_sl else "TP")
                cumret *= (1 + ret)
                position = 0.0
                entry_price = None
                entry_bar = None
                cooldown_remaining = cooldown
                equity_vals[i] = capital * cumret
                continue

        # 2. Cooldown
        if cooldown_remaining > 0:
            cooldown_remaining -= 1
            equity_vals[i] = capital * cumret
            continue

        # 3. Signal processing
        if s != 0.0:
            if position == 0.0:
                # Open new trade
                position = s
                entry_price = c
                entry_bar = i
            elif s != position:
                # Opposite signal
                if on_opposite == "reverse":
                    # Close current + open opposite
                    ret = float(position * (c - entry_price) / entry_price - cost)
                    trade_returns.append(ret)
                    if position == 1.0:
                        long_returns.append(ret)
                    else:
                        short_returns.append(ret)
                    _log_trade(i, c, ret, "signal")
                    cumret *= (1 + ret)
                    position = s
                    entry_price = c
                    entry_bar = i
                else:  # close_only
                    # Close current, go flat
                    ret = float(position * (c - entry_price) / entry_price - cost)
                    trade_returns.append(ret)
                    if position == 1.0:
                        long_returns.append(ret)
                    else:
                        short_returns.append(ret)
                    _log_trade(i, c, ret, "close_only")
                    cumret *= (1 + ret)
                    position = 0.0
                    entry_price = None
                    entry_bar = None
                    cooldown_remaining = cooldown

        equity_vals[i] = capital * cumret

    # Close last open trade at final close
    if position != 0.0 and entry_price is not None and n > 0 and entry_price != 0:
        c = close_arr[-1]
        ret = float(position * (c - entry_price) / entry_price - cost)
        trade_returns.append(ret)
        if position == 1.0:
            long_returns.append(ret)
        else:
            short_returns.append(ret)
        _log_trade(n - 1, c, ret, "end")
        cumret *= (1 + ret)
        equity_vals[-1] = capital * cumret

    # Bar returns for Sharpe (approximate)
    bar_returns = np.zeros(n)
    for i in range(1, n):
        if close_arr[i - 1] != 0 and sig_arr[i - 1] != 0:
            bar_returns[i] = sig_arr[i - 1] * (close_arr[i] - close_arr[i - 1]) / close_arr[i - 1]

    return {
        "equity": pd.Series(equity_vals, index=close.index),
        "trade_returns": trade_returns,
        "long_returns": long_returns,
        "short_returns": short_returns,
        "bar_returns": bar_returns,
        "trade_log": trade_log,
    }


# ── run_strategy: the v2 orchestrator ────────────────────────────────────

def run_strategy(feature_fn, config_fn, data_path, start_date="", end_date="",
                 validation_date="", train_split=0.7, register_model_fn=None):
    """Config-driven strategy execution. Claude writes feature_fn + config_fn,
    framework does everything else.

    Returns: results dict (same format as webapp expects)
    """
    config = config_fn()

    # Auto-correct SL/TP if Claude passed percentage instead of decimal
    for _key in ("stop_loss", "take_profit"):
        _val = config.get(_key)
        if _val is not None and _val > 0.1:  # >10% is almost certainly a percentage
            config[_key] = _val / 100.0
            print(f"[strategy] Auto-corrected {_key}: {_val} -> {config[_key]} (was percentage, converted to decimal)")

    # 1. Load data
    df, close, open_, high, low = load_ohlc(data_path, start_date, end_date)

    # 2. Feature engineering (Claude's function)
    df = feature_fn(df, close, open_, high, low)
    close = df["close"]
    open_ = df["open"]
    high = df["high"]
    low = df["low"]

    # 3. Warm-up detection: drop rows where features have NaN BEFORE any fill
    feature_cols = [c for c in df.columns if c not in ("open", "high", "low", "close")]
    raw_nans = df[feature_cols].isna().any(axis=1)
    valid_rows = ~raw_nans
    if valid_rows.any():
        first_valid = valid_rows.idxmax()
        if raw_nans.loc[:first_valid].any():
            df = df.loc[first_valid:].copy()
            close = df["close"]
            open_ = df["open"]
            high = df["high"]
            low = df["low"]

    # 4. Target
    horizon = config.get("target_horizon", 4)
    target = make_target(close, horizon=horizon)

    # 5. Split (ffill only within each partition — no bfill leak)
    mask = target.notna()
    df = df[mask].copy()
    target = target[mask]
    close = df["close"]
    high = df["high"]
    low = df["low"]

    X = df[feature_cols].copy()
    X = X.replace([np.inf, -np.inf], np.nan)

    if validation_date:
        split_idx = len(df[df.index <= validation_date])
    else:
        split_idx = int(len(df) * train_split)
    split_idx = max(1, min(split_idx, len(df) - 1))

    # ffill within train and test separately (no leak)
    X_train = X.iloc[:split_idx].ffill().fillna(0.0)
    X_test = X.iloc[split_idx:].ffill().fillna(0.0)
    X = pd.concat([X_train, X_test])

    y_train = target.iloc[:split_idx]
    y_test = target.iloc[split_idx:]
    close_train = close.iloc[:split_idx]
    close_test = close.iloc[split_idx:]
    high_test = high.iloc[split_idx:]
    low_test = low.iloc[split_idx:]

    enc = LabelEncoder()
    enc.fit([-1, 0, 1])
    y_train_enc = enc.transform(y_train)
    y_test_enc = enc.transform(y_test)

    split_dt = str(df.index[split_idx])

    sp = {
        "df": df, "X_train": X_train, "X_test": X_test,
        "y_train": y_train, "y_test": y_test,
        "y_train_enc": y_train_enc, "y_test_enc": y_test_enc,
        "enc": enc,
        "close": close, "close_train": close_train, "close_test": close_test,
        "split_idx": split_idx, "split_dt": split_dt,
        "n_train": len(X_train), "n_test": len(X_test),
    }

    # 6. Build model from config
    model = _build_model_from_config(config, X_train, y_train_enc)

    # 7. Generate signals
    threshold = config.get("signal_threshold", 0.55)
    signal_train, p_pos_train, p_neg_train = _generate_signals(model, X_train, threshold)
    signal_test, p_pos_test, p_neg_test = _generate_signals(model, X_test, threshold)

    # 8. Apply filters (order: direction → session → ATR → trend)
    direction = config.get("direction", "both")
    signal_test = _apply_direction_filter(signal_test, direction)
    signal_train = _apply_direction_filter(signal_train, direction)

    session_filter = config.get("session_filter")
    signal_test = _apply_session_filter(signal_test, signal_test.index, session_filter)
    signal_train = _apply_session_filter(signal_train, signal_train.index, session_filter)

    min_atr = config.get("min_atr")
    if min_atr is not None:
        signal_test = _apply_atr_filter(signal_test, close_test, high_test, low_test, min_atr)

    trend_filter = config.get("trend_filter")
    if trend_filter is not None:
        signal_test = _apply_trend_filter(signal_test, close_test, trend_filter)

    signal_full = pd.concat([signal_train, signal_test])

    # 9. Backtest with SL/TP/cooldown (test + train)
    high_train = high.iloc[:split_idx]
    low_train = low.iloc[:split_idx]
    has_risk = (config.get("stop_loss") is not None or
                config.get("take_profit") is not None or
                config.get("cooldown", 0) > 0 or
                config.get("on_opposite", "reverse") != "reverse")

    if has_risk:
        bt = run_backtest_v2(signal_test, close_test, high_test, low_test, config, capital=10000)
        bt_train = run_backtest_v2(signal_train, close_train, high_train, low_train, config, capital=10000)
    else:
        bt = run_backtest(signal_test, close_test, capital=10000)
        bt_train = run_backtest(signal_train, close_train, capital=10000)

    # 10. Metrics
    metrics = compute_metrics(bt, close_test, capital=10000)

    # 11. Pre-compute all trade stats (single source of truth)
    pre_stats = {
        "train_stats": compute_trade_stats(bt_train.get("trade_returns", []), capital=10000),
        "test_stats":  compute_trade_stats(bt.get("trade_returns", []), capital=10000),
        "long_stats":  compute_trade_stats(bt.get("long_returns", []), capital=10000),
        "short_stats": compute_trade_stats(bt.get("short_returns", []), capital=10000),
    }

    # 12. Register model
    if register_model_fn is not None:
        register_model_fn(model)

    # 13. Build return dict
    return build_return_dict(sp, bt, metrics, model, feature_cols,
                             signal_full, p_pos_test, p_neg_test, custom_figs=[],
                             bt_train_result=bt_train, pre_stats=pre_stats)

# ── End strategy_utils ──


DATA_PATH = '/root/Desktop/QuantifyMe/data/ohlc/GBPUSD_15min.parquet'
START_DATE = '2026-04-15'
END_DATE = '2026-05-25'
VALIDATION_DATE = ""
TRAIN_SPLIT = 0.7

# SECTION 1 — FEATURE ENGINEERING
def feature_engineering(df, close, open_, high, low):
    # ── EMA crossover core signals ──────────────────────────────────────────
    ema9  = close.ewm(span=9,  adjust=False).mean()
    ema21 = close.ewm(span=21, adjust=False).mean()
    ema50 = close.ewm(span=50, adjust=False).mean()
    ema200 = close.ewm(span=200, adjust=False).mean()

    df["ema9"]  = ema9
    df["ema21"] = ema21
    df["ema50"] = ema50
    df["ema200"] = ema200

    # Raw spread and normalised spread
    df["ema_diff"]       = ema9 - ema21
    df["ema_diff_norm"]  = (ema9 - ema21) / close

    # Cross signal: +1 when ema9 > ema21, -1 otherwise
    df["ema_cross_sign"] = np.where(ema9 > ema21, 1.0, -1.0)

    # Momentum of the spread (rate of change of spread)
    df["ema_diff_roc1"]  = df["ema_diff"].diff(1)
    df["ema_diff_roc3"]  = df["ema_diff"].diff(3)

    # Distance of price from ema50 and ema200 (normalised)
    df["dist_ema50"]  = (close - ema50)  / close
    df["dist_ema200"] = (close - ema200) / close

    # ── RSI (14) ────────────────────────────────────────────────────────────
    delta  = close.diff()
    gain   = delta.clip(lower=0.0)
    loss   = (-delta).clip(lower=0.0)
    avg_g  = gain.ewm(com=13, adjust=False).mean()
    avg_l  = loss.ewm(com=13, adjust=False).mean()
    rs     = avg_g / avg_l.replace(0.0, np.nan)
    rsi14  = 100.0 - 100.0 / (1.0 + rs)
    df["rsi14"] = rsi14

    # RSI normalised and centred
    df["rsi14_norm"] = (rsi14 - 50.0) / 50.0

    # ── MACD ────────────────────────────────────────────────────────────────
    ema12      = close.ewm(span=12, adjust=False).mean()
    ema26      = close.ewm(span=26, adjust=False).mean()
    macd_line  = ema12 - ema26
    signal_ln  = macd_line.ewm(span=9, adjust=False).mean()
    macd_hist  = macd_line - signal_ln

    df["macd_line"]    = macd_line / close
    df["macd_signal"]  = signal_ln / close
    df["macd_hist"]    = macd_hist / close
    df["macd_cross"]   = np.where(macd_line > signal_ln, 1.0, -1.0)

    # ── Bollinger Bands (20, 2) ──────────────────────────────────────────────
    bb_mid   = close.rolling(20).mean()
    bb_std   = close.rolling(20).std(ddof=0)
    bb_upper = bb_mid + 2.0 * bb_std
    bb_lower = bb_mid - 2.0 * bb_std
    bb_width = (bb_upper - bb_lower) / bb_mid.replace(0.0, np.nan)
    bb_pct   = (close - bb_lower) / (bb_upper - bb_lower).replace(0.0, np.nan)

    df["bb_width"] = bb_width
    df["bb_pct"]   = bb_pct

    # ── ATR (14) ─────────────────────────────────────────────────────────────
    tr = pd.concat([
        high - low,
        (high - close.shift(1)).abs(),
        (low  - close.shift(1)).abs()
    ], axis=1).max(axis=1)
    atr14 = tr.ewm(com=13, adjust=False).mean()
    df["atr14"]      = atr14
    df["natr14"]     = atr14 / close          # normalised ATR (volatility proxy)

    # ── Stochastic %K / %D (14, 3) ──────────────────────────────────────────
    low14  = low.rolling(14).min()
    high14 = high.rolling(14).max()
    stoch_k = 100.0 * (close - low14) / (high14 - low14).replace(0.0, np.nan)
    stoch_d = stoch_k.rolling(3).mean()
    df["stoch_k"]      = stoch_k / 100.0
    df["stoch_d"]      = stoch_d / 100.0
    df["stoch_diff"]   = (stoch_k - stoch_d) / 100.0

    # ── Rate of Change ───────────────────────────────────────────────────────
    df["roc1"]  = close.pct_change(1)
    df["roc4"]  = close.pct_change(4)
    df["roc8"]  = close.pct_change(8)
    df["roc16"] = close.pct_change(16)

    # ── Candle features ──────────────────────────────────────────────────────
    body       = (close - open_).abs()
    candle_rng = (high - low).replace(0.0, np.nan)
    df["body_ratio"]   = body / candle_rng
    df["upper_wick"]   = (high - pd.concat([close, open_], axis=1).max(axis=1)) / candle_rng
    df["lower_wick"]   = (pd.concat([close, open_], axis=1).min(axis=1) - low) / candle_rng
    df["candle_dir"]   = np.where(close >= open_, 1.0, -1.0)

    # ── Volume-like proxy: range relative to rolling average ────────────────
    df["range_ratio"]  = candle_rng / candle_rng.rolling(20).mean()

    # ── Lagged EMA diff features ─────────────────────────────────────────────
    for lag in [1, 2, 3, 4]:
        df[f"ema_diff_lag{lag}"] = df["ema_diff_norm"].shift(lag)

    # ── Lagged RSI ───────────────────────────────────────────────────────────
    for lag in [1, 2, 4]:
        df[f"rsi14_lag{lag}"] = df["rsi14_norm"].shift(lag)

    # ── Rolling volatility (std of returns) ──────────────────────────────────
    ret = close.pct_change()
    df["vol_8"]  = ret.rolling(8).std()
    df["vol_16"] = ret.rolling(16).std()
    df["vol_32"] = ret.rolling(32).std()

    # ── Trend strength: ADX-like (simplified) ────────────────────────────────
    plus_dm  = (high.diff()).clip(lower=0.0)
    minus_dm = (-low.diff()).clip(lower=0.0)
    overlap  = pd.concat([plus_dm, minus_dm], axis=1).min(axis=1)
    plus_dm  = plus_dm - overlap
    minus_dm = minus_dm - overlap
    smooth_tr  = tr.ewm(com=13, adjust=False).mean()
    plus_di    = 100.0 * plus_dm.ewm(com=13, adjust=False).mean()  / smooth_tr.replace(0.0, np.nan)
    minus_di   = 100.0 * minus_dm.ewm(com=13, adjust=False).mean() / smooth_tr.replace(0.0, np.nan)
    di_sum     = (plus_di + minus_di).replace(0.0, np.nan)
    adx        = ((plus_di - minus_di).abs() / di_sum * 100.0).ewm(com=13, adjust=False).mean()
    df["adx"]       = adx / 100.0
    df["plus_di"]   = plus_di / 100.0
    df["minus_di"]  = minus_di / 100.0

    # ── Session hour (UTC) ───────────────────────────────────────────────────
    if hasattr(df.index, "hour"):
        df["hour_sin"] = np.sin(2.0 * np.pi * df.index.hour / 24.0)
        df["hour_cos"] = np.cos(2.0 * np.pi * df.index.hour / 24.0)
    else:
        df["hour_sin"] = 0.0
        df["hour_cos"] = 1.0

    # ── Fill any NaN from warm-up periods ────────────────────────────────────
    df = df.bfill().ffill()

    return df


# SECTION 2 — STRATEGY CONFIG
def strategy_config():
    return {
        "title":            "EMA 9/21 Crossover + MACD Momentum (XGBoost)",
        "model_type":       "XGBClassifier",
        "model_params":     {
            "n_estimators":     400,
            "max_depth":        4,
            "learning_rate":    0.04,
            "subsample":        0.80,
            "colsample_bytree": 0.75,
            "min_child_weight": 3,
            "gamma":            0.10,
            "reg_alpha":        0.05,
            "reg_lambda":       1.50,
            "objective":        "binary:logistic",
            "tree_method":      "hist",
            "random_state":     42,
        },
        "signal_threshold": 0.55,
        "direction":        "both",
        "stop_loss":        0.0030,
        "take_profit":      0.0060,
        "cooldown":         0,
        "max_positions":    1,
        "on_opposite":      "reverse",
        "session_filter":   [6, 20],
        "min_atr":          None,
        "trend_filter":     "sma_50",
        "target_horizon":   4,
        "objective":        (
            "Maximize Sharpe ratio on EUR/USD 15-min data. "
            "Core signal: EMA(9) vs EMA(21) crossover enriched with MACD, RSI, "
            "Bollinger %B, Stochastic, ATR, ADX, candle structure and rolling "
            "volatility. XGBoost with moderate depth (4) and strong regularisation "
            "(gamma, alpha, lambda) prevents overfitting on ~6 weeks of intraday data. "
            "A 0.55 probability threshold filters low-confidence signals. "
            "A 2:1 TP:SL ratio (30 bp SL / 60 bp TP) improves the reward-risk "
            "balance. Session filter [6,20] UTC keeps the model away from the thin "
            "Asian pre-open. trend_filter sma_50 aligns entries with the prevailing "
            "short-term trend to reduce chop. Cooldown=0 and reverse-on-opposite "
            "allow continuous participation in trending EMA crossover moves."
        ),
        "notes":            (
            "round-trip cost 2e-5 is accounted for by the framework. "
            "target_horizon=4 bars (1 hour ahead) suits EMA crossover which "
            "generates medium-frequency signals rather than tick-level scalps. "
            "All features are normalised or expressed as ratios to minimise "
            "scale sensitivity for the logistic-objective XGBoost."
        ),
    }

# ── Framework v2: auto-generated wrapper ──
def train_and_backtest():
    _vd = VALIDATION_DATE if 'VALIDATION_DATE' in globals() else ''
    _ts = TRAIN_SPLIT if 'TRAIN_SPLIT' in globals() else 0.7
    return run_strategy(
        feature_engineering, strategy_config,
        DATA_PATH, START_DATE, END_DATE,
        _vd, _ts,
        register_model_fn=register_model
    )

# ╔══════════════════════════════════════════════════════════════╗
# ║                  STRATEGY REQUEST LOG                       ║
# ╚══════════════════════════════════════════════════════════════╝
# Generated        : 2026-05-25 02:26:36
# Model            : XGBoost
# Feature Eng.     : buy when RSI(14) crosses up from below 30, sell when it crosses down from above 70 + Auto-add features: ON
# Signal / Entry   : —
# Optimization     : —
# Risk Mgmt        : —
# Risk Filter      : —
# ══════════════════════════════════════════════════════════════

# ============================================================
# SECTION 0 — IMPORTS & CONSTANTS
import numpy as np
import pandas as pd

# ── Inlined strategy_utils ──
"""
strategy_utils.py — Standard utility functions for generated strategies.
Claude imports these instead of writing boilerplate from scratch.
This ensures consistent behavior across all generated strategies.
"""
import numpy as np
import pandas as pd
from sklearn.preprocessing import LabelEncoder


# Max backtest window per timeframe. A finer timeframe over a longer window
# blows up the results dict / parquet load / Modal train time (the 2026-05-12
# OOM was a 1-min × multi-year sweep) — and a 1-min strategy gains nothing from
# 2 years of 1-min bars. Enforced HERE because every training path (UI / API /
# Modal) funnels through run_strategy → load_ohlc. Env-overridable so a future
# "max plan" / dedicated-server tier can lift it.
_TF_MAX_DAYS = {
    "1min":  30,
    "5min":  90,
    "15min": 365,
    "1h":    730,
}


def _fetch_ohlc_from_internal(symbol: str, tf: str, start: str, end: str):
    """Phase 3.2: fetch parquet bytes from Server A's /internal/ohlc endpoint
    instead of reading a local file. Used inside Modal containers / Mac worker
    pool (Phase 3.4) so every train sees the same source of truth as the chart.

    Returns: pd.DataFrame (parquet decoded), or raises on any failure so the
    caller can fall back / surface a clear error in the job.
    """
    import hashlib as _hashlib, hmac as _hmac, io as _io, os as _os
    import urllib.request as _ur, urllib.parse as _urp

    base = (_os.environ.get("QM_INTERNAL_OHLC_BASE") or "").rstrip("/")
    secret = (_os.environ.get("INTERNAL_WS_SECRET") or "").strip()
    if not base:
        raise RuntimeError("QM_INTERNAL_OHLC_BASE not set")
    if not secret:
        raise RuntimeError("INTERNAL_WS_SECRET not set")

    msg = f"{symbol}|{tf}|{start}|{end}".encode("utf-8")
    sig = _hmac.new(secret.encode("utf-8"), msg, _hashlib.sha256).hexdigest()

    qs = _urp.urlencode({
        "symbol": symbol, "tf": tf,
        "start": start, "end": end, "sig": sig,
    })
    url = f"{base}/internal/ohlc?{qs}"

    req = _ur.Request(url, headers={"User-Agent": "qm-worker/1.0"})
    with _ur.urlopen(req, timeout=30) as resp:
        if resp.status != 200:
            raise RuntimeError(f"/internal/ohlc returned {resp.status}")
        payload = resp.read()
    print(f"[load_ohlc:internal] {symbol} {tf} fetched {len(payload)} bytes", flush=True)
    return pd.read_parquet(_io.BytesIO(payload))


def _parse_symbol_tf_from_path(data_path: str):
    """Pull SYMBOL + TF out of a path like .../EURUSD_1min.parquet."""
    import os as _os, re as _re
    base = _os.path.basename(str(data_path))
    m = _re.match(r"^([A-Z]{6})_(\d+min|\d+h)\.parquet$", base)
    if not m:
        return None, None
    return m.group(1), m.group(2)


def load_ohlc(data_path, start_date="", end_date=""):
    """Load OHLC parquet, sort index, filter dates. Always returns consistent format.

    The lower bound is clamped per timeframe (see _TF_MAX_DAYS) — a request for
    more history than the cap silently starts later.

    Phase 3.2: when env QM_USE_INTERNAL_OHLC=="1", fetch over HTTP from
    Server A's /internal/ohlc endpoint instead of pd.read_parquet on a local
    file (which on Modal is a stale Volume snapshot). The endpoint applies the
    same day-cap, so the local cap-check below is a defensive no-op in that
    path. Flag defaults to "0" → unchanged behavior.

    Returns: (df, close, open_, high, low)
    """
    import os as _os, re as _re

    _use_internal = _os.environ.get("QM_USE_INTERNAL_OHLC", "0") == "1"
    if _use_internal:
        _sym, _tf = _parse_symbol_tf_from_path(data_path)
        if not _sym or not _tf:
            raise RuntimeError(
                f"QM_USE_INTERNAL_OHLC=1 but DATA_PATH basename does not match "
                f"SYMBOL_TF.parquet: {data_path}"
            )
        df = _fetch_ohlc_from_internal(_sym, _tf, start_date or "", end_date or "")
    else:
        df = pd.read_parquet(data_path)
    df.index = pd.to_datetime(df.index)
    df = df.sort_index()

    # Per-timeframe window cap (timeframe inferred from the parquet filename).
    _m = _re.search(r"_(\d+min|\d+h)\.parquet$", _os.path.basename(str(data_path)))
    _tf = _m.group(1) if _m else None
    _max_days = _TF_MAX_DAYS.get(_tf)
    if _max_days and _max_days > 0 and len(df):
        _env_override = _os.environ.get(f"QM_MAX_DAYS_{_tf.upper()}")
        if _env_override and _env_override.isdigit():
            _max_days = int(_env_override)
        try:
            _eff_end = pd.Timestamp(end_date) if end_date else df.index.max()
            _eff_end = min(_eff_end, df.index.max())
            _floor = _eff_end - pd.Timedelta(days=_max_days)
            _req_start = pd.Timestamp(start_date) if start_date else df.index.min()
            if _req_start < _floor:
                print(f"[load_ohlc] {_tf} backtest window capped to {_max_days}d: "
                      f"start {_req_start.date()} -> {_floor.date()}", flush=True)
                start_date = _floor
        except Exception as _e:
            print(f"[load_ohlc] window-cap check skipped ({_e})", flush=True)

    if start_date:
        df = df[df.index >= start_date]
    if end_date:
        df = df[df.index <= end_date]

    return df, df["close"], df["open"], df["high"], df["low"]


def make_target(close, horizon=4):
    """Create target: direction N bars ahead. Default 4 bars = 1 hour on 15-min data.

    Returns: target (pd.Series of -1, 0, 1)
    """
    return np.sign(close.shift(-horizon) - close)


def split_data(df, target, feature_cols, train_split=0.7, validation_date=""):
    """Train/test split. Handles both ratio and date-based splits.
    Drops NaN from target before splitting. Encodes labels to [0,1,2].

    Returns: dict with keys:
        X_train, X_test, y_train, y_test,
        y_train_enc, y_test_enc, enc,
        close_train, close_test,
        split_idx, split_dt, n_train, n_test
    """
    # Drop NaN from target
    mask = target.notna()
    df = df[mask].copy()
    target = target[mask]

    close = df["close"]

    # Build feature matrix
    X = df[feature_cols].copy()
    X = X.bfill().ffill()
    X = X.replace([np.inf, -np.inf], np.nan).fillna(0.0)

    # Split
    if validation_date:
        split_idx = len(df[df.index <= validation_date])
    else:
        split_idx = int(len(df) * train_split)

    split_idx = max(1, min(split_idx, len(df) - 1))

    X_train = X.iloc[:split_idx]
    X_test = X.iloc[split_idx:]
    y_train = target.iloc[:split_idx]
    y_test = target.iloc[split_idx:]
    close_train = close.iloc[:split_idx]
    close_test = close.iloc[split_idx:]

    split_dt = str(df.index[split_idx])

    # Label encoding — always fit on [-1, 0, 1]
    enc = LabelEncoder()
    enc.fit([-1, 0, 1])
    y_train_enc = enc.transform(y_train)
    y_test_enc = enc.transform(y_test)

    return {
        "df": df, "X_train": X_train, "X_test": X_test,
        "y_train": y_train, "y_test": y_test,
        "y_train_enc": y_train_enc, "y_test_enc": y_test_enc,
        "enc": enc,
        "close": close, "close_train": close_train, "close_test": close_test,
        "split_idx": split_idx, "split_dt": split_dt,
        "n_train": len(X_train), "n_test": len(X_test),
    }


def compute_overlays(close, df_index):
    """Compute BB and MA overlays on full dataset. Always consistent.

    Returns: (bb_dict, ma_dict)
    """
    bb_mid = close.rolling(20).mean()
    bb_std = close.rolling(20).std()
    bb_upper = bb_mid + 2 * bb_std
    bb_lower = bb_mid - 2 * bb_std

    ma50 = close.rolling(50).mean()
    ma100 = close.rolling(100).mean()
    ma200 = close.rolling(200).mean()

    def _safe(s):
        s = s.reindex(df_index).bfill().ffill()
        return [float(x) if (x is not None and not np.isnan(x) and not np.isinf(x)) else None
                for x in s.values]

    bb = {"upper": _safe(bb_upper), "mid": _safe(bb_mid), "lower": _safe(bb_lower)}
    ma = {"ma50": _safe(ma50), "ma100": _safe(ma100), "ma200": _safe(ma200)}

    return bb, ma


def run_backtest(signal, close, capital=10000, cost=2e-5):
    """Run backtest with transaction costs.
    Uses price-based trade returns (same as webapp _compute_trades).
    Signal 0 = hold (keep current position), not close.

    Returns: dict with equity, trade_returns, long_returns, short_returns, bar_returns
    """
    sig_arr = signal.values
    price_arr = close.values
    idx = signal.index
    n = len(price_arr)

    # Trade returns — price-based (matches webapp _compute_trades exactly)
    trade_returns = []
    long_returns = []
    short_returns = []
    trade_log = []
    last_dir = None
    entry_price = None
    entry_bar = None

    for i in range(n):
        s = sig_arr[i]
        c = price_arr[i]
        if s != 0.0 and s != last_dir:
            # Direction change — close previous trade, open new
            if last_dir is not None and entry_price is not None and entry_price != 0:
                ret = float(last_dir * (c - entry_price) / entry_price - cost)
                trade_returns.append(ret)
                if last_dir == 1:
                    long_returns.append(ret)
                else:
                    short_returns.append(ret)
                trade_log.append({
                    "type": "Buy" if last_dir == 1 else "Sell",
                    "entry_time": str(idx[entry_bar]),
                    "exit_time": str(idx[i]),
                    "entry_price": round(entry_price, 5),
                    "exit_price": round(c, 5),
                    "pnl": round(last_dir * (c - entry_price), 5),
                    "pnl_pct": round(ret * 100, 3),
                    "exit_reason": "signal",
                })
            entry_price = c
            entry_bar = i
            last_dir = s

    # Close last open trade
    if last_dir is not None and entry_price is not None and n > 0 and entry_price != 0:
        c = price_arr[-1]
        ret = float(last_dir * (c - entry_price) / entry_price - cost)
        trade_returns.append(ret)
        if last_dir == 1:
            long_returns.append(ret)
        else:
            short_returns.append(ret)
        trade_log.append({
            "type": "Buy" if last_dir == 1 else "Sell",
            "entry_time": str(idx[entry_bar]),
            "exit_time": str(idx[-1]),
            "entry_price": round(entry_price, 5),
            "exit_price": round(c, 5),
            "pnl": round(last_dir * (c - entry_price), 5),
            "pnl_pct": round(ret * 100, 3),
            "exit_reason": "end",
        })

    # Equity curve from trade returns
    cumret = 1.0
    equity_vals = np.full(n, float(capital))
    trade_idx = 0
    in_trade = False
    t_entry_price = None
    t_dir = None

    for i in range(n):
        s = sig_arr[i]
        c = price_arr[i]
        if s != 0.0 and s != t_dir:
            if t_dir is not None and t_entry_price is not None and t_entry_price != 0:
                t_ret = t_dir * (c - t_entry_price) / t_entry_price - cost
                cumret *= (1 + t_ret)
            t_entry_price = c
            t_dir = s
        equity_vals[i] = capital * cumret

    # Bar returns for Sharpe
    bar_returns = np.zeros(n)
    for i in range(1, n):
        if price_arr[i - 1] != 0 and last_dir is not None:
            bar_returns[i] = sig_arr[i - 1] * (price_arr[i] - price_arr[i - 1]) / price_arr[i - 1] if sig_arr[i - 1] != 0 else 0.0

    return {
        "equity": pd.Series(equity_vals, index=close.index),
        "trade_returns": trade_returns,
        "long_returns": long_returns,
        "short_returns": short_returns,
        "bar_returns": bar_returns,
        "trade_log": trade_log,
    }


def compute_trade_stats(trades, capital=10000):
    """Single source of truth for trade statistics.
    Every display path reads from this — no recomputation anywhere.
    All values are rounded and JSON-safe (no inf/nan).
    """
    if not trades:
        return {"n": 0, "wins": 0, "losses": 0, "wr": 0, "avg": 0,
                "best": 0, "worst": 0, "ret": 0, "np": 0, "mdd": 0,
                "pf": 0, "rr": 0, "expect": 0}
    w = [r for r in trades if r > 0]
    l = [r for r in trades if r < 0]
    cumret = 1.0
    for r in trades:
        cumret *= (1 + r)
    net_p = capital * (cumret - 1)
    # Max drawdown
    eq = np.cumprod([1.0] + [1 + r for r in trades])
    peak = np.maximum.accumulate(eq)
    mdd = float(((eq - peak) / peak).min()) if len(eq) > 1 else 0.0
    # Profit Factor
    gross_w = sum(w) if w else 0
    gross_l = abs(sum(l)) if l else 0
    pf = gross_w / gross_l if gross_l > 0 else (9999.0 if gross_w > 0 else 0)
    # Risk:Reward
    avg_w = float(np.mean(w)) if w else 0
    avg_l = abs(float(np.mean(l))) if l else 0
    rr = avg_w / avg_l if avg_l > 0 else (9999.0 if avg_w > 0 else 0)
    # Expectancy
    expect = net_p / len(trades)
    return {
        "n": len(trades), "wins": len(w), "losses": len(l),
        "wr": round(len(w) / len(trades), 4),
        "avg": round(float(np.mean(trades)), 6),
        "best": round(max(w), 6) if w else 0,
        "worst": round(min(l), 6) if l else 0,
        "ret": round(cumret - 1, 6),
        "np": round(net_p, 2),
        "mdd": round(mdd, 6),
        "pf": round(pf, 2),
        "rr": round(rr, 2),
        "expect": round(expect, 2),
    }


def compute_metrics(bt_result, close_test, capital=10000):
    """Compute all standard metrics from backtest result.
    Uses trade-level compounding (same as webapp _trade_stats) for accuracy.

    Returns: dict with total_ret, bh_ret, sharpe_strat, sharpe_bh, mdd, n_trades
    """
    equity = bt_result["equity"]
    trade_returns = bt_result["trade_returns"]

    # Total return — trade-level compounding (matches webapp)
    if trade_returns:
        cumret = 1.0
        for r in trade_returns:
            cumret *= (1 + r)
        total_ret = cumret - 1
    else:
        total_ret = 0.0

    # Buy and hold
    bh_equity = capital * (close_test / close_test.iloc[0])
    bh_ret = (bh_equity.iloc[-1] - capital) / capital if capital != 0 else 0.0

    # Sharpe ratio — trade-level (matches webapp: sqrt(252*26) annualization)
    if len(trade_returns) >= 2 and float(np.std(trade_returns)) > 0:
        sharpe_strat = float(np.mean(trade_returns) / np.std(trade_returns) * np.sqrt(252 * 26))
    else:
        sharpe_strat = 0.0

    bh_rets = bh_equity.pct_change().dropna()
    if len(bh_rets) > 1 and bh_rets.std() != 0:
        sharpe_bh = float((bh_rets.mean() / bh_rets.std()) * np.sqrt(252 * 24 * 4))
    else:
        sharpe_bh = 0.0

    # Max drawdown — trade-level (matches webapp)
    if trade_returns:
        eq = np.cumprod([1.0] + [1 + r for r in trade_returns])
        peak = np.maximum.accumulate(eq)
        mdd = float(((eq - peak) / peak).min()) if len(eq) > 1 else 0.0
    else:
        mdd = 0.0

    return {
        "total_ret": float(total_ret),
        "bh_ret": float(bh_ret),
        "sharpe_strat": float(sharpe_strat) if not np.isnan(sharpe_strat) else 0.0,
        "sharpe_bh": float(sharpe_bh) if not np.isnan(sharpe_bh) else 0.0,
        "mdd": float(mdd),
        "n_trades": len(trade_returns),
    }


# Diagnostics line/histogram series (equity / drawdown / rolling_acc / conf_hist)
# only feed the small Diagnostics charts — they're never used by the price chart
# or scroll-back. On a 1-min model trained over the (2.2-capped) window these are
# still ~30k points each; downsample to a visually-identical resolution before the
# dict leaves the trainer so it doesn't carry that into Server-A RAM / Postgres.
_RESULTS_SERIES_MAX = 5000


def _downsample_idx(n, cap=_RESULTS_SERIES_MAX):
    """Evenly-spaced index list spanning [0, n-1] (first+last always kept), or
    None when no downsampling is needed (n <= cap)."""
    if n <= cap:
        return None
    return np.unique(np.linspace(0, n - 1, cap).astype(int)).tolist()


def _take(arr, idx):
    """Subset a list by an index list (idx may be None → return arr unchanged)."""
    if idx is None or not isinstance(arr, list):
        return arr
    return [arr[i] for i in idx]


# trade_log / train_trade_log are lists of per-trade dicts (display-only — the
# Trade Log tab). They scale with TRADE count, not bar count, so the bar-window
# cap (Phase 2.2) doesn't bound them — a degenerate near-every-bar model can put
# 10k+ trade dicts in the blob (>3 MB). Cap each (independently — a small-N model
# keeps every trade) to the most-recent N, recording `*_total` + `*_truncated`
# so the true count is still reported. Real strategies have far fewer than
# _TRADE_LOG_MAX trades, so this only ever bites pathological models.
_TRADE_LOG_MAX = 5000


def _cap_trade_log(tl):
    """Return (capped_list, original_len, was_truncated)."""
    if not isinstance(tl, list) or len(tl) <= _TRADE_LOG_MAX:
        return tl, (len(tl) if isinstance(tl, list) else 0), False
    return tl[-_TRADE_LOG_MAX:], len(tl), True


def build_return_dict(split_result, bt_result, metrics, model, feature_cols,
                      signal_full, p_pos_test, p_neg_test, custom_figs=None,
                      bt_train_result=None, pre_stats=None):
    """Assemble the complete return dict. Handles ALL serialization.
    Never returns Timestamps, numpy arrays, or non-JSON types.

    Returns: JSON-safe dict with all required keys
    """
    df = split_result["df"]
    close = split_result["close"]
    close_test = split_result["close_test"]
    X_test = split_result["X_test"]
    y_test = split_result["y_test"]
    equity = bt_result["equity"]
    bar_returns = bt_result["bar_returns"]

    # OHLC
    ohlc_dates = [str(x) for x in df.index.tolist()]

    def _safe_list(arr):
        return [float(x) if (x is not None and not np.isnan(x) and not np.isinf(x)) else None
                for x in arr]

    # Overlays
    bb, ma = compute_overlays(close, df.index)

    # Buy and hold equity
    capital = equity.iloc[0] if len(equity) > 0 else 10000
    bh_equity = capital * (close_test / close_test.iloc[0])

    # Confusion matrix
    from sklearn.metrics import confusion_matrix
    pred_test = model.predict(X_test)
    y_test_arr = np.asarray(y_test)
    cm = confusion_matrix(y_test_arr, pred_test, labels=[-1, 0, 1])

    # Rolling accuracy
    sig_arr = signal_full.reindex(close_test.index).values
    correct = pd.Series((pred_test == y_test_arr).astype(float), index=X_test.index)
    active_test = pd.Series(sig_arr != 0, index=close_test.index) if len(sig_arr) == len(close_test) else pd.Series(True, index=close_test.index)
    correct_active = correct.where(active_test, other=np.nan)
    rolling_acc = correct_active.rolling(30, min_periods=1).mean()

    # Feature importance
    importances = model.feature_importances_
    fi_pairs = sorted(zip(feature_cols, importances), key=lambda x: x[1])[-15:]

    # Drawdown
    rolling_max = equity.cummax()
    drawdown = (equity - rolling_max) / rolling_max.replace(0, np.nan)
    drawdown = drawdown.fillna(0.0)

    # ── Downsample the Diagnostics-only series (see _downsample_idx) ──────────
    _eq_dates = [str(x) for x in close_test.index.tolist()]
    _eq_strat = _safe_list(equity.values)
    _eq_bh    = _safe_list(bh_equity.values)
    _eq_idx   = _downsample_idx(len(_eq_dates))
    _eq_dates, _eq_strat, _eq_bh = _take(_eq_dates, _eq_idx), _take(_eq_strat, _eq_idx), _take(_eq_bh, _eq_idx)

    _ra_dates = [str(x) for x in rolling_acc.index.tolist()]
    _ra_vals  = [float(x) if (not np.isnan(x) and not np.isinf(x)) else None for x in rolling_acc.values]
    _ra_idx   = _downsample_idx(len(_ra_dates))
    _ra_dates, _ra_vals = _take(_ra_dates, _ra_idx), _take(_ra_vals, _ra_idx)

    _dd_dates = [str(x) for x in drawdown.index.tolist()]
    _dd_vals  = _safe_list(drawdown.values)
    _dd_idx   = _downsample_idx(len(_dd_dates))
    _dd_dates, _dd_vals = _take(_dd_dates, _dd_idx), _take(_dd_vals, _dd_idx)

    _cp_pos = [float(x) for x in (p_pos_test.tolist() if hasattr(p_pos_test, 'tolist') else list(p_pos_test))]
    _cp_neg = [float(x) for x in (p_neg_test.tolist() if hasattr(p_neg_test, 'tolist') else list(p_neg_test))]
    _cp_pos = _take(_cp_pos, _downsample_idx(len(_cp_pos)))
    _cp_neg = _take(_cp_neg, _downsample_idx(len(_cp_neg)))

    # ── Trade logs — display-only (Trade Log tab); cap to most-recent N with a
    #    `_total` field so the true count is still reported (see _cap_trade_log).
    #    NB: ret_dist arrays are left FULL — a downstream path in callbacks.py
    #    recomputes n_trades/win-rate from len(ret_dist), so a sample would skew
    #    the displayed counts; they're small anyway and gzip handles them.
    _tl_test, _tl_test_n, _tl_test_tr = _cap_trade_log(bt_result.get("trade_log", []))
    _tl_tr,   _tl_tr_n,   _tl_tr_tr   = _cap_trade_log(bt_train_result.get("trade_log", []) if bt_train_result else [])

    return {
        "ohlc": {
            "dates": ohlc_dates,
            "open": _safe_list(df["open"].values),
            "high": _safe_list(df["high"].values),
            "low": _safe_list(df["low"].values),
            "close": _safe_list(df["close"].values),
        },
        "signals": {
            "dates": [str(x) for x in signal_full.index.tolist()],
            "values": [float(x) for x in signal_full.values],
        },
        "bb": bb,
        "ma": ma,
        "equity": {
            "dates": _eq_dates,
            "strategy": _eq_strat,
            "bh": _eq_bh,
        },
        "feature_importance": {
            "names": [p[0] for p in fi_pairs],
            "values": [float(p[1]) for p in fi_pairs],
        },
        "conf_matrix": cm.tolist(),
        "conf_hist": {
            "p_pos": _cp_pos,
            "p_neg": _cp_neg,
        },
        "rolling_acc": {
            "dates": _ra_dates,
            "values": _ra_vals,
        },
        "drawdown": {
            "dates": _dd_dates,
            "values": _dd_vals,
        },
        "ret_dist": [float(x) for x in bt_result["trade_returns"]],
        "ret_dist_long": [float(x) for x in bt_result["long_returns"]],
        "ret_dist_short": [float(x) for x in bt_result["short_returns"]],
        "train_ret_dist": [float(x) for x in bt_train_result["trade_returns"]] if bt_train_result else [],
        "train_ret_dist_long": [float(x) for x in bt_train_result["long_returns"]] if bt_train_result else [],
        "train_ret_dist_short": [float(x) for x in bt_train_result["short_returns"]] if bt_train_result else [],
        "trade_log": _tl_test,
        "train_trade_log": _tl_tr,
        "trade_log_total": _tl_test_n,
        "train_trade_log_total": _tl_tr_n,
        "trade_log_truncated": _tl_test_tr,
        "train_trade_log_truncated": _tl_tr_tr,
        **(pre_stats or {}),
        "metrics": metrics,
        "split_dt": split_result["split_dt"],
        "split_idx": int(split_result["split_idx"]),
        "n_train": int(split_result["n_train"]),
        "n_test": int(split_result["n_test"]),
        "feature_cols": list(feature_cols),
        "custom_figs": custom_figs or [],
    }


# ════════════════════════════════════════════════════════════════════════════
# STRATEGY FRAMEWORK v2 — Config-driven architecture
# Claude writes feature_engineering() + strategy_config(). Framework does rest.
# ════════════════════════════════════════════════════════════════════════════

import importlib

_MODEL_REGISTRY = {
    "XGBClassifier":              ("xgboost", "XGBClassifier"),
    "RandomForestClassifier":     ("sklearn.ensemble", "RandomForestClassifier"),
    "GradientBoostingClassifier": ("sklearn.ensemble", "GradientBoostingClassifier"),
    "LogisticRegression":         ("sklearn.linear_model", "LogisticRegression"),
    "ExtraTreesClassifier":       ("sklearn.ensemble", "ExtraTreesClassifier"),
    "AdaBoostClassifier":         ("sklearn.ensemble", "AdaBoostClassifier"),
}


def _build_model_from_config(config, X_train, y_train_enc):
    """Build, fit, and wrap a model from strategy_config dict."""
    model_type = config.get("model_type", "RandomForestClassifier")
    model_params = dict(config.get("model_params", {}))

    if model_type not in _MODEL_REGISTRY:
        raise ValueError(f"Unknown model_type '{model_type}'. Valid: {list(_MODEL_REGISTRY.keys())}")

    module_path, class_name = _MODEL_REGISTRY[model_type]
    mod = importlib.import_module(module_path)
    cls = getattr(mod, class_name)

    # XGBoost defaults
    if class_name == "XGBClassifier":
        model_params.setdefault("use_label_encoder", False)
        model_params.setdefault("eval_metric", "mlogloss")
        model_params.setdefault("tree_method", "hist")
        # Determinism > speed (2026-05-25). XGBoost hist with n_jobs=-1 is
        # NON-reproducible even with random_state set — the parallel histogram
        # gradient-sum order varies across threads, so the SAME code + data
        # gives a slightly different model (and backtest) every run. Forcing
        # single-thread makes training bit-reproducible so: (a) a user who
        # copies a strategy and reruns it gets identical numbers, (b) the
        # community "Live" score matches a redeploy, (c) "same code, different
        # result" support reports go away. Cost: single-threaded XGB (a few
        # seconds slower on large windows; hist is fast so it's minor). FORCED
        # (not setdefault) so the guarantee can't be silently broken by a
        # strategy passing n_jobs. Exact reproducibility holds within the
        # platform (pinned versions / same Modal image); a user's own machine
        # with different xgboost/numpy/CPU can still differ in low-order bits.
        model_params["n_jobs"] = 1

    # Common defaults
    model_params.setdefault("random_state", 42)

    from model_wrapper import ModelWrapper

    clf = cls(**model_params)
    clf.fit(X_train, y_train_enc)

    enc = LabelEncoder()
    enc.fit([-1, 0, 1])
    return ModelWrapper(clf, original_classes=enc.classes_, n_features=X_train.shape[1])


def _generate_signals(model, X, threshold):
    """Framework-owned signal generation. Deterministic threshold logic."""
    proba = model.predict_proba(X)
    classes = list(model.classes_)

    idx_pos = classes.index(1) if 1 in classes else None
    idx_neg = classes.index(-1) if -1 in classes else None

    p_pos = proba[:, idx_pos] if idx_pos is not None else np.zeros(len(X))
    p_neg = proba[:, idx_neg] if idx_neg is not None else np.zeros(len(X))

    signal_vals = np.zeros(len(X))
    signal_vals = np.where(p_pos >= threshold, 1.0, signal_vals)
    signal_vals = np.where(p_neg >= threshold, -1.0, signal_vals)
    # Both exceed: pick stronger
    both = (p_pos >= threshold) & (p_neg >= threshold)
    signal_vals[both] = np.where(p_pos[both] >= p_neg[both], 1.0, -1.0)

    return pd.Series(signal_vals, index=X.index), p_pos, p_neg


# ── Filter functions (all no-ops when config value is None) ──────────────

def _apply_direction_filter(signal, direction):
    """Zero out signals that don't match allowed direction."""
    if direction is None or direction == "both":
        return signal
    s = signal.copy()
    if direction == "long":
        s[s < 0] = 0.0
    elif direction == "short":
        s[s > 0] = 0.0
    return s


def _apply_session_filter(signal, index, session_hours):
    """Zero out signals outside session hours [start, end] UTC."""
    if session_hours is None:
        return signal
    s = signal.copy()
    start_h, end_h = session_hours[0], session_hours[1]
    hours = index.hour
    if start_h <= end_h:
        mask = (hours >= start_h) & (hours < end_h)
    else:  # wrap around midnight, e.g. [22, 6]
        mask = (hours >= start_h) | (hours < end_h)
    s[~mask] = 0.0
    return s


def _apply_atr_filter(signal, close, high, low, min_atr):
    """Zero out signals when NATR(14) is below threshold."""
    if min_atr is None:
        return signal
    hl = high - low
    hc = (high - close.shift(1)).abs()
    lc = (low - close.shift(1)).abs()
    tr = pd.concat([hl, hc, lc], axis=1).max(axis=1)
    atr14 = tr.ewm(com=13, adjust=False).mean()
    natr = atr14 / close.replace(0, np.nan)
    s = signal.copy()
    s[natr < min_atr] = 0.0
    return s


def _apply_trend_filter(signal, close, trend_filter):
    """Only allow signals aligned with trend. e.g. 'sma_50': longs above SMA, shorts below."""
    if trend_filter is None:
        return signal
    # Parse: "sma_50" → SMA with period 50
    parts = trend_filter.lower().replace("-", "_").split("_")
    if len(parts) >= 2 and parts[0] in ("sma", "ema"):
        period = int(parts[1])
    else:
        return signal  # unknown filter, skip
    if parts[0] == "sma":
        trend_line = close.rolling(period).mean()
    else:
        trend_line = close.ewm(span=period, adjust=False).mean()
    s = signal.copy()
    # Longs only above trend, shorts only below
    s[(s > 0) & (close < trend_line)] = 0.0
    s[(s < 0) & (close > trend_line)] = 0.0
    return s


# ── run_backtest_v2: framework-owned SL/TP/cooldown/position management ──

def run_backtest_v2(signal, close, high, low, config, capital=10000, cost=2e-5):
    """Backtest with SL/TP/cooldown/direction handling built into the engine.
    Unlike run_backtest (v1), this function handles position exits internally.

    Returns: same dict shape as run_backtest()
    """
    stop_loss = config.get("stop_loss")
    take_profit = config.get("take_profit")
    cooldown = config.get("cooldown", 0)
    on_opposite = config.get("on_opposite", "reverse")

    sig_arr = signal.values
    close_arr = close.values
    high_arr = high.values
    low_arr = low.values
    idx = signal.index
    n = len(close_arr)

    trade_returns = []
    long_returns = []
    short_returns = []
    trade_log = []
    equity_vals = np.full(n, float(capital))
    cumret = 1.0

    position = 0.0       # current direction: 1.0, -1.0, or 0.0 (flat)
    entry_price = None
    entry_bar = None      # index into arrays for entry time
    cooldown_remaining = 0

    def _log_trade(exit_bar, exit_px, ret, reason):
        trade_log.append({
            "type": "Buy" if position == 1.0 else "Sell",
            "entry_time": str(idx[entry_bar]),
            "exit_time": str(idx[exit_bar]),
            "entry_price": round(entry_price, 5),
            "exit_price": round(exit_px, 5),
            "pnl": round(position * (exit_px - entry_price), 5),
            "pnl_pct": round(ret * 100, 3),
            "exit_reason": reason,
        })

    for i in range(n):
        c = close_arr[i]
        h = high_arr[i]
        lo = low_arr[i]
        s = sig_arr[i]

        # 1. Check SL/TP if in trade
        if position != 0.0 and entry_price is not None:
            hit_sl = False
            hit_tp = False
            exit_price = None

            if position == 1.0:  # long
                if stop_loss is not None and lo <= entry_price * (1 - stop_loss):
                    hit_sl = True
                    exit_price = entry_price * (1 - stop_loss)
                elif take_profit is not None and h >= entry_price * (1 + take_profit):
                    hit_tp = True
                    exit_price = entry_price * (1 + take_profit)
            else:  # short
                if stop_loss is not None and h >= entry_price * (1 + stop_loss):
                    hit_sl = True
                    exit_price = entry_price * (1 + stop_loss)
                elif take_profit is not None and lo <= entry_price * (1 - take_profit):
                    hit_tp = True
                    exit_price = entry_price * (1 - take_profit)

            if hit_sl or hit_tp:
                ret = float(position * (exit_price - entry_price) / entry_price - cost)
                trade_returns.append(ret)
                if position == 1.0:
                    long_returns.append(ret)
                else:
                    short_returns.append(ret)
                _log_trade(i, exit_price, ret, "SL" if hit_sl else "TP")
                cumret *= (1 + ret)
                position = 0.0
                entry_price = None
                entry_bar = None
                cooldown_remaining = cooldown
                equity_vals[i] = capital * cumret
                continue

        # 2. Cooldown
        if cooldown_remaining > 0:
            cooldown_remaining -= 1
            equity_vals[i] = capital * cumret
            continue

        # 3. Signal processing
        if s != 0.0:
            if position == 0.0:
                # Open new trade
                position = s
                entry_price = c
                entry_bar = i
            elif s != position:
                # Opposite signal
                if on_opposite == "reverse":
                    # Close current + open opposite
                    ret = float(position * (c - entry_price) / entry_price - cost)
                    trade_returns.append(ret)
                    if position == 1.0:
                        long_returns.append(ret)
                    else:
                        short_returns.append(ret)
                    _log_trade(i, c, ret, "signal")
                    cumret *= (1 + ret)
                    position = s
                    entry_price = c
                    entry_bar = i
                else:  # close_only
                    # Close current, go flat
                    ret = float(position * (c - entry_price) / entry_price - cost)
                    trade_returns.append(ret)
                    if position == 1.0:
                        long_returns.append(ret)
                    else:
                        short_returns.append(ret)
                    _log_trade(i, c, ret, "close_only")
                    cumret *= (1 + ret)
                    position = 0.0
                    entry_price = None
                    entry_bar = None
                    cooldown_remaining = cooldown

        equity_vals[i] = capital * cumret

    # Close last open trade at final close
    if position != 0.0 and entry_price is not None and n > 0 and entry_price != 0:
        c = close_arr[-1]
        ret = float(position * (c - entry_price) / entry_price - cost)
        trade_returns.append(ret)
        if position == 1.0:
            long_returns.append(ret)
        else:
            short_returns.append(ret)
        _log_trade(n - 1, c, ret, "end")
        cumret *= (1 + ret)
        equity_vals[-1] = capital * cumret

    # Bar returns for Sharpe (approximate)
    bar_returns = np.zeros(n)
    for i in range(1, n):
        if close_arr[i - 1] != 0 and sig_arr[i - 1] != 0:
            bar_returns[i] = sig_arr[i - 1] * (close_arr[i] - close_arr[i - 1]) / close_arr[i - 1]

    return {
        "equity": pd.Series(equity_vals, index=close.index),
        "trade_returns": trade_returns,
        "long_returns": long_returns,
        "short_returns": short_returns,
        "bar_returns": bar_returns,
        "trade_log": trade_log,
    }


# ── run_strategy: the v2 orchestrator ────────────────────────────────────

def run_strategy(feature_fn, config_fn, data_path, start_date="", end_date="",
                 validation_date="", train_split=0.7, register_model_fn=None):
    """Config-driven strategy execution. Claude writes feature_fn + config_fn,
    framework does everything else.

    Returns: results dict (same format as webapp expects)
    """
    config = config_fn()

    # Auto-correct SL/TP if Claude passed percentage instead of decimal
    for _key in ("stop_loss", "take_profit"):
        _val = config.get(_key)
        if _val is not None and _val > 0.1:  # >10% is almost certainly a percentage
            config[_key] = _val / 100.0
            print(f"[strategy] Auto-corrected {_key}: {_val} -> {config[_key]} (was percentage, converted to decimal)")

    # 1. Load data
    df, close, open_, high, low = load_ohlc(data_path, start_date, end_date)

    # 2. Feature engineering (Claude's function)
    df = feature_fn(df, close, open_, high, low)
    close = df["close"]
    open_ = df["open"]
    high = df["high"]
    low = df["low"]

    # 3. Warm-up detection: drop rows where features have NaN BEFORE any fill
    feature_cols = [c for c in df.columns if c not in ("open", "high", "low", "close")]
    raw_nans = df[feature_cols].isna().any(axis=1)
    valid_rows = ~raw_nans
    if valid_rows.any():
        first_valid = valid_rows.idxmax()
        if raw_nans.loc[:first_valid].any():
            df = df.loc[first_valid:].copy()
            close = df["close"]
            open_ = df["open"]
            high = df["high"]
            low = df["low"]

    # 4. Target
    horizon = config.get("target_horizon", 4)
    target = make_target(close, horizon=horizon)

    # 5. Split (ffill only within each partition — no bfill leak)
    mask = target.notna()
    df = df[mask].copy()
    target = target[mask]
    close = df["close"]
    high = df["high"]
    low = df["low"]

    X = df[feature_cols].copy()
    X = X.replace([np.inf, -np.inf], np.nan)

    if validation_date:
        split_idx = len(df[df.index <= validation_date])
    else:
        split_idx = int(len(df) * train_split)
    split_idx = max(1, min(split_idx, len(df) - 1))

    # ffill within train and test separately (no leak)
    X_train = X.iloc[:split_idx].ffill().fillna(0.0)
    X_test = X.iloc[split_idx:].ffill().fillna(0.0)
    X = pd.concat([X_train, X_test])

    y_train = target.iloc[:split_idx]
    y_test = target.iloc[split_idx:]
    close_train = close.iloc[:split_idx]
    close_test = close.iloc[split_idx:]
    high_test = high.iloc[split_idx:]
    low_test = low.iloc[split_idx:]

    enc = LabelEncoder()
    enc.fit([-1, 0, 1])
    y_train_enc = enc.transform(y_train)
    y_test_enc = enc.transform(y_test)

    split_dt = str(df.index[split_idx])

    sp = {
        "df": df, "X_train": X_train, "X_test": X_test,
        "y_train": y_train, "y_test": y_test,
        "y_train_enc": y_train_enc, "y_test_enc": y_test_enc,
        "enc": enc,
        "close": close, "close_train": close_train, "close_test": close_test,
        "split_idx": split_idx, "split_dt": split_dt,
        "n_train": len(X_train), "n_test": len(X_test),
    }

    # 6. Build model from config
    model = _build_model_from_config(config, X_train, y_train_enc)

    # 7. Generate signals
    threshold = config.get("signal_threshold", 0.55)
    signal_train, p_pos_train, p_neg_train = _generate_signals(model, X_train, threshold)
    signal_test, p_pos_test, p_neg_test = _generate_signals(model, X_test, threshold)

    # 8. Apply filters (order: direction → session → ATR → trend)
    direction = config.get("direction", "both")
    signal_test = _apply_direction_filter(signal_test, direction)
    signal_train = _apply_direction_filter(signal_train, direction)

    session_filter = config.get("session_filter")
    signal_test = _apply_session_filter(signal_test, signal_test.index, session_filter)
    signal_train = _apply_session_filter(signal_train, signal_train.index, session_filter)

    min_atr = config.get("min_atr")
    if min_atr is not None:
        signal_test = _apply_atr_filter(signal_test, close_test, high_test, low_test, min_atr)

    trend_filter = config.get("trend_filter")
    if trend_filter is not None:
        signal_test = _apply_trend_filter(signal_test, close_test, trend_filter)

    signal_full = pd.concat([signal_train, signal_test])

    # 9. Backtest with SL/TP/cooldown (test + train)
    high_train = high.iloc[:split_idx]
    low_train = low.iloc[:split_idx]
    has_risk = (config.get("stop_loss") is not None or
                config.get("take_profit") is not None or
                config.get("cooldown", 0) > 0 or
                config.get("on_opposite", "reverse") != "reverse")

    if has_risk:
        bt = run_backtest_v2(signal_test, close_test, high_test, low_test, config, capital=10000)
        bt_train = run_backtest_v2(signal_train, close_train, high_train, low_train, config, capital=10000)
    else:
        bt = run_backtest(signal_test, close_test, capital=10000)
        bt_train = run_backtest(signal_train, close_train, capital=10000)

    # 10. Metrics
    metrics = compute_metrics(bt, close_test, capital=10000)

    # 11. Pre-compute all trade stats (single source of truth)
    pre_stats = {
        "train_stats": compute_trade_stats(bt_train.get("trade_returns", []), capital=10000),
        "test_stats":  compute_trade_stats(bt.get("trade_returns", []), capital=10000),
        "long_stats":  compute_trade_stats(bt.get("long_returns", []), capital=10000),
        "short_stats": compute_trade_stats(bt.get("short_returns", []), capital=10000),
    }

    # 12. Register model
    if register_model_fn is not None:
        register_model_fn(model)

    # 13. Build return dict
    return build_return_dict(sp, bt, metrics, model, feature_cols,
                             signal_full, p_pos_test, p_neg_test, custom_figs=[],
                             bt_train_result=bt_train, pre_stats=pre_stats)

# ── End strategy_utils ──


DATA_PATH = '/root/Desktop/QuantifyMe/data/ohlc/EURUSD_15min.parquet'
START_DATE = '2026-04-15'
END_DATE = '2026-05-25'
VALIDATION_DATE = ""
TRAIN_SPLIT = 0.7

# SECTION 1 — FEATURE ENGINEERING
def feature_engineering(df, close, open_, high, low):

    # --- RSI(14) ---
    period = 14
    delta = close.diff()
    gain = delta.clip(lower=0)
    loss = -delta.clip(upper=0)
    avg_gain = gain.ewm(alpha=1.0 / period, min_periods=period, adjust=False).mean()
    avg_loss = loss.ewm(alpha=1.0 / period, min_periods=period, adjust=False).mean()
    rs = avg_gain / avg_loss.replace(0, np.nan)
    df["rsi_14"] = 100.0 - (100.0 / (1.0 + rs))

    # --- RSI crossover signals: cross up from below 30, cross down from above 70 ---
    rsi_prev = df["rsi_14"].shift(1)
    df["rsi_cross_up30"] = np.where(
        (rsi_prev < 30) & (df["rsi_14"] >= 30), 1.0, 0.0
    )
    df["rsi_cross_dn70"] = np.where(
        (rsi_prev > 70) & (df["rsi_14"] <= 70), 1.0, 0.0
    )

    # --- RSI distance from thresholds (signed) ---
    df["rsi_dist_30"] = df["rsi_14"] - 30.0
    df["rsi_dist_70"] = df["rsi_14"] - 70.0
    df["rsi_dist_50"] = df["rsi_14"] - 50.0

    # --- RSI(5) for short-term momentum ---
    period5 = 5
    delta5 = close.diff()
    gain5 = delta5.clip(lower=0)
    loss5 = -delta5.clip(upper=0)
    avg_gain5 = gain5.ewm(alpha=1.0 / period5, min_periods=period5, adjust=False).mean()
    avg_loss5 = loss5.ewm(alpha=1.0 / period5, min_periods=period5, adjust=False).mean()
    rs5 = avg_gain5 / avg_loss5.replace(0, np.nan)
    df["rsi_5"] = 100.0 - (100.0 / (1.0 + rs5))

    # --- RSI(28) for longer-term regime ---
    period28 = 28
    delta28 = close.diff()
    gain28 = delta28.clip(lower=0)
    loss28 = -delta28.clip(upper=0)
    avg_gain28 = gain28.ewm(alpha=1.0 / period28, min_periods=period28, adjust=False).mean()
    avg_loss28 = loss28.ewm(alpha=1.0 / period28, min_periods=period28, adjust=False).mean()
    rs28 = avg_gain28 / avg_loss28.replace(0, np.nan)
    df["rsi_28"] = 100.0 - (100.0 / (1.0 + rs28))

    # --- Bollinger Bands (20, 2) ---
    bb_period = 20
    bb_mid = close.rolling(bb_period).mean()
    bb_std = close.rolling(bb_period).std()
    bb_upper = bb_mid + 2.0 * bb_std
    bb_lower = bb_mid - 2.0 * bb_std
    df["bb_mid"] = bb_mid
    df["bb_width"] = np.where(bb_mid != 0, (bb_upper - bb_lower) / bb_mid, np.nan)
    df["bb_pct_b"] = np.where(
        (bb_upper - bb_lower) != 0,
        (close - bb_lower) / (bb_upper - bb_lower),
        0.5
    )
    df["price_vs_bb_mid"] = close - bb_mid

    # --- ATR(14) for volatility ---
    tr1 = high - low
    tr2 = (high - close.shift(1)).abs()
    tr3 = (low - close.shift(1)).abs()
    true_range = pd.concat([tr1, tr2, tr3], axis=1).max(axis=1)
    df["atr_14"] = true_range.ewm(alpha=1.0 / 14, min_periods=14, adjust=False).mean()
    df["natr_14"] = np.where(close != 0, df["atr_14"] / close, np.nan)

    # --- MACD (12, 26, 9) ---
    ema12 = close.ewm(span=12, min_periods=12, adjust=False).mean()
    ema26 = close.ewm(span=26, min_periods=26, adjust=False).mean()
    macd_line = ema12 - ema26
    signal_line = macd_line.ewm(span=9, min_periods=9, adjust=False).mean()
    df["macd"] = macd_line
    df["macd_signal"] = signal_line
    df["macd_hist"] = macd_line - signal_line
    df["macd_cross_up"] = np.where(
        (macd_line.shift(1) < signal_line.shift(1)) & (macd_line >= signal_line), 1.0, 0.0
    )
    df["macd_cross_dn"] = np.where(
        (macd_line.shift(1) > signal_line.shift(1)) & (macd_line <= signal_line), 1.0, 0.0
    )

    # --- Stochastic Oscillator (14, 3) ---
    stoch_period = 14
    lowest_low = low.rolling(stoch_period).min()
    highest_high = high.rolling(stoch_period).max()
    denom = (highest_high - lowest_low).replace(0, np.nan)
    stoch_k = 100.0 * (close - lowest_low) / denom
    stoch_d = stoch_k.rolling(3).mean()
    df["stoch_k"] = stoch_k
    df["stoch_d"] = stoch_d
    df["stoch_cross_up"] = np.where(
        (stoch_k.shift(1) < stoch_d.shift(1)) & (stoch_k >= stoch_d) & (stoch_k < 30), 1.0, 0.0
    )
    df["stoch_cross_dn"] = np.where(
        (stoch_k.shift(1) > stoch_d.shift(1)) & (stoch_k <= stoch_d) & (stoch_k > 70), 1.0, 0.0
    )

    # --- SMA trend features ---
    df["sma_20"] = close.rolling(20).mean()
    df["sma_50"] = close.rolling(50).mean()
    df["sma_200"] = close.rolling(200).mean()
    df["price_vs_sma20"] = (close - df["sma_20"]) / df["sma_20"].replace(0, np.nan)
    df["price_vs_sma50"] = (close - df["sma_50"]) / df["sma_50"].replace(0, np.nan)
    df["sma20_vs_sma50"] = (df["sma_20"] - df["sma_50"]) / df["sma_50"].replace(0, np.nan)

    # --- Price momentum (returns) ---
    df["ret_1"] = close.pct_change(1)
    df["ret_4"] = close.pct_change(4)
    df["ret_8"] = close.pct_change(8)
    df["ret_16"] = close.pct_change(16)

    # --- Candle body and wick features ---
    body = (close - open_).abs()
    upper_wick = high - pd.concat([close, open_], axis=1).max(axis=1)
    lower_wick = pd.concat([close, open_], axis=1).min(axis=1) - low
    candle_range = (high - low).replace(0, np.nan)
    df["body_ratio"] = body / candle_range
    df["upper_wick_ratio"] = upper_wick / candle_range
    df["lower_wick_ratio"] = lower_wick / candle_range
    df["candle_direction"] = np.where(close >= open_, 1.0, -1.0)

    # --- Volume of consecutive bars in same direction ---
    df["consec_up"] = (
        df["candle_direction"]
        .groupby((df["candle_direction"] != df["candle_direction"].shift(1)).cumsum())
        .cumcount() + 1
    ) * np.where(df["candle_direction"] > 0, 1.0, 0.0)
    df["consec_dn"] = (
        df["candle_direction"]
        .groupby((df["candle_direction"] != df["candle_direction"].shift(1)).cumsum())
        .cumcount() + 1
    ) * np.where(df["candle_direction"] < 0, 1.0, 0.0)

    # --- RSI slope (rate of change) ---
    df["rsi_slope_3"] = df["rsi_14"].diff(3)
    df["rsi_slope_5"] = df["rsi_14"].diff(5)

    # --- Rolling high/low channel breakout context ---
    df["high_20"] = high.rolling(20).max()
    df["low_20"] = low.rolling(20).min()
    df["price_pos_in_range"] = np.where(
        (df["high_20"] - df["low_20"]) != 0,
        (close - df["low_20"]) / (df["high_20"] - df["low_20"]),
        0.5
    )

    # --- RSI oversold/overbought binary flags ---
    df["rsi_oversold"] = np.where(df["rsi_14"] < 30, 1.0, 0.0)
    df["rsi_overbought"] = np.where(df["rsi_14"] > 70, 1.0, 0.0)
    df["rsi_neutral"] = np.where((df["rsi_14"] >= 40) & (df["rsi_14"] <= 60), 1.0, 0.0)

    # --- Rolling RSI min/max to track extremes ---
    df["rsi_min_10"] = df["rsi_14"].rolling(10).min()
    df["rsi_max_10"] = df["rsi_14"].rolling(10).max()
    df["rsi_range_10"] = df["rsi_max_10"] - df["rsi_min_10"]

    # --- Time-of-day features (sin/cos encoding for session awareness) ---
    if hasattr(df.index, "hour"):
        hour = df.index.hour + df.index.minute / 60.0
        df["hour_sin"] = np.sin(2.0 * np.pi * hour / 24.0)
        df["hour_cos"] = np.cos(2.0 * np.pi * hour / 24.0)
    else:
        df["hour_sin"] = 0.0
        df["hour_cos"] = 1.0

    # --- Fill NaN from warm-up periods ---
    df = df.bfill().ffill()

    return df

# SECTION 2 — STRATEGY CONFIG
def strategy_config():
    return {
        "title": "RSI Crossover Mean-Reversion (XGBoost, Sharpe)",
        "model_type": "XGBClassifier",
        "model_params": {
            "n_estimators": 400,
            "max_depth": 4,
            "learning_rate": 0.04,
            "subsample": 0.8,
            "colsample_bytree": 0.7,
            "min_child_weight": 3,
            "gamma": 0.1,
            "reg_alpha": 0.05,
            "reg_lambda": 1.2,
            "objective": "binary:logistic",
            "n_jobs": -1,
            "random_state": 42,
        },
        "signal_threshold": 0.55,
        "direction": "both",
        "stop_loss": 0.004,
        "take_profit": 0.008,
        "cooldown": 0,
        "max_positions": 1,
        "on_opposite": "reverse",
        "session_filter": [7, 18],
        "min_atr": None,
        "trend_filter": None,
        "target_horizon": 4,
        "objective": (
            "Maximize Sharpe ratio by capturing mean-reversion when RSI crosses "
            "back from oversold (<30) or overbought (>70) extremes. XGBoost with "
            "moderate depth and shrinkage prevents overfitting on the short EUR/USD "
            "window. A 2:1 TP:SL ratio (0.8%/0.4%) on 15-min bars targets clean "
            "risk-adjusted returns. Session filter restricts to liquid London/NY hours."
        ),
        "notes": (
            "Feature set combines the core RSI crossover signal with multi-period RSI, "
            "MACD histogram, Stochastic, Bollinger %B, ATR normalised volatility, "
            "price momentum across 4 horizons, candle structure ratios, and time encoding. "
            "colsample_bytree=0.7 adds diversity across trees; subsample=0.8 reduces "
            "variance. min_child_weight=3 avoids splitting on noisy one-off RSI spikes. "
            "No trend_filter so the model can express both long and short mean-reversion "
            "signals symmetrically via the 'both' direction setting."
        ),
    }

# ── Framework v2: auto-generated wrapper ──
def train_and_backtest():
    _vd = VALIDATION_DATE if 'VALIDATION_DATE' in globals() else ''
    _ts = TRAIN_SPLIT if 'TRAIN_SPLIT' in globals() else 0.7
    return run_strategy(
        feature_engineering, strategy_config,
        DATA_PATH, START_DATE, END_DATE,
        _vd, _ts,
        register_model_fn=register_model
    )