scripts/score_test_predictions.py

"""Score test-split predictions from a Spatial-Qwen bench run.

This script is **evaluation-only** — it does not load the model. It consumes a
`predictions.jsonl` (produced by `scripts/bench_test_generate.py` or the older
`batch_bench_spatial_beats_qa.py`) plus the original QA split, joins on
`pair_id`, and computes per-task metrics.

Supported task_names (see `all_qa_llm_by_difficulty_v2`):
    estimate_azimuth / estimate_elevation  - numeric angle
    identify_source_by_doa                  - single source label
    identify_source_by_location             - single source label
    detect_time                             - one time span per event
    detect_source                           - list of (event, start, end)

Key features:
    * Robust regex-based extractors with per-task error categories.
    * Optional LLM judge (OpenAI-compatible, uses `gpt4o_api.py` endpoint style)
      for semantic equivalence on source-identification tasks where
      surface-form exact match fails but the prediction might still be correct
      (e.g. "bell" vs "church_bell"). LLM is ALSO used as a last-resort
      answer-extractor for verbose generations before scoring.
    * Detailed parse-fail tracking: every record records a
      `parse_status` ∈ {"ok", "fail_regex", "fail_llm_extract", "fail_empty",
      "fail_no_answer_meta"}.
    * Aggregate report distinguishes task-correctness, parse rate, and
      LLM-assist rate.

Usage (no LLM):
    python scripts/score_test_predictions.py \\
        --predictions-jsonl runs/.../bench/test/<ckpt>/predictions.jsonl \\
        --qa-root /apdcephfs.../easy_filtered \\
        --output-json runs/.../bench/test/<ckpt>/result.json

Usage (LLM judge on ambiguous source identification):
    python scripts/score_test_predictions.py \\
        --predictions-jsonl ... --qa-root ... \\
        --llm-judge --llm-model gemini-3.1-pro-preview \\
        --llm-concurrency 8

The LLM judge is only invoked when:
    1. exact match fails,
    2. the task is `identify_source_by_doa` / `identify_source_by_location`
       (where synonyms are common) or the user passes `--llm-judge-all-tasks`.
    3. both prediction and answer are non-empty after cleaning.

Fully deterministic regex path still runs regardless of --llm-judge, so the
non-LLM `correct` column is always populated as a fallback.
"""

from __future__ import annotations

import argparse
import json
import math
import os
import re
import sys
import time
from collections import defaultdict
from concurrent.futures import ThreadPoolExecutor, as_completed
from dataclasses import dataclass, field
from statistics import median
from typing import Any, Callable, Dict, List, Optional, Tuple


# ---------------------------------------------------------------------------
# Regex + small parsers
# ---------------------------------------------------------------------------

FLOAT_RE = re.compile(r"[-+]?\d+(?:\.\d+)?")
# "0.1s to 5.8s" / "from 0.1 s to 5.8 s" / "0.1 - 5.8 seconds"
TIME_SPAN_RE = re.compile(
    r"(?P<start>[-+]?\d+(?:\.\d+)?)\s*(?:s|sec|seconds)?\s*(?:to|-|–|—|until)\s*(?P<end>[-+]?\d+(?:\.\d+)?)\s*(?:s|sec|seconds)?",
    re.IGNORECASE,
)
# Event spans embedded in a longer sentence, e.g.
# "The audio contains a camera from 0.0s to 9.6s, glass from 0.0s to 0.3s, ..."
EVENT_SPAN_RE = re.compile(
    # Captures "<label> from X to Y" OR "<label> active from X to Y"
    r"(?P<label>[A-Za-z][A-Za-z\s_'-]{0,40})\s+"
    r"(?:active\s+)?(?:from\s+)?(?P<start>[-+]?\d+(?:\.\d+)?)\s*(?:s|sec|seconds)?\s*"
    r"(?:to|-|–|—|until)\s*(?P<end>[-+]?\d+(?:\.\d+)?)\s*(?:s|sec|seconds)?",
    re.IGNORECASE,
)
STOPWORDS = {
    "the", "a", "an", "sound", "source", "is", "at", "from", "to", "and",
    "of", "that", "which", "are", "can", "be", "heard", "coming", "audio",
    "clip", "contains", "features", "active", "during", "this", "in", "with",
}


def normalize_text(text: Any) -> str:
    return " ".join(str(text).strip().lower().split())


def strip_stopwords(text: str) -> str:
    return " ".join(w for w in normalize_text(text).split() if w not in STOPWORDS)


def canonicalize_label(text: str) -> str:
    """Canonicalize a source/event label for comparison.

    Lowercase + underscore -> space + strip trailing punctuation + singularize
    common plurals. This is intentionally simple; LLM judge handles harder
    synonym cases.
    """
    text = normalize_text(text).replace("_", " ")
    text = text.strip(".,;:!?\"'()[]")
    # Drop articles.
    for prefix in ("a ", "an ", "the "):
        if text.startswith(prefix):
            text = text[len(prefix):]
    # Very naive plural -> singular. Guard against -es / -ches / -ies.
    if text.endswith("ies") and len(text) > 4:
        text = text[:-3] + "y"
    elif text.endswith(("sses", "shes", "ches", "xes")):
        text = text[:-2]
    elif text.endswith("s") and len(text) > 2 and not text.endswith(("ss", "us", "is", "os")):
        text = text[:-1]
    return text.strip()


_LABEL_LEADING_JUNK = re.compile(
    r"^(?:the\s+audio\s+(?:contains|features|includes)|audio\s+(?:contains|features|includes)|"
    r"the\s+clip\s+(?:contains|features)|this\s+(?:audio\s+)?(?:clip\s+)?(?:contains|features)|"
    r"and\s+|followed\s+by\s+|then\s+|next\s+)\s*",
    re.IGNORECASE,
)
_LABEL_TRAILING_JUNK = re.compile(
    r"\s*(?:active|audible|heard|happening|occurring|present|from)\s*$",
    re.IGNORECASE,
)


def _clean_event_label(raw: str) -> str:
    """Trim connector phrases / filler from a raw label span."""
    s = raw.strip()
    # Remove a few common leading phrases.
    while True:
        new = _LABEL_LEADING_JUNK.sub("", s)
        if new == s:
            break
        s = new
    s = _LABEL_TRAILING_JUNK.sub("", s)
    return canonicalize_label(s)


def parse_first_float(text: Any) -> Optional[float]:
    match = FLOAT_RE.search(str(text))
    if match is None:
        return None
    return float(match.group(0))


def parse_time_span_first(text: Any) -> Optional[Tuple[float, float]]:
    m = TIME_SPAN_RE.search(str(text))
    if m is None:
        # Fallback: take the first two floats.
        floats = [float(x) for x in FLOAT_RE.findall(str(text))]
        if len(floats) < 2:
            return None
        s, e = floats[0], floats[1]
        if e < s:
            s, e = e, s
        return s, e
    s = float(m.group("start"))
    e = float(m.group("end"))
    if e < s:
        s, e = e, s
    return s, e


def parse_all_events(text: Any) -> List[Tuple[str, float, float]]:
    """Extract (label, start, end) tuples from a detect_source-style answer."""
    events: List[Tuple[str, float, float]] = []
    for m in EVENT_SPAN_RE.finditer(str(text)):
        label_raw = m.group("label")
        label = _clean_event_label(label_raw)
        if not label or label in STOPWORDS:
            continue
        try:
            s = float(m.group("start"))
            e = float(m.group("end"))
        except ValueError:
            continue
        if e < s:
            s, e = e, s
        events.append((label, s, e))
    return events


def angle_err_deg(pred: float, target: float) -> float:
    d = pred - target
    while d > 180.0:
        d -= 360.0
    while d <= -180.0:
        d += 360.0
    return abs(d)


def interval_iou(ps: float, pe: float, gs: float, ge: float) -> float:
    inter = max(0.0, min(pe, ge) - max(ps, gs))
    union = max(pe, ge) - min(ps, gs)
    if union <= 0:
        return 0.0
    return inter / union


def mean_or_none(xs: List[float]) -> Optional[float]:
    if not xs:
        return None
    return float(sum(xs) / len(xs))


def median_or_none(xs: List[float]) -> Optional[float]:
    if not xs:
        return None
    return float(median(xs))


# ---------------------------------------------------------------------------
# LLM judge / extractor (optional, threaded)
# ---------------------------------------------------------------------------


@dataclass
class LLMConfig:
    enabled: bool = False
    model: str = "gemini-3.1-pro-preview"
    base_url: str = "https://yunwu.ai/v1"
    api_key: str = os.environ.get(
        "SPATIAL_QWEN_LLM_API_KEY",
        "sk-uLLRM86XBL9hx5CJUfpW8POle3KFS5mJ1iAiGgiMOY6Xxbjj",
    )
    concurrency: int = 4
    max_retries: int = 3
    timeout_s: float = 60.0
    judge_all_tasks: bool = False
    # Safety: don't let the judge rewrite history; we ONLY use it for
    # borderline exact-match misses on single-label tasks.
    judge_max_calls: int = 5000


class LLMJudge:
    """Thin OpenAI-compatible wrapper. Lazily imports openai to keep the
    scorer runnable with no API deps when --llm-judge is off.
    """

    def __init__(self, cfg: LLMConfig) -> None:
        self.cfg = cfg
        self._client = None
        self._lock_calls = 0

    def _client_or_none(self):
        if not self.cfg.enabled:
            return None
        if self._client is not None:
            return self._client
        try:
            from openai import OpenAI  # type: ignore
        except ImportError as exc:
            raise RuntimeError(
                "--llm-judge requires `pip install openai`"
            ) from exc
        self._client = OpenAI(base_url=self.cfg.base_url, api_key=self.cfg.api_key)
        return self._client

    def _chat(self, prompt: str) -> Optional[str]:
        if self._lock_calls >= self.cfg.judge_max_calls:
            return None
        self._lock_calls += 1
        client = self._client_or_none()
        if client is None:
            return None
        last_exc: Optional[Exception] = None
        for attempt in range(self.cfg.max_retries):
            try:
                completion = client.chat.completions.create(
                    model=self.cfg.model,
                    messages=[{"role": "user", "content": prompt}],
                    timeout=self.cfg.timeout_s,
                )
                return (completion.choices[0].message.content or "").strip()
            except Exception as exc:  # noqa: BLE001
                last_exc = exc
                time.sleep(min(2.0 * (attempt + 1), 10.0))
        sys.stderr.write(f"[llm-judge] call failed after {self.cfg.max_retries}: {last_exc}\n")
        return None

    # ---- task-specific prompts ----

    def extract_label(self, prediction: str, candidates: Optional[List[str]] = None) -> Optional[str]:
        """Ask the LLM to boil a verbose prediction down to a single short
        label (1-3 words).
        """
        hint = ""
        if candidates:
            hint = (
                "The sound source should be one of these canonical names "
                f"if possible: {', '.join(sorted(set(candidates))[:80])}. "
                "If the text mentions something not in the list, return the "
                "closest English name, 1-3 words, lowercase, no punctuation.\n"
            )
        prompt = (
            "Extract the single sound-source name mentioned as the ANSWER "
            "from the following text. Return ONLY the label in 1-3 English "
            "words, lowercase, no punctuation, no explanation. If nothing "
            "identifiable is mentioned, return the literal token UNKNOWN.\n"
            f"{hint}"
            f"TEXT:\n{prediction}\n"
            "ANSWER:"
        )
        out = self._chat(prompt)
        if out is None:
            return None
        out = out.strip().splitlines()[0].strip(" \t\"'.,`").lower()
        if not out or out == "unknown":
            return None
        return out

    def judge_equivalent(self, prediction_label: str, gold_label: str,
                          task_name: str) -> Optional[bool]:
        """Ask: is `prediction_label` a semantically valid rewording of
        `gold_label`? Returns True/False or None on API failure.
        """
        prompt = (
            "You are evaluating a spatial-audio QA system. Decide whether "
            "the MODEL ANSWER and the GOLD ANSWER refer to the SAME sound "
            "source / event category. Surface wording may differ "
            "(e.g. 'footstep' vs 'footsteps', 'bell' vs 'church_bell' vs "
            "'bell ringing'). Output exactly one token: YES or NO.\n\n"
            f"TASK: {task_name}\n"
            f"MODEL ANSWER: {prediction_label}\n"
            f"GOLD ANSWER:  {gold_label}\n\n"
            "VERDICT:"
        )
        out = self._chat(prompt)
        if out is None:
            return None
        head = out.strip().splitlines()[0].upper()
        if head.startswith("YES"):
            return True
        if head.startswith("NO"):
            return False
        return None


# ---------------------------------------------------------------------------
# Per-task scorers
# ---------------------------------------------------------------------------


@dataclass
class TaskScore:
    """Result of scoring a single (prediction, qa) pair."""
    pair_id: Any
    task_name: str
    prediction: str
    answer: str
    canonical_answer: Optional[str]
    correct: float = 0.0                # 0.0 / 1.0 (or IoU for spans)
    parse_status: str = "ok"            # ok, fail_regex, fail_llm_extract, fail_empty, fail_no_answer_meta
    metric_type: str = "exact_match"
    details: Dict[str, Any] = field(default_factory=dict)
    llm_used: bool = False


def score_estimate_angle(record: Dict[str, Any], pred_text: str,
                           is_azimuth: bool, angle_threshold_deg: float,
                           ) -> TaskScore:
    task = str(record["task_name"])
    meta = record.get("answer_meta") or {}
    key = "azimuth_deg" if is_azimuth else "elevation_deg"
    target = meta.get(key)
    # Fallback: parse from answer text.
    if target is None:
        target = parse_first_float(record.get("answer", ""))

    ts = TaskScore(
        pair_id=record.get("pair_id"),
        task_name=task,
        prediction=pred_text,
        answer=str(record.get("answer", "")),
        canonical_answer=record.get("canonical_answer"),
        metric_type=("er%d_angle" % int(angle_threshold_deg)
                     if is_azimuth
                     else "abs%d_angle" % int(angle_threshold_deg)),
    )

    if target is None:
        ts.parse_status = "fail_no_answer_meta"
        return ts
    if not str(pred_text).strip():
        ts.parse_status = "fail_empty"
        return ts

    pred = parse_first_float(pred_text)
    if pred is None:
        ts.parse_status = "fail_regex"
        return ts

    err = angle_err_deg(float(pred), float(target)) if is_azimuth \
        else abs(float(pred) - float(target))
    ts.details = {
        "predicted_deg": float(pred),
        "target_deg": float(target),
        "error_deg": float(err),
        "threshold_deg": angle_threshold_deg,
    }
    ts.correct = float(err <= angle_threshold_deg)
    return ts


def score_identify_source(record: Dict[str, Any], pred_text: str,
                            llm: LLMJudge, llm_allowed: bool,
                            candidate_labels: Optional[List[str]] = None,
                            ) -> TaskScore:
    task = str(record["task_name"])
    gold = record.get("canonical_answer") or record.get("answer", "")
    gold_norm = canonicalize_label(str(gold))

    ts = TaskScore(
        pair_id=record.get("pair_id"),
        task_name=task,
        prediction=pred_text,
        answer=str(record.get("answer", "")),
        canonical_answer=record.get("canonical_answer"),
        metric_type="source_label_match",
    )

    if not gold_norm:
        ts.parse_status = "fail_no_answer_meta"
        return ts
    if not str(pred_text).strip():
        ts.parse_status = "fail_empty"
        return ts

    pred_norm = canonicalize_label(pred_text)
    # Stage 1: direct canonical label match.
    if pred_norm == gold_norm:
        ts.correct = 1.0
        ts.details = {"pred_label": pred_norm, "gold_label": gold_norm, "match_stage": "canonical"}
        return ts

    # Stage 2: substring match (either direction).
    if gold_norm and gold_norm in pred_norm:
        ts.correct = 1.0
        ts.details = {"pred_label": pred_norm, "gold_label": gold_norm, "match_stage": "substring"}
        return ts

    # Stage 3: stopword-stripped match.
    pred_s = strip_stopwords(pred_text)
    gold_s = strip_stopwords(str(gold))
    if pred_s and gold_s and (pred_s == gold_s or gold_s in pred_s):
        ts.correct = 1.0
        ts.details = {"pred_label": pred_s, "gold_label": gold_s, "match_stage": "stopword_stripped"}
        return ts

    # Stage 4: LLM extractor — boil verbose prediction to a single label,
    # then compare normalized.
    details: Dict[str, Any] = {"pred_label": pred_norm, "gold_label": gold_norm, "match_stage": "none"}
    if llm_allowed and llm.cfg.enabled:
        extracted = llm.extract_label(pred_text, candidates=candidate_labels)
        if extracted is not None:
            ts.llm_used = True
            extracted_norm = canonicalize_label(extracted)
            details["llm_extracted"] = extracted_norm
            if extracted_norm == gold_norm:
                ts.correct = 1.0
                details["match_stage"] = "llm_extract"
                ts.details = details
                return ts
            # Stage 5: LLM judge semantic equivalence.
            verdict = llm.judge_equivalent(extracted_norm or pred_norm, gold_norm, task)
            if verdict is True:
                ts.correct = 1.0
                details["match_stage"] = "llm_judge"
                ts.details = details
                return ts
            if verdict is None and extracted is None:
                ts.parse_status = "fail_llm_extract"
    # Fell through all stages.
    ts.details = details
    return ts


def score_detect_time(record: Dict[str, Any], pred_text: str,
                        iou_threshold: float) -> TaskScore:
    task = str(record["task_name"])
    meta = record.get("answer_meta") or {}
    refs = record.get("source_refs") or []

    # Pick the gold span: prefer answer_meta.time_span / start_time+end_time,
    # else source_refs[0], else parse the answer text.
    gold_span: Optional[Tuple[float, float]] = None
    span_meta = meta.get("time_span")
    if isinstance(span_meta, (list, tuple)) and len(span_meta) == 2:
        gold_span = (float(span_meta[0]), float(span_meta[1]))
    elif meta.get("start_time") is not None and meta.get("end_time") is not None:
        gold_span = (float(meta["start_time"]), float(meta["end_time"]))
    elif refs and isinstance(refs[0], dict):
        r0 = refs[0]
        if r0.get("start_time") is not None and r0.get("end_time") is not None:
            gold_span = (float(r0["start_time"]), float(r0["end_time"]))
    if gold_span is None:
        gold_span = parse_time_span_first(record.get("answer", ""))

    ts = TaskScore(
        pair_id=record.get("pair_id"),
        task_name=task,
        prediction=pred_text,
        answer=str(record.get("answer", "")),
        canonical_answer=record.get("canonical_answer"),
        metric_type="time_span_iou",
    )

    if gold_span is None:
        ts.parse_status = "fail_no_answer_meta"
        return ts
    if not str(pred_text).strip():
        ts.parse_status = "fail_empty"
        return ts
    pred_span = parse_time_span_first(pred_text)
    if pred_span is None:
        ts.parse_status = "fail_regex"
        return ts
    ps, pe = pred_span
    gs, ge = gold_span
    iou = interval_iou(ps, pe, gs, ge)
    ts.details = {
        "predicted_span": [ps, pe],
        "target_span": [gs, ge],
        "iou": iou,
        "start_error_s": abs(ps - gs),
        "end_error_s": abs(pe - ge),
        "iou_threshold": iou_threshold,
    }
    # Use IoU directly as soft-correct (0..1), plus a binary @threshold.
    ts.correct = float(iou)
    ts.details["correct_binary"] = int(iou >= iou_threshold)
    return ts


def score_detect_source(record: Dict[str, Any], pred_text: str,
                          llm: LLMJudge, llm_allowed: bool,
                          iou_threshold: float) -> TaskScore:
    """Detect-source: list of (label, start, end). Score with event-level F1
    under (label_match AND iou>=thr). Label match uses canonical normalization;
    optional LLM synonym matching on a label-by-label basis (disabled by default
    because of API cost on long lists — enable via --llm-judge-all-tasks).
    """
    task = str(record["task_name"])
    refs = record.get("source_refs") or []
    gold_events: List[Tuple[str, float, float]] = []
    for r in refs:
        if not isinstance(r, dict):
            continue
        label = canonicalize_label(str(r.get("class_name") or ""))
        if label and r.get("start_time") is not None and r.get("end_time") is not None:
            gold_events.append((label, float(r["start_time"]), float(r["end_time"])))
    if not gold_events:
        # Fallback: parse the answer text.
        gold_events = parse_all_events(record.get("answer", ""))

    ts = TaskScore(
        pair_id=record.get("pair_id"),
        task_name=task,
        prediction=pred_text,
        answer=str(record.get("answer", "")),
        canonical_answer=record.get("canonical_answer"),
        metric_type="detect_source_f1",
    )

    if not gold_events:
        ts.parse_status = "fail_no_answer_meta"
        return ts
    if not str(pred_text).strip():
        ts.parse_status = "fail_empty"
        return ts
    pred_events = parse_all_events(pred_text)
    if not pred_events:
        ts.parse_status = "fail_regex"
        return ts

    # Greedy matching: for each gold event, find the highest-IoU pred event
    # whose label matches.
    matched_pred = [False] * len(pred_events)
    tp = 0
    ious: List[float] = []
    for (gl, gs, ge) in gold_events:
        best_idx = -1
        best_iou = 0.0
        for i, (pl, ps, pe) in enumerate(pred_events):
            if matched_pred[i]:
                continue
            label_ok = (pl == gl)
            if not label_ok and llm_allowed and llm.cfg.enabled:
                verdict = llm.judge_equivalent(pl, gl, task)
                if verdict is True:
                    label_ok = True
                    ts.llm_used = True
            if not label_ok:
                continue
            iou = interval_iou(ps, pe, gs, ge)
            if iou > best_iou:
                best_iou = iou
                best_idx = i
        if best_idx >= 0 and best_iou >= iou_threshold:
            matched_pred[best_idx] = True
            tp += 1
            ious.append(best_iou)
    n_gold = len(gold_events)
    n_pred = len(pred_events)
    precision = tp / max(n_pred, 1)
    recall = tp / max(n_gold, 1)
    f1 = 0.0 if precision + recall == 0 else 2 * precision * recall / (precision + recall)
    ts.details = {
        "n_gold_events": n_gold,
        "n_pred_events": n_pred,
        "tp": tp,
        "precision": precision,
        "recall": recall,
        "f1": f1,
        "matched_iou_mean": mean_or_none(ious),
        "iou_threshold": iou_threshold,
    }
    ts.correct = float(f1)
    return ts


# ---------------------------------------------------------------------------
# Medium-split per-task scorers (count_sources, classify_motion,
# estimate_distance, onset_from_location)
# ---------------------------------------------------------------------------


def _parse_first_int(text: Any) -> Optional[int]:
    """Return the first integer found in text. Used by count_sources where
    the GT is a small integer ('1', '2', '3', ...) and the prediction may
    be embedded in a sentence ('There are 2 sources active.')."""
    m = re.search(r"-?\d+", str(text))
    if m is None:
        return None
    try:
        return int(m.group(0))
    except ValueError:
        return None


def score_count_sources(record: Dict[str, Any], pred_text: str) -> "TaskScore":
    """count_sources: GT is an integer (active_count). Just extract the
    first integer from prediction and compare. No LLM needed.
    """
    task = str(record["task_name"])
    meta = record.get("answer_meta") or {}
    target = meta.get("active_count")
    if target is None:
        target = _parse_first_int(record.get("answer", ""))

    ts = TaskScore(
        pair_id=record.get("pair_id"),
        task_name=task,
        prediction=pred_text,
        answer=str(record.get("answer", "")),
        canonical_answer=record.get("canonical_answer"),
        metric_type="count_exact_match",
    )

    if target is None:
        ts.parse_status = "fail_no_answer_meta"
        return ts
    if not str(pred_text).strip():
        ts.parse_status = "fail_empty"
        return ts
    pred = _parse_first_int(pred_text)
    if pred is None:
        ts.parse_status = "fail_regex"
        return ts
    err = abs(int(pred) - int(target))
    ts.details = {
        "predicted_count": int(pred),
        "target_count": int(target),
        "abs_error": int(err),
    }
    ts.correct = float(int(pred) == int(target))
    return ts


def score_estimate_distance(record: Dict[str, Any], pred_text: str,
                              rel_threshold: float = 0.3) -> "TaskScore":
    """estimate_distance: GT is a float (meters). The "correct" signal is
    the relative error: rel_err = |pred - gt| / max(|gt|, eps), with a
    threshold (default 0.3, i.e. within 30% of the true distance).

    Why relative instead of absolute: a 1m error is huge for a near-source
    (gt=1.5m -> 67% off) but acceptable for a far source (gt=8m -> 12.5%
    off). A relative threshold gives a fair signal across the [near, far]
    range.

    Aggregates report:
        - mean / median absolute error (m)
        - mean / median relative error
        - acc within `rel_threshold` (the binary "correct" signal)
        - extra acc points at relaxed/tight thresholds (rel<0.2 / rel<0.5)
    """
    task = str(record["task_name"])
    meta = record.get("answer_meta") or {}
    target = meta.get("distance_m")
    if target is None:
        target = parse_first_float(record.get("answer", ""))

    ts = TaskScore(
        pair_id=record.get("pair_id"),
        task_name=task,
        prediction=pred_text,
        answer=str(record.get("answer", "")),
        canonical_answer=record.get("canonical_answer"),
        metric_type=f"rel{rel_threshold:.2f}_distance",
    )

    if target is None:
        ts.parse_status = "fail_no_answer_meta"
        return ts
    if not str(pred_text).strip():
        ts.parse_status = "fail_empty"
        return ts
    pred = parse_first_float(pred_text)
    if pred is None:
        ts.parse_status = "fail_regex"
        return ts
    err = abs(float(pred) - float(target))
    rel_err = err / max(abs(float(target)), 1e-3)
    ts.details = {
        "predicted_m": float(pred),
        "target_m": float(target),
        "abs_error_m": float(err),
        "rel_error": float(rel_err),
        "rel_threshold": float(rel_threshold),
    }
    ts.correct = float(rel_err <= rel_threshold)
    return ts


def score_onset_from_location(record: Dict[str, Any], pred_text: str,
                                 within_s: float = 0.4) -> "TaskScore":
    """onset_from_location: GT is the onset time (seconds). Score reports
    the time error and acc within `within_s` (default 0.4s)."""
    task = str(record["task_name"])
    meta = record.get("answer_meta") or {}
    target = meta.get("onset_time")
    # Fallback: try to pull a "X.Xs" pattern from canonical/answer text.
    if target is None:
        target = parse_first_float(record.get("canonical_answer") or
                                    record.get("answer", ""))

    ts = TaskScore(
        pair_id=record.get("pair_id"),
        task_name=task,
        prediction=pred_text,
        answer=str(record.get("answer", "")),
        canonical_answer=record.get("canonical_answer"),
        metric_type=f"abs{within_s}s_onset",
    )

    if target is None:
        ts.parse_status = "fail_no_answer_meta"
        return ts
    if not str(pred_text).strip():
        ts.parse_status = "fail_empty"
        return ts
    pred = parse_first_float(pred_text)
    if pred is None:
        ts.parse_status = "fail_regex"
        return ts
    err = abs(float(pred) - float(target))
    ts.details = {
        "predicted_s": float(pred),
        "target_s": float(target),
        "abs_error_s": float(err),
        "within_s_threshold": float(within_s),
    }
    ts.correct = float(err <= within_s)
    return ts


# Default canonical motion labels. The QA generator uses 3 categories:
# stationary / moving / approaching|receding (some variants may say
# 'approaching' or 'moving towards'). Substring + LLM judge handle synonyms.
_MOTION_CANONICAL_LABELS = (
    "stationary", "moving", "approaching", "receding",
    "moving towards", "moving away",
)


def score_classify_motion(record: Dict[str, Any], pred_text: str,
                            llm: "LLMJudge") -> "TaskScore":
    """classify_motion: GT is a short label ('stationary' / 'moving' / ...).
    Strategy: substring match first (covers 90% of cases like
    'The laughter remains stationary throughout its duration.'),
    then fall back to LLM judge for ambiguous synonyms.
    """
    task = str(record["task_name"])
    meta = record.get("answer_meta") or {}
    gold = meta.get("motion_label") or record.get("canonical_answer") or \
        record.get("answer", "")
    gold_norm = canonicalize_label(str(gold))

    ts = TaskScore(
        pair_id=record.get("pair_id"),
        task_name=task,
        prediction=pred_text,
        answer=str(record.get("answer", "")),
        canonical_answer=record.get("canonical_answer"),
        metric_type="motion_label_match",
    )

    if not gold_norm:
        ts.parse_status = "fail_no_answer_meta"
        return ts
    if not str(pred_text).strip():
        ts.parse_status = "fail_empty"
        return ts

    pred_norm = canonicalize_label(pred_text)
    details: Dict[str, Any] = {
        "predicted_norm": pred_norm,
        "gold_norm": gold_norm,
    }

    # Stage 1: exact / substring match on canonicalized labels.
    if pred_norm == gold_norm:
        details["match_stage"] = "exact"
        ts.correct = 1.0
        ts.details = details
        return ts
    if gold_norm in pred_norm.split():
        details["match_stage"] = "word"
        ts.correct = 1.0
        ts.details = details
        return ts
    if gold_norm in pred_norm:
        details["match_stage"] = "substring"
        ts.correct = 1.0
        ts.details = details
        return ts

    # Stage 2: LLM judge. classify_motion has many phrasings ("staying still"
    # vs "stationary", "approaching" vs "moving towards the listener", etc.)
    # so we always allow the LLM to break ties when substring fails.
    if llm.cfg.enabled:
        verdict = llm.judge_equivalent(pred_norm, gold_norm, task)
        ts.llm_used = True
        if verdict is True:
            details["match_stage"] = "llm_judge"
            ts.correct = 1.0
        elif verdict is False:
            details["match_stage"] = "none"
            ts.correct = 0.0
        else:
            ts.parse_status = "fail_llm_extract"
            details["match_stage"] = "llm_failed"
        ts.details = details
        return ts

    # No LLM available -> mark as miss but keep parse_status=ok.
    details["match_stage"] = "none"
    ts.correct = 0.0
    ts.details = details
    return ts


# ---------------------------------------------------------------------------
# Dispatch + aggregate
# ---------------------------------------------------------------------------


def score_record(qa: Dict[str, Any], pred_text: str, llm: LLMJudge,
                   thresholds: Dict[str, float],
                   candidate_labels: Optional[List[str]],
                   ) -> TaskScore:
    task = str(qa.get("task_name") or "")
    if task == "estimate_azimuth":
        return score_estimate_angle(qa, pred_text, True,
                                      thresholds.get("azimuth_deg", thresholds["angle_deg"]))
    if task == "estimate_elevation":
        return score_estimate_angle(qa, pred_text, False,
                                      thresholds.get("elevation_deg", thresholds["angle_deg"]))
    if task in ("identify_source_by_doa", "identify_source_by_location"):
        return score_identify_source(qa, pred_text, llm,
                                       llm_allowed=True,
                                       candidate_labels=candidate_labels)
    if task == "detect_time":
        return score_detect_time(qa, pred_text, thresholds["iou"])
    if task == "detect_source":
        return score_detect_source(qa, pred_text, llm,
                                     llm_allowed=llm.cfg.judge_all_tasks,
                                     iou_threshold=thresholds["iou"])
    # Medium-split tasks.
    if task == "count_sources":
        return score_count_sources(qa, pred_text)
    if task == "estimate_distance":
        return score_estimate_distance(qa, pred_text,
                                          rel_threshold=thresholds.get("distance_rel", 0.3))
    if task == "onset_from_location":
        return score_onset_from_location(qa, pred_text,
                                            within_s=thresholds.get("onset_s", 0.4))
    if task == "classify_motion":
        # classify_motion always tries LLM-judge fallback for synonym
        # robustness, regardless of --llm-judge-all-tasks. The LLM call only
        # actually fires if --llm-judge is enabled (LLMJudge.cfg.enabled).
        return score_classify_motion(qa, pred_text, llm)
    # Unknown task -> fall back to normalized exact match.
    ts = TaskScore(
        pair_id=qa.get("pair_id"),
        task_name=task or "unknown",
        prediction=pred_text,
        answer=str(qa.get("answer", "")),
        canonical_answer=qa.get("canonical_answer"),
        metric_type="normalized_exact_match",
    )
    if not pred_text.strip():
        ts.parse_status = "fail_empty"
        return ts
    ts.correct = float(normalize_text(pred_text) == normalize_text(qa.get("answer", "")))
    return ts


def summarize(scored: List[TaskScore], thresholds: Dict[str, float],
                parse_status_order: List[str]) -> Dict[str, Any]:
    by_task: Dict[str, List[TaskScore]] = defaultdict(list)
    for s in scored:
        by_task[s.task_name].append(s)

    parse_ok = [s for s in scored if s.parse_status == "ok"]
    correct_all = mean_or_none([float(s.correct) for s in scored])
    correct_parseable = mean_or_none([float(s.correct) for s in parse_ok])

    summary: Dict[str, Any] = {
        "examples": len(scored),
        "overall_correct": correct_all,
        "overall_correct_parseable": correct_parseable,
        "parse_rate": mean_or_none([1.0 if s.parse_status == "ok" else 0.0 for s in scored]),
        "llm_used_rate": mean_or_none([1.0 if s.llm_used else 0.0 for s in scored]),
        "thresholds": thresholds,
        "parse_status_counts": {
            status: sum(1 for s in scored if s.parse_status == status)
            for status in parse_status_order
        },
        "per_task": {},
    }

    for task_name, records in sorted(by_task.items()):
        n = len(records)
        parse_ok_records = [r for r in records if r.parse_status == "ok"]
        per_task: Dict[str, Any] = {
            "examples": n,
            "metric_type": records[0].metric_type,
            "correct_all": mean_or_none([float(r.correct) for r in records]),
            "correct_parseable": mean_or_none([float(r.correct) for r in parse_ok_records]),
            "parse_rate": mean_or_none([1.0 if r.parse_status == "ok" else 0.0 for r in records]),
            "parse_status_counts": {
                status: sum(1 for r in records if r.parse_status == status)
                for status in parse_status_order
            },
            "llm_used_rate": mean_or_none([1.0 if r.llm_used else 0.0 for r in records]),
        }
        # Task-specific aggregates.
        if task_name in ("estimate_azimuth", "estimate_elevation"):
            errs = [float(r.details.get("error_deg"))
                    for r in parse_ok_records if "error_deg" in r.details]
            per_task["error_deg_mean"] = mean_or_none(errs)
            per_task["error_deg_median"] = median_or_none(errs)
            per_task["correct_is_within_threshold"] = per_task["correct_all"]
            # ------- collapse / template-answer diagnostics -------
            # Purpose: distinguish a model that genuinely localizes from one
            # that emits a constant / few-value template and rides the GT
            # distribution bias (elevation is centered near 0° -- a constant
            # 0° predictor already hits 33% acc@10° without reading the audio).
            preds = [float(r.details.get("predicted_deg"))
                     for r in parse_ok_records if "predicted_deg" in r.details]
            tgts = [float(r.details.get("target_deg"))
                    for r in parse_ok_records if "target_deg" in r.details]
            n_pred = len(preds)
            if n_pred > 0:
                # Output diversity: unique predictions / n. <0.1 means the
                # model is emitting a very small vocabulary of angles.
                uniq = len(set(round(p, 1) for p in preds))
                per_task["unique_prediction_ratio"] = uniq / n_pred
                per_task["unique_prediction_count"] = uniq
                # Top-1 prediction frequency: if one value dominates (>30%
                # for elevation, >15% for azimuth), the model is likely
                # template-answering.
                from collections import Counter
                pc = Counter(round(p, 1) for p in preds)
                top1_val, top1_cnt = pc.most_common(1)[0]
                per_task["top1_prediction"] = float(top1_val)
                per_task["top1_prediction_share"] = top1_cnt / n_pred
                # Constant-baseline comparison: how does the best single
                # constant prediction (median of GT) compare? This is the
                # "free floor" that any trivial model can hit. If the model's
                # acc / median_err is close to this floor, it hasn't learned
                # spatial localization.
                if tgts:
                    import statistics
                    gt_med = statistics.median(tgts)
                    if task_name == "estimate_azimuth":
                        # Wrap-around aware
                        def _err(p, g):
                            return abs(((p - g + 180.0) % 360.0) - 180.0)
                    else:
                        def _err(p, g):
                            return abs(p - g)
                    const_errs = [_err(gt_med, g) for g in tgts]
                    thr = (thresholds.get("azimuth_deg") if task_name == "estimate_azimuth"
                           else thresholds.get("elevation_deg")) or thresholds.get("angle_deg") or 20.0
                    per_task["const_median_baseline_value_deg"] = float(gt_med)
                    per_task["const_median_baseline_median_err"] = median_or_none(const_errs)
                    per_task["const_median_baseline_mean_err"] = mean_or_none(const_errs)
                    per_task["const_median_baseline_acc"] = mean_or_none(
                        [1.0 if e <= thr else 0.0 for e in const_errs])
                    # Gain vs constant baseline: positive = model actually
                    # learned; near-zero or negative = template collapse.
                    per_task["acc_gain_over_constant"] = (
                        (per_task["correct_all"] or 0.0)
                        - (per_task["const_median_baseline_acc"] or 0.0))
                    per_task["median_err_reduction_vs_constant"] = (
                        (per_task["const_median_baseline_median_err"] or 0.0)
                        - (per_task["error_deg_median"] or 0.0))
                # Stratified accuracy: bin predictions by GT angle and compute
                # acc@threshold within each bin, then macro-average. A model
                # that only predicts near-zero (common for elevation) will
                # pass easy bins but fail hard bins, and macro-avg penalizes
                # that bias.
                if task_name == "estimate_elevation":
                    bins = [(-90, -30), (-30, -10), (-10, 10), (10, 30), (30, 90)]
                else:  # estimate_azimuth
                    bins = [(-180, -90), (-90, -30), (-30, 30), (30, 90), (90, 180)]
                thr = (thresholds.get("azimuth_deg") if task_name == "estimate_azimuth"
                       else thresholds.get("elevation_deg")) or thresholds.get("angle_deg") or 20.0
                bin_accs: List[float] = []
                bin_counts: List[int] = []
                for lo, hi in bins:
                    mask_errs = []
                    for p, g in zip(preds, tgts):
                        if lo <= g < hi:
                            if task_name == "estimate_azimuth":
                                e = abs(((p - g + 180.0) % 360.0) - 180.0)
                            else:
                                e = abs(p - g)
                            mask_errs.append(1.0 if e <= thr else 0.0)
                    bin_counts.append(len(mask_errs))
                    if mask_errs:
                        bin_accs.append(sum(mask_errs) / len(mask_errs))
                if bin_accs:
                    per_task["acc_macro_by_gt_bin"] = sum(bin_accs) / len(bin_accs)
                    per_task["acc_per_gt_bin"] = {
                        f"[{lo},{hi})": {
                            "n": bin_counts[i],
                            "acc": (bin_accs[i] if i < len(bin_accs) else None),
                        }
                        for i, (lo, hi) in enumerate(bins)
                    }
        elif task_name == "detect_time":
            ious = [float(r.details.get("iou"))
                     for r in parse_ok_records if "iou" in r.details]
            starts = [float(r.details.get("start_error_s"))
                     for r in parse_ok_records if "start_error_s" in r.details]
            ends = [float(r.details.get("end_error_s"))
                     for r in parse_ok_records if "end_error_s" in r.details]
            per_task["iou_mean"] = mean_or_none(ious)
            per_task["iou_median"] = median_or_none(ious)
            per_task["start_error_mean_s"] = mean_or_none(starts)
            per_task["end_error_mean_s"] = mean_or_none(ends)
            per_task["iou_at_threshold"] = mean_or_none(
                [float(r.details.get("correct_binary") or 0)
                 for r in parse_ok_records])
        elif task_name == "detect_source":
            f1s = [float(r.details.get("f1"))
                    for r in parse_ok_records if "f1" in r.details]
            precs = [float(r.details.get("precision"))
                    for r in parse_ok_records if "precision" in r.details]
            recs = [float(r.details.get("recall"))
                    for r in parse_ok_records if "recall" in r.details]
            per_task["f1_mean"] = mean_or_none(f1s)
            per_task["precision_mean"] = mean_or_none(precs)
            per_task["recall_mean"] = mean_or_none(recs)
        elif task_name in ("identify_source_by_doa", "identify_source_by_location"):
            stages: Dict[str, int] = defaultdict(int)
            for r in records:
                stages[str(r.details.get("match_stage", "none"))] += 1
            per_task["match_stage_counts"] = dict(stages)
        elif task_name == "count_sources":
            errs = [float(r.details.get("abs_error"))
                    for r in parse_ok_records if "abs_error" in r.details]
            per_task["abs_error_mean"] = mean_or_none(errs)
            per_task["abs_error_median"] = median_or_none(errs)
            # acc within tolerance N for sanity
            for tol in (0, 1):
                per_task[f"acc_within_{tol}"] = mean_or_none(
                    [1.0 if e <= tol else 0.0 for e in errs])
        elif task_name == "estimate_distance":
            errs = [float(r.details.get("abs_error_m"))
                    for r in parse_ok_records if "abs_error_m" in r.details]
            rels = [float(r.details.get("rel_error"))
                    for r in parse_ok_records if "rel_error" in r.details]
            per_task["abs_error_m_mean"] = mean_or_none(errs)
            per_task["abs_error_m_median"] = median_or_none(errs)
            per_task["rel_error_mean"] = mean_or_none(rels)
            per_task["rel_error_median"] = median_or_none(rels)
            # The "correct" signal is rel_err <= 0.3 (the binary metric).
            per_task["acc_rel_within_0.3"] = per_task["correct_all"]
            # Extra context: tighter (0.2) and looser (0.5) thresholds.
            per_task["acc_rel_within_0.2"] = mean_or_none(
                [1.0 if r <= 0.2 else 0.0 for r in rels])
            per_task["acc_rel_within_0.5"] = mean_or_none(
                [1.0 if r <= 0.5 else 0.0 for r in rels])
            # Also keep the old absolute-distance acc as a sanity reference.
            per_task["acc_abs_within_1m"] = mean_or_none(
                [1.0 if e <= 1.0 else 0.0 for e in errs])
        elif task_name == "onset_from_location":
            errs = [float(r.details.get("abs_error_s"))
                    for r in parse_ok_records if "abs_error_s" in r.details]
            per_task["abs_error_s_mean"] = mean_or_none(errs)
            per_task["abs_error_s_median"] = median_or_none(errs)
            per_task["acc_within_0.4s"] = per_task["correct_all"]
            per_task["acc_within_0.2s"] = mean_or_none(
                [1.0 if e <= 0.2 else 0.0 for e in errs])
            per_task["acc_within_1.0s"] = mean_or_none(
                [1.0 if e <= 1.0 else 0.0 for e in errs])
        elif task_name == "classify_motion":
            stages: Dict[str, int] = defaultdict(int)
            for r in records:
                stages[str(r.details.get("match_stage", "none"))] += 1
            per_task["match_stage_counts"] = dict(stages)
        summary["per_task"][task_name] = per_task

    return summary


# ---------------------------------------------------------------------------
# I/O helpers
# ---------------------------------------------------------------------------


def load_jsonl(path: str) -> List[Dict[str, Any]]:
    with open(path, "r", encoding="utf-8") as handle:
        return [json.loads(line) for line in handle if line.strip()]


def resolve_qa_split(qa_root: str, split: str) -> str:
    for ext in (".jsonl", ".json"):
        p = os.path.join(qa_root, f"{split}{ext}")
        if os.path.exists(p):
            return p
    raise FileNotFoundError(f"Missing {split}.jsonl or {split}.json under {qa_root}")


def load_qa_split(qa_root: str, split: str) -> List[Dict[str, Any]]:
    path = resolve_qa_split(qa_root, split)
    if path.endswith(".jsonl"):
        return load_jsonl(path)
    with open(path, "r", encoding="utf-8") as handle:
        payload = json.load(handle)
    if isinstance(payload, list):
        return payload
    return payload.get("records") or payload.get("data") or []


def build_candidate_labels(qa_records: List[Dict[str, Any]]) -> List[str]:
    """Collect all canonical source labels seen in the split — used as
    hints for the LLM extractor."""
    labels: set = set()
    for r in qa_records:
        c = r.get("canonical_answer")
        if c:
            labels.add(canonicalize_label(str(c)))
        for ref in (r.get("source_refs") or []):
            if isinstance(ref, dict) and ref.get("class_name"):
                labels.add(canonicalize_label(str(ref["class_name"])))
    labels.discard("")
    return sorted(labels)


def clean_generated(text: str) -> str:
    """Trim the typical decoder tail so scorers see just the answer."""
    s = str(text).replace("\r\n", "\n").strip()
    for marker in ("Human:", "Question:", "\nHuman:", "\nQuestion:"):
        if marker in s:
            s = s.split(marker, 1)[0].strip()
    # Keep multi-line for detect_source / detect_time (they can have multiple
    # events). Only trim if the first non-empty line looks like a short
    # label-style answer.
    lines = [ln.strip() for ln in s.splitlines() if ln.strip()]
    if not lines:
        return ""
    # Heuristic: if first line is short and later lines start with "Answer:" /
    # "Explanation:", take only the first line. Otherwise keep all.
    if (len(lines) > 1 and len(lines[0]) <= 120
            and any(ln.lower().startswith(("explanation:", "reason:", "note:", "because")) for ln in lines[1:])):
        return lines[0]
    return "\n".join(lines).strip()


# ---------------------------------------------------------------------------
# CLI
# ---------------------------------------------------------------------------


def parse_args() -> argparse.Namespace:
    p = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
    p.add_argument("--predictions-jsonl", required=True,
                    help="Path to predictions.jsonl produced by the bench script.")
    p.add_argument("--qa-root", default=None,
                    help="QA root containing <split>.jsonl with the gold fields. "
                         "Optional: if omitted, scorer uses the `answer` / "
                         "`canonical_answer` / `answer_meta` / `source_refs` "
                         "fields embedded in predictions.jsonl directly "
                         "(self-scoring mode).")
    p.add_argument("--split", default="test")
    p.add_argument("--output-json", default=None,
                    help="Summary output path. Defaults to <predictions-dir>/score_result.json.")
    p.add_argument("--per-record-jsonl", default=None,
                    help="Optional path to dump per-record scoring detail.")
    p.add_argument("--angle-threshold-deg", type=float, default=20.0,
                    help="Default angle tolerance. Used only if the per-task "
                         "--azimuth/--elevation flags are left unset.")
    p.add_argument("--azimuth-threshold-deg", type=float, default=20.0,
                    help="estimate_azimuth tolerance (deg). Default: 20.")
    p.add_argument("--elevation-threshold-deg", type=float, default=10.0,
                    help="estimate_elevation tolerance (deg). Default: 10. "
                         "Elevation uses a tighter threshold because GT is "
                         "concentrated near 0° (a constant-0 predictor "
                         "already hits ~34%% at 10°; at 20° it would hit ~55%% "
                         "which obscures template-answer baselines).")
    p.add_argument("--iou-threshold", type=float, default=0.5,
                    help="IoU threshold used for detect_time/detect_source binary metrics.")
    p.add_argument("--distance-rel-threshold", type=float, default=0.3,
                    help="estimate_distance tolerance as a relative error "
                         "(|pred-gt|/|gt|). Default: 0.3 (i.e. correct if "
                         "predicted distance is within 30%% of the truth). "
                         "Relative error is fairer than absolute meters when "
                         "GT spans both near (<2m) and far (>5m) sources.")
    p.add_argument("--onset-threshold-s", type=float, default=0.4,
                    help="onset_from_location tolerance (seconds). Default: 0.4.")
    p.add_argument("--llm-judge", action="store_true",
                    help="Enable OpenAI-compatible LLM judge / extractor for "
                         "identify_source_* when regex match fails.")
    p.add_argument("--llm-judge-all-tasks", action="store_true",
                    help="Also use LLM for detect_source label synonym matching. Slow.")
    p.add_argument("--llm-model", default="gemini-3.1-pro-preview")
    p.add_argument("--llm-base-url", default="https://yunwu.ai/v1")
    p.add_argument("--llm-concurrency", type=int, default=4,
                    help="Number of parallel LLM calls. Use --llm-concurrency 1 for ordered debugging.")
    p.add_argument("--llm-max-calls", type=int, default=5000,
                    help="Hard cap on LLM calls (safety valve for cost).")
    p.add_argument("--keep-duplicate-pair-ids", action="store_true")
    return p.parse_args()


def main() -> int:
    args = parse_args()

    predictions = load_jsonl(args.predictions_jsonl)
    if args.qa_root:
        qa_records = load_qa_split(args.qa_root, args.split)
    else:
        # Self-scoring mode: synthesize qa_records from predictions themselves.
        # The `answer` field in preds IS the ground truth; the `canonical_answer`
        # / `answer_meta` / `source_refs` fields are absent only for legacy runs.
        print("[score] self-scoring mode (no --qa-root): using answer fields "
              "from predictions.jsonl as ground truth.")
        qa_records = [
            {
                "pair_id": r.get("pair_id"),
                "task_name": r.get("task_name"),
                "question": r.get("question"),
                "answer": r.get("answer"),
                "audio_path": r.get("audio_path"),
                "scene_id": r.get("scene_id"),
                "segment_stem": r.get("segment_stem"),
                "canonical_answer": r.get("canonical_answer"),
                "answer_meta": r.get("answer_meta"),
                "source_refs": r.get("source_refs"),
            }
            for r in predictions
        ]

    # Synthesize a stable pair_id for any record (pred or QA) whose source
    # split has pair_id=None. Must match the formula used by the bench
    # collator (scripts/batch_bench_spatial_beats_qa.py) so predictions and
    # QA align exactly.
    def _ensure_pair_id(r: Dict[str, Any]) -> None:
        pid = r.get("pair_id")
        if pid is None or pid == "":
            import hashlib
            key = "|".join(
                str(r.get(k, ""))
                for k in ("scene_id", "segment_stem", "task_name", "question", "audio_path")
            )
            r["pair_id"] = "auto_" + hashlib.sha1(key.encode("utf-8")).hexdigest()[:16]

    for r in qa_records:
        _ensure_pair_id(r)
    for r in predictions:
        _ensure_pair_id(r)

    # Build a join key. Prefer pair_id, but many generated QA splits have
    # pair_id = None for every record (the `easy_filtered` dataset is one of
    # them). In that case fall back to (question + task_name) which is
    # unique in practice (and further (question + task_name + audio_path)
    # if even that collides).
    def _primary_key(r):
        pid = r.get("pair_id")
        # Only trust pair_id if it's a "real" (non-synthetic) id. Synthetic
        # ids start with "auto_" and are only stable when both sides saw the
        # same audio_path/scene_id, which is not true for older bench runs
        # that dropped audio_path. Skip to richer fallback in that case.
        if pid is not None and not str(pid).startswith("auto_"):
            return ("id", str(pid))
        # Best: (task, question, audio_path) — uniquely identifies the sample.
        ap = r.get("audio_path")
        if ap:
            return ("tqa", str(r.get("task_name", "")), str(r.get("question", "")), str(ap))
        # Fallback for legacy predictions that lack audio_path: use the
        # generated answer string, which together with (task,question)
        # uniquely picks the right QA record in ~91% of easy_filtered test.
        return (
            "tqans",
            str(r.get("task_name", "")),
            str(r.get("question", "")),
            str(r.get("answer", "")),
        )

    def _fallback_key(r):
        return (
            "qta",
            str(r.get("question", "")),
            str(r.get("task_name", "")),
            str(r.get("audio_path", "")),
        )

    qa_index: Dict[Any, Dict[str, Any]] = {}
    qa_collisions = 0
    for r in qa_records:
        # Register the record under EVERY possible key that a prediction
        # might produce, so lookups succeed regardless of which fields the
        # prediction file carries (legacy runs dropped audio_path).
        keys = []
        pid = r.get("pair_id")
        if pid is not None and not str(pid).startswith("auto_"):
            keys.append(("id", str(pid)))
        ap = r.get("audio_path")
        if ap:
            keys.append(("tqa", str(r.get("task_name", "")),
                         str(r.get("question", "")), str(ap)))
        # Legacy key: (task, question, answer). Useful when preds lack audio_path.
        keys.append(("tqans", str(r.get("task_name", "")),
                     str(r.get("question", "")), str(r.get("answer", ""))))
        for k in keys:
            if k in qa_index:
                qa_collisions += 1
            else:
                qa_index[k] = r

    if qa_collisions:
        print(f"[score] WARN: {qa_collisions} QA records collided on "
              f"(question,task_name); falling back to (question,task_name,audio_path) "
              f"for those.")

    keyed_mode = "pair_id"
    if all(r.get("pair_id") is None for r in qa_records[:200]):
        keyed_mode = "(question, task_name)"
    print(f"[score] join key: {keyed_mode}; "
          f"qa_records={len(qa_records)} predictions={len(predictions)}")

    # Deduplicate predictions by join key (DistributedSampler padding emits dups).
    if not args.keep_duplicate_pair_ids:
        seen = set()
        deduped = []
        for rec in predictions:
            k = _primary_key(rec)
            if k in seen:
                continue
            seen.add(k)
            deduped.append(rec)
        predictions = deduped
        print(f"[score] after dedup: {len(predictions)} predictions")

    thresholds = {
        "angle_deg": args.angle_threshold_deg,
        "azimuth_deg": args.azimuth_threshold_deg,
        "elevation_deg": args.elevation_threshold_deg,
        "iou": args.iou_threshold,
        "distance_rel": args.distance_rel_threshold,
        "onset_s": args.onset_threshold_s,
    }

    llm_cfg = LLMConfig(
        enabled=bool(args.llm_judge or args.llm_judge_all_tasks),
        model=args.llm_model,
        base_url=args.llm_base_url,
        concurrency=max(1, args.llm_concurrency),
        judge_all_tasks=bool(args.llm_judge_all_tasks),
        judge_max_calls=args.llm_max_calls,
    )
    llm = LLMJudge(llm_cfg)
    candidate_labels = build_candidate_labels(qa_records) if llm_cfg.enabled else None

    # Score in parallel when LLM is on (LLM calls dominate cost); otherwise
    # score serially to keep logs readable.
    parse_status_order = ["ok", "fail_regex", "fail_llm_extract",
                           "fail_empty", "fail_no_answer_meta"]

    scored: List[TaskScore] = []
    unmatched = [0]  # counter, mutable via closure

    def _do_one(rec: Dict[str, Any]) -> Optional[TaskScore]:
        k = _primary_key(rec)
        qa = qa_index.get(k)
        if qa is None:
            # Try fallback key for collisions resolved above.
            qa = qa_index.get(_fallback_key(rec))
        if qa is None:
            # If prediction record itself already carries the answer, build a
            # lightweight qa dict from it — that way we still get a score even
            # on splits where no pair_id / question match is possible.
            if rec.get("answer") is not None and rec.get("task_name"):
                qa = {
                    "pair_id": rec.get("pair_id"),
                    "task_name": rec.get("task_name"),
                    "question": rec.get("question"),
                    "answer": rec.get("answer"),
                    "canonical_answer": rec.get("canonical_answer"),
                    "answer_meta": rec.get("answer_meta"),
                    "source_refs": rec.get("source_refs"),
                }
            else:
                unmatched[0] += 1
                return None
        raw_pred = rec.get("prediction_cleaned") or rec.get("prediction") or ""
        pred_text = clean_generated(raw_pred)
        ts = score_record(qa, pred_text, llm, thresholds, candidate_labels)
        return ts

    if llm_cfg.enabled and llm_cfg.concurrency > 1:
        with ThreadPoolExecutor(max_workers=llm_cfg.concurrency) as pool:
            futures = [pool.submit(_do_one, rec) for rec in predictions]
            for i, fut in enumerate(as_completed(futures)):
                ts = fut.result()
                if ts is not None:
                    scored.append(ts)
                if (i + 1) % 500 == 0:
                    print(f"  scored {i+1}/{len(predictions)}", flush=True)
    else:
        for i, rec in enumerate(predictions):
            ts = _do_one(rec)
            if ts is not None:
                scored.append(ts)
            if (i + 1) % 2000 == 0:
                print(f"  scored {i+1}/{len(predictions)}", flush=True)

    summary = summarize(scored, thresholds, parse_status_order)

    out_json = args.output_json or os.path.join(
        os.path.dirname(os.path.abspath(args.predictions_jsonl)), "score_result.json")
    os.makedirs(os.path.dirname(out_json), exist_ok=True)
    with open(out_json, "w", encoding="utf-8") as handle:
        json.dump(summary, handle, indent=2, ensure_ascii=False, sort_keys=True)
    print(f"\n[score] wrote {out_json}")

    if args.per_record_jsonl:
        with open(args.per_record_jsonl, "w", encoding="utf-8") as handle:
            for s in scored:
                handle.write(json.dumps({
                    "pair_id": s.pair_id,
                    "task_name": s.task_name,
                    "correct": s.correct,
                    "parse_status": s.parse_status,
                    "metric_type": s.metric_type,
                    "llm_used": s.llm_used,
                    "prediction": s.prediction,
                    "answer": s.answer,
                    "canonical_answer": s.canonical_answer,
                    "details": s.details,
                }, ensure_ascii=False) + "\n")
        print(f"[score] wrote per-record scoring to {args.per_record_jsonl}")

    # Print a compact tabular summary for humans.
    def _fmt(v, spec=".4f"):
        if v is None:
            return "n/a"
        try:
            return format(v, spec)
        except Exception:
            return str(v)

    print(f"\n=== overall ({summary['examples']} records) ===")
    if unmatched[0] > 0:
        print(f"  unmatched predictions: {unmatched[0]} (no QA join key)")
    print(f"  parse_rate               = {_fmt(summary['parse_rate'])}")
    print(f"  correct (all records)    = {_fmt(summary['overall_correct'])}")
    print(f"  correct (parseable only) = {_fmt(summary['overall_correct_parseable'])}")
    print(f"  llm_used_rate            = {_fmt(summary['llm_used_rate'])}")
    print("  parse_status_counts      = " + json.dumps(summary["parse_status_counts"]))

    for task_name, t in summary["per_task"].items():
        print(f"\n=== {task_name}  (n={t['examples']}, metric={t['metric_type']}) ===")
        print(f"  correct_all       = {_fmt(t['correct_all'])}")
        print(f"  correct_parseable = {_fmt(t.get('correct_parseable'))}")
        print(f"  parse_rate        = {_fmt(t['parse_rate'])}")
        print("  parse_status      = " + json.dumps(t["parse_status_counts"]))
        for key in ("error_deg_mean", "error_deg_median",
                     "iou_mean", "iou_median", "iou_at_threshold",
                     "start_error_mean_s", "end_error_mean_s",
                     "f1_mean", "precision_mean", "recall_mean",
                     "match_stage_counts",
                     # count_sources
                     "abs_error_mean", "abs_error_median",
                     "acc_within_0", "acc_within_1",
                     # estimate_distance
                     "abs_error_m_mean", "abs_error_m_median",
                     "rel_error_mean", "rel_error_median",
                     "acc_rel_within_0.2", "acc_rel_within_0.3",
                     "acc_rel_within_0.5", "acc_abs_within_1m",
                     # onset_from_location
                     "abs_error_s_mean", "abs_error_s_median",
                     "acc_within_0.2s", "acc_within_0.4s", "acc_within_1.0s",
                     # diagnostic for angle tasks
                     "const_median_baseline_acc", "acc_gain_over_constant",
                     "acc_macro_by_gt_bin", "unique_prediction_ratio",
                     "top1_prediction_share"):
            if key in t:
                val = t[key]
                if isinstance(val, float):
                    print(f"  {key:30s}= {_fmt(val)}")
                elif val is None:
                    print(f"  {key:30s}= n/a")
                else:
                    print(f"  {key:30s}= {val}")
    return 0


if __name__ == "__main__":
    sys.exit(main())