mitai-jinkendo/backend/data_layer/correlations.py

"""
Correlation Metrics Data Layer

Provides structured correlation analysis and plateau detection functions.

Functions:
    - calculate_lag_correlation(): Lag correlation between variables
    - calculate_correlation_sleep_recovery(): Sleep-recovery correlation
    - calculate_plateau_detected(): Plateau detection (weight, strength, endurance)
    - calculate_top_drivers(): Top drivers for current goals
    - calculate_correlation_confidence(): Confidence level for correlations

All functions return structured data (dict) or simple values.
Use placeholder_resolver.py for formatted strings for AI.

Phase 0c: Multi-Layer Architecture
Version: 1.0
"""

from typing import Any, Dict, List, Optional, Tuple

from datetime import datetime, timedelta, date
from db import get_db, get_cursor, r2d
import statistics

from data_layer.nutrition_body_merge import build_merged_daily_nutrition_body_rows
from data_layer.nutrition_metrics import estimate_tdee_kcal_from_latest_weight

# Lag-Korrelation (Issue #53): gleiche TDEE-Logik wie nutrition_metrics / nutrition_viz
MIN_PAIRS_LAG_CORR = 15
LAG_CORR_LOOKBACK_DAYS = 120

def calculate_lag_correlation(profile_id: str, var1: str, var2: str, max_lag_days: int = 14) -> Optional[Dict]:
    """
    Pearson-Korrelation mit Lag-Sweep (Issue 53, Data-Layer).

    C1: Tagesbilanz (kcal − TDEE wie ``estimate_tdee_kcal_from_latest_weight``) vs. ΔGewicht [t→t+L], L≥1.
    C2: Protein (g) vs. ΔMager [t→t+L] aus ``build_merged_daily_nutrition_body_rows``, L≥1.
    C3: Summe ``duration_min`` pro Tag vs. HRV oder Ruhepuls am Tag t+L (L≥0).

    Rückgabe enthält u. a. ``best_lag`` / ``best_lag_days``, ``correlation``, ``interpretation``,
    optional ``lag_details`` (r, n je Lag), mindestens ``MIN_PAIRS_LAG_CORR`` Paare am besten Lag.
    """
    v1 = (var1 or "").strip().lower()
    if v1 in ("energy", "energy_balance"):
        v1n = "energy"
    elif v1 in ("training_load", "load"):
        v1n = "training_load"
    elif v1 == "protein":
        v1n = "protein"
    else:
        v1n = v1

    if v1n == 'energy' and var2 == 'weight':
        return _normalize_lag_payload(_correlate_energy_weight(profile_id, max_lag_days))
    elif v1n == 'protein' and var2 == 'lbm':
        return _normalize_lag_payload(_correlate_protein_lbm(profile_id, max_lag_days))
    elif v1n == 'training_load' and var2 in ['hrv', 'rhr']:
        return _normalize_lag_payload(_correlate_load_vitals(profile_id, var2, max_lag_days))
    else:
        return None


def _normalize_lag_payload(raw: Optional[Dict]) -> Optional[Dict]:
    """Charts erwarten u. a. ``best_lag_days``; Layer liefert teils ``best_lag``."""
    if not raw:
        return None
    out = dict(raw)
    if out.get("best_lag_days") is None and out.get("best_lag") is not None:
        out["best_lag_days"] = out["best_lag"]
    return out


def _iso_date_key(d: Any) -> str:
    if d is None:
        return ""
    if hasattr(d, "isoformat"):
        return str(d.isoformat())[:10]
    s = str(d)
    return s[:10] if len(s) >= 10 else s


def _parse_iso_to_date(ds: str) -> Optional[date]:
    if not ds or len(ds) < 10:
        return None
    try:
        return date.fromisoformat(ds[:10])
    except ValueError:
        return None


def _pearson_r(xs: List[float], ys: List[float]) -> Optional[float]:
    """Pearson-Korrelation; mindestens ``MIN_PAIRS_LAG_CORR`` Paare."""
    n = len(xs)
    if n < MIN_PAIRS_LAG_CORR or n != len(ys):
        return None
    mx = sum(xs) / n
    my = sum(ys) / n
    num = sum((xs[i] - mx) * (ys[i] - my) for i in range(n))
    dx = sum((xs[i] - mx) ** 2 for i in range(n))
    dy = sum((ys[i] - my) ** 2 for i in range(n))
    if dx <= 1e-12 or dy <= 1e-12:
        return None
    r = num / ((dx**0.5) * (dy**0.5))
    return float(max(-1.0, min(1.0, r)))


def _direction_from_r(r: float) -> str:
    if r > 0.05:
        return "positive"
    if r < -0.05:
        return "negative"
    return "none"


def _lag_confidence(n_pairs: int, r: float) -> str:
    return calculate_correlation_confidence(n_pairs, abs(r))


def _correlate_energy_weight(profile_id: str, max_lag: int) -> Optional[Dict]:
    """
    Pearson: Tagesbilanz (kcal − TDEE wie nutrition_metrics) vs. Gewichtsdifferenz
    vom Tag t zu Tag t+L (L = 0 … max_lag). Bestes Lag nach maximalem |r|.
    """
    tdee = estimate_tdee_kcal_from_latest_weight(profile_id)
    if tdee is None or float(tdee) <= 0:
        return {
            "best_lag": None,
            "correlation": None,
            "direction": "none",
            "confidence": "insufficient",
            "data_points": 0,
            "interpretation": "Keine TDEE-Schätzung möglich (Gewicht/Demografie).",
            "reason": "no_tdee",
        }

    tdee_f = float(tdee)
    cutoff = (datetime.now() - timedelta(days=LAG_CORR_LOOKBACK_DAYS)).strftime("%Y-%m-%d")

    with get_db() as conn:
        cur = get_cursor(conn)
        cur.execute(
            """
            SELECT date::date AS d, SUM(kcal)::float AS kcal
            FROM nutrition_log
            WHERE profile_id = %s AND date >= %s::date AND kcal IS NOT NULL
            GROUP BY date
            ORDER BY date
            """,
            (profile_id, cutoff),
        )
        kcal_rows = cur.fetchall()
        cur.execute(
            """
            SELECT date::date AS d, weight::float AS weight
            FROM weight_log
            WHERE profile_id = %s AND date >= %s::date AND weight IS NOT NULL
            ORDER BY date
            """,
            (profile_id, cutoff),
        )
        w_rows = cur.fetchall()

    kcal_by: Dict[str, float] = {}
    for r in kcal_rows:
        kcal_by[_iso_date_key(r["d"])] = float(r["kcal"] or 0)
    weight_by: Dict[str, float] = {}
    for r in w_rows:
        weight_by[_iso_date_key(r["d"])] = float(r["weight"])

    balance_by = {d: kcal_by[d] - tdee_f for d in kcal_by}

    best: Optional[Tuple[int, float, int, List[Tuple[int, float]]]] = None
    lag_details: List[Dict[str, Any]] = []

    max_l = max(0, min(int(max_lag), 28))
    # Lag 0: ΔGewicht am selben Tag ist immer 0 → sinnvoll erst ab Tag 1
    for lag in range(1, max_l + 1):
        xs: List[float] = []
        ys: List[float] = []
        for ds in sorted(balance_by.keys()):
            d0 = _parse_iso_to_date(ds)
            if d0 is None:
                continue
            d1 = d0 + timedelta(days=lag)
            ds1 = d1.isoformat()
            w0 = weight_by.get(ds)
            w1 = weight_by.get(ds1)
            if w0 is None or w1 is None:
                continue
            xs.append(balance_by[ds])
            ys.append(w1 - w0)
        r = _pearson_r(xs, ys)
        n_p = len(xs)
        lag_details.append({"lag": lag, "n_pairs": n_p, "r": None if r is None else round(r, 4)})
        if r is None:
            continue
        if best is None or abs(r) > abs(best[1]):
            best = (lag, r, n_p)

    if best is None:
        return {
            "best_lag": None,
            "correlation": None,
            "direction": "none",
            "confidence": "insufficient",
            "data_points": 0,
            "interpretation": "Zu wenige gepaarte Tage mit Ernährung, Gewicht und gewähltem Lag.",
            "reason": "insufficient_pairs",
            "lag_details": lag_details,
            "tdee_kcal_used": round(tdee_f, 0),
        }

    lag_b, r_b, n_b, _ = best
    direction = _direction_from_r(r_b)
    conf = _lag_confidence(n_b, r_b)
    interp = (
        f"Tagesbilanz (kcal − TDEE ~{tdee_f:.0f}) vs. Gewichtsänderung nach {lag_b} Tagen: "
        f"r ≈ {r_b:.2f} ({direction}). "
        f"Basierend auf {n_b} Kalendertagen mit vollständigen Paaren."
    )

    return {
        "best_lag": lag_b,
        "correlation": round(r_b, 4),
        "direction": direction,
        "confidence": conf,
        "data_points": n_b,
        "interpretation": interp,
        "lag_details": lag_details,
        "tdee_kcal_used": round(tdee_f, 0),
    }


def _correlate_protein_lbm(profile_id: str, max_lag: int) -> Optional[Dict]:
    """
    Pearson: Protein (g/Tag) vs. Magermasse-Differenz (kg) vom Tag t zu t+L.
    Datenbasis: nutrition_body_merge (Caliper-LBM forward-filled wie Ernährungs-Verlauf).
    """
    merged = build_merged_daily_nutrition_body_rows(profile_id)
    if not merged:
        return {
            "best_lag": None,
            "correlation": None,
            "direction": "none",
            "confidence": "insufficient",
            "data_points": 0,
            "interpretation": "Keine zusammengeführten Ernährungs-/Körperdaten.",
            "reason": "no_merged_rows",
        }

    protein_by: Dict[str, float] = {}
    lbm_by: Dict[str, float] = {}
    for row in merged:
        ds = _iso_date_key(row.get("date"))
        if not ds:
            continue
        pg = row.get("protein_g")
        lm = row.get("lean_mass")
        if pg is not None:
            protein_by[ds] = float(pg)
        if lm is not None:
            lbm_by[ds] = float(lm)

    best: Optional[Tuple[int, float, int]] = None
    lag_details: List[Dict[str, Any]] = []
    max_l = max(0, min(int(max_lag), 28))

    for lag in range(1, max_l + 1):
        xs: List[float] = []
        ys: List[float] = []
        for ds in sorted(protein_by.keys()):
            if ds not in lbm_by:
                continue
            d0 = _parse_iso_to_date(ds)
            if d0 is None:
                continue
            d1 = d0 + timedelta(days=lag)
            ds1 = d1.isoformat()
            if ds1 not in lbm_by:
                continue
            xs.append(protein_by[ds])
            ys.append(lbm_by[ds1] - lbm_by[ds])
        r = _pearson_r(xs, ys)
        n_p = len(xs)
        lag_details.append({"lag": lag, "n_pairs": n_p, "r": None if r is None else round(r, 4)})
        if r is None:
            continue
        if best is None or abs(r) > abs(best[1]):
            best = (lag, r, n_p)

    if best is None:
        return {
            "best_lag": None,
            "correlation": None,
            "direction": "none",
            "confidence": "insufficient",
            "data_points": 0,
            "interpretation": "Zu wenige Tage mit Protein und Magermasse (Caliper) für die gewählten Lags.",
            "reason": "insufficient_pairs",
            "lag_details": lag_details,
        }

    lag_b, r_b, n_b = best
    direction = _direction_from_r(r_b)
    conf = _lag_confidence(n_b, r_b)
    interp = (
        f"Protein (g/Tag) vs. Magermasse-Änderung nach {lag_b} Tagen: r ≈ {r_b:.2f} ({direction}). "
        f"{n_b} gepaarte Tage."
    )

    return {
        "best_lag": lag_b,
        "correlation": round(r_b, 4),
        "direction": direction,
        "confidence": conf,
        "data_points": n_b,
        "interpretation": interp,
        "lag_details": lag_details,
    }


def _correlate_load_vitals(profile_id: str, vital: str, max_lag: int) -> Optional[Dict]:
    """
    Pearson: Tages-Trainingslast (Summe duration_min) vs. Vitals (HRV ms oder Ruhepuls)
    am Kalendertag t+Lag (typisch: Belastung am Vortag, Vitalwert am Folgetag bei Lag ≥ 1).
    """
    col = "hrv" if vital == "hrv" else "resting_hr"
    cutoff = (datetime.now() - timedelta(days=LAG_CORR_LOOKBACK_DAYS)).strftime("%Y-%m-%d")

    with get_db() as conn:
        cur = get_cursor(conn)
        cur.execute(
            """
            SELECT date::text AS d, COALESCE(SUM(duration_min), 0)::float AS minutes
            FROM activity_log
            WHERE profile_id = %s AND date >= %s::date
              AND duration_min IS NOT NULL AND duration_min > 0
            GROUP BY date
            ORDER BY date
            """,
            (profile_id, cutoff),
        )
        load_rows = cur.fetchall()
        cur.execute(
            f"""
            SELECT date::text AS d, {col}::float AS v
            FROM vitals_baseline
            WHERE profile_id = %s AND date >= %s::date AND {col} IS NOT NULL
            ORDER BY date
            """,
            (profile_id, cutoff),
        )
        vit_rows = cur.fetchall()

    load_by = {str(r["d"])[:10]: float(r["minutes"] or 0) for r in load_rows}
    vital_by = {str(r["d"])[:10]: float(r["v"]) for r in vit_rows}

    best: Optional[Tuple[int, float, int]] = None
    lag_details: List[Dict[str, Any]] = []
    max_l = max(0, min(int(max_lag), 28))
    vlabel = "HRV (ms)" if vital == "hrv" else "Ruhepuls (bpm)"

    for lag in range(0, max_l + 1):
        xs: List[float] = []
        ys: List[float] = []
        for ds in sorted(load_by.keys()):
            d0 = _parse_iso_to_date(ds)
            if d0 is None:
                continue
            d1 = d0 + timedelta(days=lag)
            ds1 = d1.isoformat()
            if ds1 not in vital_by:
                continue
            xs.append(load_by[ds])
            ys.append(vital_by[ds1])
        r = _pearson_r(xs, ys)
        n_p = len(xs)
        lag_details.append({"lag": lag, "n_pairs": n_p, "r": None if r is None else round(r, 4)})
        if r is None:
            continue
        if best is None or abs(r) > abs(best[1]):
            best = (lag, r, n_p)

    if best is None:
        return {
            "best_lag": None,
            "correlation": None,
            "direction": "none",
            "confidence": "insufficient",
            "data_points": 0,
            "interpretation": f"Zu wenige gepaarte Tage mit Training und {vlabel}.",
            "reason": "insufficient_pairs",
            "lag_details": lag_details,
            "vital": vital,
        }

    lag_b, r_b, n_b = best
    direction = _direction_from_r(r_b)
    conf = _lag_confidence(n_b, r_b)
    interp = (
        f"Trainingsminuten/Tag vs. {vlabel} nach {lag_b} Tagen Lag: r ≈ {r_b:.2f} ({direction}). "
        f"{n_b} Paare."
    )

    return {
        "best_lag": lag_b,
        "correlation": round(r_b, 4),
        "direction": direction,
        "confidence": conf,
        "data_points": n_b,
        "interpretation": interp,
        "lag_details": lag_details,
        "vital": vital,
    }


# ============================================================================
# C4: Sleep vs. Recovery Correlation
# ============================================================================

def calculate_correlation_sleep_recovery(profile_id: str) -> Optional[Dict]:
    """
    Correlate sleep quality/duration with recovery score
    """
    # TODO: Implement full correlation
    return {
        'correlation': 0.65,  # Strong positive (expected)
        'direction': 'positive',
        'confidence': 'high',
        'data_points': 28
    }


# ============================================================================
# C6: Plateau Detector
# ============================================================================

def calculate_plateau_detected(profile_id: str) -> Optional[Dict]:
    """
    Detect if user is in a plateau based on goal mode

    Returns:
        {
            'plateau_detected': True/False,
            'plateau_type': 'weight_loss'/'strength'/'endurance'/None,
            'confidence': 'high'/'medium'/'low',
            'duration_days': X,
            'top_factors': [list of potential causes]
        }
    """
    from data_layer.scores import get_user_focus_weights

    focus_weights = get_user_focus_weights(profile_id)

    if not focus_weights:
        return None

    # Determine primary focus area
    top_focus = max(focus_weights, key=focus_weights.get)

    # Check for plateau based on focus area
    if top_focus in ['körpergewicht', 'körperfett']:
        return _detect_weight_plateau(profile_id)
    elif top_focus == 'kraftaufbau':
        return _detect_strength_plateau(profile_id)
    elif top_focus == 'cardio':
        return _detect_endurance_plateau(profile_id)
    else:
        return None


def _detect_weight_plateau(profile_id: str) -> Dict:
    """Detect weight loss plateau"""
    from data_layer.body_metrics import calculate_weight_28d_slope
    from data_layer.nutrition_metrics import calculate_nutrition_score

    slope = calculate_weight_28d_slope(profile_id)
    nutrition_score = calculate_nutrition_score(profile_id)

    if slope is None:
        return {'plateau_detected': False, 'reason': 'Insufficient data'}

    # Plateau = flat weight for 28 days despite adherence
    is_plateau = abs(slope) < 0.02 and nutrition_score and nutrition_score > 70

    if is_plateau:
        factors = []

        # Check potential factors
        if nutrition_score > 85:
            factors.append('Hohe Adhärenz trotz Stagnation → mögliche Anpassung des Stoffwechsels')

        # Check if deficit is too small
        from data_layer.nutrition_metrics import calculate_energy_balance_7d
        balance = calculate_energy_balance_7d(profile_id)
        if balance and balance > -200:
            factors.append('Energiedefizit zu gering (<200 kcal/Tag)')

        # Check water retention (if waist is shrinking but weight stable)
        from data_layer.body_metrics import calculate_waist_28d_delta
        waist_delta = calculate_waist_28d_delta(profile_id)
        if waist_delta and waist_delta < -1:
            factors.append('Taillenumfang sinkt → mögliche Wasserretention maskiert Fettabbau')

        return {
            'plateau_detected': True,
            'plateau_type': 'weight_loss',
            'confidence': 'high' if len(factors) >= 2 else 'medium',
            'duration_days': 28,
            'top_factors': factors[:3]
        }
    else:
        return {'plateau_detected': False}


def _detect_strength_plateau(profile_id: str) -> Dict:
    """Detect strength training plateau"""
    from data_layer.body_metrics import calculate_lbm_28d_change
    from data_layer.activity_metrics import calculate_activity_score
    from data_layer.recovery_metrics import calculate_recovery_score_v2

    lbm_change = calculate_lbm_28d_change(profile_id)
    activity_score = calculate_activity_score(profile_id)
    recovery_score = calculate_recovery_score_v2(profile_id)

    if lbm_change is None:
        return {'plateau_detected': False, 'reason': 'Insufficient data'}

    # Plateau = flat LBM despite high activity score
    is_plateau = abs(lbm_change) < 0.3 and activity_score and activity_score > 75

    if is_plateau:
        factors = []

        if recovery_score and recovery_score < 60:
            factors.append('Recovery Score niedrig → möglicherweise Übertraining')

        from data_layer.nutrition_metrics import calculate_protein_adequacy_28d
        protein_score = calculate_protein_adequacy_28d(profile_id)
        if protein_score and protein_score < 70:
            factors.append('Proteinzufuhr unter Zielbereich')

        from data_layer.activity_metrics import calculate_monotony_score
        monotony = calculate_monotony_score(profile_id)
        if monotony and monotony > 2.0:
            factors.append('Hohe Trainingsmonotonie → Stimulus-Anpassung')

        return {
            'plateau_detected': True,
            'plateau_type': 'strength',
            'confidence': 'medium',
            'duration_days': 28,
            'top_factors': factors[:3]
        }
    else:
        return {'plateau_detected': False}


def _detect_endurance_plateau(profile_id: str) -> Dict:
    """Detect endurance plateau"""
    from data_layer.activity_metrics import calculate_training_minutes_week, calculate_monotony_score
    from data_layer.recovery_metrics import calculate_vo2max_trend_28d

    # TODO: Implement when vitals_baseline.vo2_max is populated
    return {'plateau_detected': False, 'reason': 'VO2max tracking not yet implemented'}


# ============================================================================
# C7: Multi-Factor Driver Panel
# ============================================================================

def calculate_top_drivers(profile_id: str) -> Optional[List[Dict]]:
    """
    Calculate top influencing factors for goal progress

    Returns list of drivers:
    [
        {
            'factor': 'Energiebilanz',
            'status': 'förderlich'/'neutral'/'hinderlich',
            'evidence': 'hoch'/'mittel'/'niedrig',
            'reason': '1-sentence explanation'
        },
        ...
    ]
    """
    drivers = []

    # 1. Energy balance
    from data_layer.nutrition_metrics import calculate_energy_balance_7d
    balance = calculate_energy_balance_7d(profile_id)
    if balance is not None:
        if -500 <= balance <= -200:
            status = 'förderlich'
            reason = f'Moderates Defizit ({int(balance)} kcal/Tag) unterstützt Fettabbau'
        elif balance < -800:
            status = 'hinderlich'
            reason = f'Sehr großes Defizit ({int(balance)} kcal/Tag) → Risiko für Magermasseverlust'
        elif -200 < balance < 200:
            status = 'neutral'
            reason = 'Energiebilanz ausgeglichen'
        else:
            status = 'neutral'
            reason = f'Energieüberschuss ({int(balance)} kcal/Tag)'

        drivers.append({
            'factor': 'Energiebilanz',
            'status': status,
            'evidence': 'hoch',
            'reason': reason
        })

    # 2. Protein adequacy
    from data_layer.nutrition_metrics import calculate_protein_adequacy_28d
    protein_score = calculate_protein_adequacy_28d(profile_id)
    if protein_score is not None:
        if protein_score >= 80:
            status = 'förderlich'
            reason = f'Proteinzufuhr konstant im Zielbereich (Score: {protein_score})'
        elif protein_score >= 60:
            status = 'neutral'
            reason = f'Proteinzufuhr teilweise im Zielbereich (Score: {protein_score})'
        else:
            status = 'hinderlich'
            reason = f'Proteinzufuhr häufig unter Zielbereich (Score: {protein_score})'

        drivers.append({
            'factor': 'Proteinzufuhr',
            'status': status,
            'evidence': 'hoch',
            'reason': reason
        })

    # 3. Sleep duration
    from data_layer.recovery_metrics import calculate_sleep_avg_duration_7d
    sleep_hours = calculate_sleep_avg_duration_7d(profile_id)
    if sleep_hours is not None:
        if sleep_hours >= 7:
            status = 'förderlich'
            reason = f'Schlafdauer ausreichend ({sleep_hours:.1f}h/Nacht)'
        elif sleep_hours >= 6.5:
            status = 'neutral'
            reason = f'Schlafdauer knapp ausreichend ({sleep_hours:.1f}h/Nacht)'
        else:
            status = 'hinderlich'
            reason = f'Schlafdauer zu gering ({sleep_hours:.1f}h/Nacht < 7h Empfehlung)'

        drivers.append({
            'factor': 'Schlafdauer',
            'status': status,
            'evidence': 'hoch',
            'reason': reason
        })

    # 4. Sleep regularity
    from data_layer.recovery_metrics import calculate_sleep_regularity_proxy
    regularity = calculate_sleep_regularity_proxy(profile_id)
    if regularity is not None:
        if regularity <= 45:
            status = 'förderlich'
            reason = f'Schlafrhythmus regelmäßig (Abweichung: {int(regularity)} min)'
        elif regularity <= 75:
            status = 'neutral'
            reason = f'Schlafrhythmus moderat variabel (Abweichung: {int(regularity)} min)'
        else:
            status = 'hinderlich'
            reason = f'Schlafrhythmus stark variabel (Abweichung: {int(regularity)} min)'

        drivers.append({
            'factor': 'Schlafregelmäßigkeit',
            'status': status,
            'evidence': 'mittel',
            'reason': reason
        })

    # 5. Training consistency
    from data_layer.activity_metrics import calculate_training_frequency_7d
    frequency = calculate_training_frequency_7d(profile_id)
    if frequency is not None:
        if 3 <= frequency <= 6:
            status = 'förderlich'
            reason = f'Trainingsfrequenz im Zielbereich ({frequency}× pro Woche)'
        elif frequency <= 2:
            status = 'hinderlich'
            reason = f'Trainingsfrequenz zu niedrig ({frequency}× pro Woche)'
        else:
            status = 'neutral'
            reason = f'Trainingsfrequenz sehr hoch ({frequency}× pro Woche) → Recovery beachten'

        drivers.append({
            'factor': 'Trainingskonsistenz',
            'status': status,
            'evidence': 'hoch',
            'reason': reason
        })

    # 6. Quality sessions
    from data_layer.activity_metrics import calculate_quality_sessions_pct
    quality_pct = calculate_quality_sessions_pct(profile_id)
    if quality_pct is not None:
        if quality_pct >= 75:
            status = 'förderlich'
            reason = f'{quality_pct}% der Trainings mit guter Qualität'
        elif quality_pct >= 50:
            status = 'neutral'
            reason = f'{quality_pct}% der Trainings mit guter Qualität'
        else:
            status = 'hinderlich'
            reason = f'Nur {quality_pct}% der Trainings mit guter Qualität'

        drivers.append({
            'factor': 'Trainingsqualität',
            'status': status,
            'evidence': 'mittel',
            'reason': reason
        })

    # 7. Recovery score
    from data_layer.recovery_metrics import calculate_recovery_score_v2
    recovery = calculate_recovery_score_v2(profile_id)
    if recovery is not None:
        if recovery >= 70:
            status = 'förderlich'
            reason = f'Recovery Score gut ({recovery}/100)'
        elif recovery >= 50:
            status = 'neutral'
            reason = f'Recovery Score moderat ({recovery}/100)'
        else:
            status = 'hinderlich'
            reason = f'Recovery Score niedrig ({recovery}/100) → mehr Erholung nötig'

        drivers.append({
            'factor': 'Recovery',
            'status': status,
            'evidence': 'hoch',
            'reason': reason
        })

    # 8. Rest day compliance
    from data_layer.activity_metrics import calculate_rest_day_compliance
    compliance = calculate_rest_day_compliance(profile_id)
    if compliance is not None:
        if compliance >= 80:
            status = 'förderlich'
            reason = f'Ruhetage gut eingehalten ({compliance}%)'
        elif compliance >= 60:
            status = 'neutral'
            reason = f'Ruhetage teilweise eingehalten ({compliance}%)'
        else:
            status = 'hinderlich'
            reason = f'Ruhetage häufig ignoriert ({compliance}%) → Übertrainingsrisiko'

        drivers.append({
            'factor': 'Ruhetagsrespekt',
            'status': status,
            'evidence': 'mittel',
            'reason': reason
        })

    # Sort by importance: hinderlich first, then förderlich, then neutral
    priority = {'hinderlich': 0, 'förderlich': 1, 'neutral': 2}
    drivers.sort(key=lambda d: priority[d['status']])

    return drivers[:8]  # Top 8 drivers


# ============================================================================
# Confidence/Evidence Levels
# ============================================================================

def calculate_correlation_confidence(data_points: int, correlation: float) -> str:
    """
    Determine confidence level for correlation

    Returns: 'high', 'medium', or 'low'
    """
    # Need sufficient data points
    if data_points < 20:
        return 'low'

    # Strong correlation with good data
    if data_points >= 40 and abs(correlation) >= 0.5:
        return 'high'
    elif data_points >= 30 and abs(correlation) >= 0.4:
        return 'medium'
    else:
        return 'low'