"""
Evaluation metrics shared by both baselines (and, later, any Whisper-based
model), so every experiment in experiments/EXPERIMENTS.md is measured the
same way.

Definitions used throughout this project (stated once, here, so every doc
and script agrees):

  - Positive class = END (endpoint_bool = True / turn complete).
  - False END  = model predicted END but the true label was CONTINUE.
                 This is a "premature endpoint" — the costly error in a live
                 voice agent (interrupts the user).
  - False CONTINUE = model predicted CONTINUE but the true label was END.
                 This is a "delayed endpoint" — the agent waits too long
                 before responding.

True wall-clock endpoint latency cannot be computed from this dataset (no
conversation-level timestamps — docs/INITIAL_ANALYSIS.md §7). Anywhere that
number would normally go, code in this module reports
`"endpoint_latency_ms": None` with a reason string, rather than a fabricated
number or a silently missing field.
"""

from __future__ import annotations

import time
from typing import Callable

import numpy as np
from sklearn.metrics import confusion_matrix, precision_recall_fscore_support


def classification_metrics(y_true: np.ndarray, y_pred: np.ndarray) -> dict:
    y_true = np.asarray(y_true, dtype=bool)
    y_pred = np.asarray(y_pred, dtype=bool)

    if len(y_true) == 0:
        return {"n": 0, "note": "empty evaluation set"}

    accuracy = float(np.mean(y_true == y_pred))
    precision, recall, f1, _ = precision_recall_fscore_support(
        y_true, y_pred, average="binary", zero_division=0
    )
    cm = confusion_matrix(y_true, y_pred, labels=[False, True])
    tn, fp, fn, tp = cm.ravel()

    n_actual_continue = tn + fp  # true label = CONTINUE
    n_actual_end = fn + tp       # true label = END

    false_end_rate = float(fp / n_actual_continue) if n_actual_continue > 0 else None
    false_continue_rate = float(fn / n_actual_end) if n_actual_end > 0 else None

    return {
        "n": int(len(y_true)),
        "accuracy": accuracy,
        "precision": float(precision),
        "recall": float(recall),
        "f1": float(f1),
        "confusion_matrix": {"tn": int(tn), "fp": int(fp), "fn": int(fn), "tp": int(tp)},
        "false_end_rate": false_end_rate,          # premature endpoint rate
        "false_continue_rate": false_continue_rate,  # delayed endpoint rate
        "endpoint_latency_ms": None,
        "endpoint_latency_note": (
            "Not measurable: dataset has no conversation-level timestamps "
            "(see docs/INITIAL_ANALYSIS.md §7)."
        ),
    }


def slice_metrics(
    y_true: np.ndarray,
    y_pred: np.ndarray,
    slice_labels: np.ndarray,
    min_samples: int = 20,
) -> dict:
    """Per-slice metrics (e.g. by language, by synthetic flag). Slices with
    fewer than `min_samples` are excluded and reported as such, per the
    Phase 2 brief's "only report slices with sufficient samples" rule —
    exclusion is explicit, not silent.
    """
    y_true = np.asarray(y_true, dtype=bool)
    y_pred = np.asarray(y_pred, dtype=bool)
    slice_labels = np.asarray(slice_labels)

    results = {}
    excluded = []
    for val in sorted(set(slice_labels.tolist()), key=str):
        mask = slice_labels == val
        n = int(mask.sum())
        if n < min_samples:
            excluded.append({"slice": str(val), "n": n})
            continue
        results[str(val)] = classification_metrics(y_true[mask], y_pred[mask])
    return {"slices": results, "excluded_insufficient_n": excluded, "min_samples": min_samples}


def measure_inference_latency(
    predict_fn: Callable,
    inputs: list,
    n_warmup: int = 3,
    n_repeats: int = 1,
) -> dict:
    """Wall-clock latency per single-item inference call, on THIS machine
    (CPU, exact spec unstated — see caveat below). Reported per-call, not
    batched, since production turn-detection is a per-utterance call.

    Honesty caveat this function always attaches: measured latency is
    environment-specific (CPU model, load, Python/library versions all
    matter). It should be reported as "measured on this dev/eval machine",
    never presented as a universal number, and never compared directly to
    a different environment's numbers (e.g. the upstream Smart Turn
    project's reported ~65ms Pipecat Cloud figure) without noting that
    caveat explicitly in the write-up.
    """
    for x in inputs[:n_warmup]:
        predict_fn(x)

    times_ms = []
    for _ in range(n_repeats):
        for x in inputs:
            t0 = time.perf_counter()
            predict_fn(x)
            t1 = time.perf_counter()
            times_ms.append((t1 - t0) * 1000.0)

    times_ms = np.array(times_ms)
    return {
        "n_calls": int(len(times_ms)),
        "mean_ms": float(np.mean(times_ms)),
        "median_ms": float(np.median(times_ms)),
        "p95_ms": float(np.percentile(times_ms, 95)),
        "p99_ms": float(np.percentile(times_ms, 99)),
        "min_ms": float(np.min(times_ms)),
        "max_ms": float(np.max(times_ms)),
        "caveat": "Measured on this dev/eval machine; not directly comparable across environments.",
    }