sv_gpu.py

"""
Score Vision SN44 — Unified miner v3.29 (2026-04-08). R9c vehicle FP16 (mAP50=0.929). Person: TTA consensus + 15% box shrink + NMS 0.35.
Dual-model: vehicle (YOLO11m INT8 1280) + person (YOLO12s FP16 960 TRT).
Pose model: YOLOv8n-pose FP16 640 for false-positive filtering + keypoint box refinement.
Vehicle weights loaded from secondary HF repo (meaculpitt/ScoreVision-Vehicle).
Person weights loaded from primary HF repo (template downloads automatically).

Vehicle model (vehicle_weights.onnx):
  Trained classes: 0=car, 1=bus, 2=truck, 3=motorcycle
  Output: 0=bus, 1=car, 2=truck, 3=motorcycle. All classes scored (v3.20 bus fix).
  Per-class confidence thresholds: car 0.45, truck 0.45, motorcycle 0.35.
  Per-class aspect ratio bounds for FP filtering.
  Single-pass (v3.19) — flip TTA removed for RTF improvement.

Person model (person_weights.onnx):
  YOLO12s FP16 960px end2end [1,300,6]. Single class: 0=person.
  Background TRT build: starts on CUDA immediately, builds TRT FP16 engine in background
  thread (~18min on fresh node), swaps to TRT atomically when ready. Cached thereafter.
  SAHI-style tiling: full + 2 adaptive tiles + flip TTA, max-conf NMS merge.

Pose model (pose_weights.onnx):
  YOLOv8n-pose FP16 640px [1,56,8400]. 17 COCO keypoints.
  Runs once on full image after person detection.
  Anatomical keypoint scoring: weighted per-keypoint sum (head 0.38, upper 0.32, lower 0.30).
  1. Head keypoints visible → never suppress, always refine box.
  2. Score >= 0.15 → keep + refine. Score > 0 → keep as-is. Score == 0 + large + low-conf → suppress.
  3. Box refinement: blend detected box with tight keypoint bbox for better fit.
  Face detector (optional): if face_session loaded, face inside box → never suppress.

Vehicle + person models run on every image when hint='both'. All detections merged.
Vehicle eval uses cls_id 1-3. Person eval uses cls_id 0 only.
"""

import os
import ctypes
import glob as _glob
import logging as _logging

_cuda_log = _logging.getLogger(__name__)

def _preload_cuda_libs():
    """Pre-load CUDA + TensorRT libs from pip packages so ORT GPU/TRT providers work.

    Search order for TRT libs (libnvinfer.so, libnvonnxparser.so):
      1. sys.path entries containing tensorrt_libs/ subdirectory
      2. site.getsitepackages() + user site-packages for tensorrt_libs/ or tensorrt/
      3. ctypes.util.find_library('nvinfer') as system-wide fallback
    If not found, logs clearly and skips TRT — never attempts pip operations.
    """
    try:
        import ctypes.util as _ctypes_util
        lib_dirs = []
        loaded = set()

        # ── CUDA libs from nvidia pip packages ──
        for mod_name in ['nvidia.cudnn', 'nvidia.cublas', 'nvidia.cuda_runtime',
                         'nvidia.cufft', 'nvidia.curand', 'nvidia.cusolver',
                         'nvidia.cusparse', 'nvidia.nvjitlink']:
            try:
                mod = __import__(mod_name, fromlist=['__file__'])
                lib_dir = os.path.join(os.path.dirname(mod.__file__), 'lib')
                if os.path.isdir(lib_dir) and lib_dir not in lib_dirs:
                    lib_dirs.append(lib_dir)
            except ImportError:
                pass

        # ── TensorRT libs — multi-strategy search ──
        import sys as _sys
        _trt_dir = None

        # Strategy 1: sys.path (covers standard pip installs)
        for p in _sys.path:
            for subdir in ('tensorrt_libs', 'tensorrt'):
                candidate = os.path.join(p, subdir)
                if os.path.isdir(candidate) and _glob.glob(os.path.join(candidate, 'libnvinfer*')):
                    _trt_dir = candidate
                    break
            if _trt_dir:
                break

        # Strategy 2: site-packages directories (covers user installs, venvs)
        if not _trt_dir:
            import site
            search_dirs = list(site.getsitepackages()) if hasattr(site, 'getsitepackages') else []
            user_site = getattr(site, 'getusersitepackages', lambda: None)()
            if user_site:
                search_dirs.append(user_site)
            # Also check common paths not always in site
            search_dirs.extend([
                '/usr/local/lib/python3.12/dist-packages',
                os.path.expanduser('~/.local/lib/python3.12/site-packages'),
                '/home/miner/.local/lib/python3.12/site-packages',
            ])
            for sp in search_dirs:
                for subdir in ('tensorrt_libs', 'tensorrt'):
                    candidate = os.path.join(sp, subdir)
                    if os.path.isdir(candidate) and _glob.glob(os.path.join(candidate, 'libnvinfer*')):
                        _trt_dir = candidate
                        break
                if _trt_dir:
                    break

        # Strategy 3: ctypes.util.find_library (system-wide LD search)
        if not _trt_dir:
            nvinfer_path = _ctypes_util.find_library('nvinfer')
            if nvinfer_path:
                _cuda_log.info('TRT found via system library: %s', nvinfer_path)
                try:
                    ctypes.CDLL(nvinfer_path, mode=ctypes.RTLD_GLOBAL)
                    loaded.add('nvinfer')
                except OSError as e:
                    _cuda_log.warning('Failed to load system nvinfer: %s', e)

        if _trt_dir:
            if _trt_dir not in lib_dirs:
                lib_dirs.append(_trt_dir)
            _cuda_log.info('TRT libs directory: %s', _trt_dir)
        elif 'nvinfer' not in loaded:
            _cuda_log.info('TensorRT libs not found — TRT EP will be unavailable (CUDA EP still works)')

        if not lib_dirs and not loaded:
            _cuda_log.warning('No CUDA or TRT libs found to preload')
            return

        # Set LD_LIBRARY_PATH for any child processes / dlopen fallbacks
        existing = os.environ.get('LD_LIBRARY_PATH', '')
        os.environ['LD_LIBRARY_PATH'] = ':'.join(lib_dirs + ([existing] if existing else []))

        # Load CUDA libs (glob all .so in nvidia dirs)
        for lib_dir in lib_dirs:
            if 'tensorrt' in lib_dir:
                continue  # TRT libs loaded selectively below
            for so in sorted(_glob.glob(os.path.join(lib_dir, 'lib*.so*'))):
                try:
                    ctypes.CDLL(so, mode=ctypes.RTLD_GLOBAL)
                except OSError:
                    pass

        # Load TRT libs selectively (only the essentials, not builder resources)
        if _trt_dir:
            for lib_name in ['libnvinfer.so', 'libnvinfer_plugin.so', 'libnvonnxparser.so']:
                matches = _glob.glob(os.path.join(_trt_dir, lib_name + '*'))
                if matches:
                    try:
                        ctypes.CDLL(matches[0], mode=ctypes.RTLD_GLOBAL)
                        loaded.add(lib_name.split('.')[0])
                    except OSError as e:
                        _cuda_log.warning('Failed to load %s: %s', lib_name, e)
                else:
                    _cuda_log.info('%s not found in %s', lib_name, _trt_dir)

        if loaded:
            _cuda_log.info('Preloaded libs: %s', ', '.join(sorted(loaded)))
    except Exception as e:
        _cuda_log.warning('CUDA/TRT preload error: %s', e)

_preload_cuda_libs()



from pathlib import Path
import math
import time
import logging

import cv2
import numpy as np
import onnxruntime as ort
from numpy import ndarray
from pydantic import BaseModel

import json
import threading
from datetime import datetime, timezone
from concurrent.futures import ThreadPoolExecutor, as_completed
import inspect

# ── Latency logger (per-request timing) ─────────────────────────────────
import logging as _lat_logging
_lat_logger = _lat_logging.getLogger("sv_latency")
_lat_logger.setLevel(_lat_logging.INFO)
_lat_logger.propagate = False
if not _lat_logger.handlers:
    try:
        import tempfile as _lat_tempfile
        # Try /home/miner first (Lium), fall back to /tmp (Chutes cloud)
        for _lat_path in ["/home/miner/latency.log", _lat_tempfile.gettempdir() + "/latency.log"]:
            try:
                _lat_fh = _lat_logging.FileHandler(_lat_path)
                _lat_fh.setFormatter(_lat_logging.Formatter(
                    "%(asctime)s.%(msecs)03d %(message)s", datefmt="%Y-%m-%d %H:%M:%S"))
                _lat_logger.addHandler(_lat_fh)
                break
            except (OSError, PermissionError):
                continue
    except Exception:
        pass  # No file logging — latency still logged via main logger

logger = logging.getLogger(__name__)

# ── Vehicle config ──────────────────────────────────────────────────────────
VEH_MODEL_TO_OUT: dict[int, int] = {0: 1, 1: 0, 2: 2, 3: 3}  # bus→0 (validator expects bus at idx 0)
VEH_SKIP_CLS = set()          # v3.20: bus now scored (cls_id=0). Element detection prevents collision.
VEH_NUM_CLASSES = 4
VEH_CONF_THRES = 0.30         # Low decode threshold for TTA (final filter is per-class)
VEH_TTA_CONF = 0.20           # TTA flip pass decode threshold
VEH_NMS_IOU = 0.50

# ── Per-class vehicle confidence thresholds (output cls_id) ────────────────
# Raising from uniform 0.35: reduces FP (avg 4.1 FFPI → target <2.0)
VEH_CLASS_CONF: dict[int, float] = {
    1: 0.60,    # car — raised from 0.50, most FP-prone class (75% of training data)
    2: 0.45,    # truck — keep
    3: 0.50,    # motorcycle — raised from 0.45, small targets prone to FP
    0: 0.45,    # bus — keep
}

# ── Per-class vehicle aspect ratio bounds (min_ratio, max_ratio) ───────────
# ratio = max(w,h) / min(w,h). Generous bounds to avoid suppressing valid detections.
VEH_CLASS_ASPECT: dict[int, float] = {
    1: 5.0,     # car — rarely > 5:1 from any angle
    2: 6.0,     # truck — can be elongated
    3: 4.5,     # motorcycle — compact, rarely very elongated
    0: 8.0,     # bus — elongated body
}

# ── Per-class minimum area (pixels) ───────────────────────────────────────
VEH_CLASS_MIN_AREA: dict[int, int] = {
    1: 196,     # car — 14x14 min
    2: 256,     # truck — 16x16 min (should be at least medium-sized)
    3: 100,     # motorcycle — 10x10 min (can be very small in distance)
    0: 400,     # bus — 20x20 min
}

# ── Vehicle box sanity filters (global fallbacks) ─────────────────────────
VEH_MIN_WH = 20  # was 8. Kills tiny horizon artifacts (confirmed: h<25 extras on block 7900800)
VEH_MIN_AREA = 100
VEH_MAX_ASPECT = 8.0
VEH_MAX_AREA_RATIO = 0.95
VEH_MAX_DET = 40

# ── Vehicle parts confirmation config ────────────────────────────────────
# Cross-validates vehicle detections using person detections, OpenCV analysis,
# and optional license plate detector. Small/distant vehicles exempt.
VEH_PARTS_ENABLED = True          # Master switch for parts confirmation
VEH_PARTS_SMALL_AREA = 0.004     # Below this area ratio: exempt from suppression
VEH_PARTS_FP_CONF = 0.50         # Below this conf + large + unconfirmed → suppress
VEH_PARTS_FP_CONF_STRICT = 0.55  # Stricter threshold when plate model loaded but no plate
VEH_PARTS_FP_AREA = 0.03         # Above this area ratio → eligible for FP suppression
# Confidence boosts for confirmed parts (additive)
VEH_PARTS_BOOST_DRIVER = 0.08    # Person in driver/passenger region
VEH_PARTS_BOOST_RIDER = 0.10     # Person on motorcycle (overlap + optional lean)
VEH_PARTS_BOOST_HL = 0.05        # Headlight pair detected
VEH_PARTS_BOOST_PLATE = 0.12     # License plate detected (Phase 2)
VEH_PARTS_BOOST_WINDOW = 0.06    # Bus window pattern on truck
# Headlight detection thresholds
VEH_PARTS_HL_MIN_PX = 60         # Min vehicle width (px) for headlight check
VEH_PARTS_HL_BRIGHT = 200        # Grayscale threshold for bright spots
VEH_PARTS_HL_MIN_BLOB = 15       # Min contour area for headlight candidate
# Window pattern detection (bus/coach)
VEH_PARTS_WINDOW_MIN_PX = 100    # Min vehicle width for window pattern check
VEH_PARTS_WINDOW_MIN_PEAKS = 3   # Min periodic edge peaks for window confirmation
# Motorcycle rider pose
VEH_PARTS_RIDER_LEAN_DEG = 15.0  # Min torso lean from vertical (degrees) for rider pose
# Plate detection thresholds
VEH_PARTS_PLATE_MIN_PX = 80   # plates visible at ~80px vehicle width (was 120)
VEH_PARTS_PLATE_CONF = 0.35      # Min plate detection confidence

# ── Person config (TTA consensus) ───────────────────────────────────────────
PER_CONF_LOW = 0.60  # Was 0.55. Raised 2026-04-05 to match top peer precision floor after
                     # observing the 3-way tied 52-box group (conf_min=0.585, composite=0.280) was
                     # beaten by top peer's 44-box response (conf_min=0.716, composite=0.377).
                     # 0.60 targets the precision/recall inflection point without the full 0.65+
                     # aggression that might cost recall on sparse scenes.
PER_CONF_HIGH = 0.58  # NOTE: dead code, not referenced anywhere. Kept for reference only.
PER_CONSENSUS_IOU = 0.50
PER_RTF_BUDGET = 8.0

# ── Person box sanity filters ──────────────────────────────────────────────
PER_MIN_WH = 8
PER_MIN_AREA = 14 * 14
PER_MAX_ASPECT = 6.0
PER_MAX_AREA_RATIO = 0.80

# ── Person tiling config (SAHI-inspired) ────────────────────────────────────
PER_TILE_OVERLAP = 0.20          # 20% overlap between tiles
PER_TILE_MIN_DIM_RATIO = 1.15   # tile when image dim > model_dim * this (~1104px for 960 model)
PER_TILE_CONF = 0.55            # raised from 0.40 to match PER_CONF_LOW
PER_NMS_IOU = 0.35              # NMS IoU for merging across passes — tightened to reduce FP duplicates
PER_MAX_DET = 100               # Loose safety ceiling ONLY — not a count cap. Strategy is confidence-floor:
                                # PER_CONF_LOW=0.60 is the real filter; any box above threshold passes.
                                # Raised from 50 after 2026-04-05 investigation: top peers emit 77+ boxes on
                                # crowd eval images, and the currently-running chute (rev 6b9d0d6) caps at 30
                                # which is demonstrably hitting mAP50 0.39 on person crowd blocks. 50 would
                                # still clip. 100 gives real headroom — only triggers on pathological runaway
                                # FP cases where NMS has already failed. Previous values (10 spec'd, 50 first
                                # fix) were too tight. See FAILURE_ANALYSIS.md (2026-04-05).

# ── TTA consensus thresholds (DMSC19-inspired graduated approach) ────────────
# Cross-view confirmation eliminates the soft-NMS confidence decay bug.
# Instead of concatenate+soft-NMS (which decayed confs below floor), we match
# boxes across original+flip views and apply graduated confidence thresholds.
PER_TTA_MATCH_IOU = 0.50        # IoU threshold for cross-view box matching
PER_TTA_CONF_BOTH = 0.50        # Confirmed by both views: lower threshold (high confidence)
PER_TTA_CONF_ORIG = 0.60        # Original-only: standard threshold (PER_CONF_LOW)
PER_TTA_CONF_FLIP = 0.75        # Flip-only: strict (flip-only detections are likely FP)

# ── Frame quality gating (Laplacian variance) ───────────────────────────────
PER_BLUR_THRESHOLD = 50.0       # Laplacian variance below this = severely blurry
PER_BLUR_CONF_PENALTY = 0.85    # multiply confs by this for blurry frames (reduce FP)

# ── Adaptive CLAHE config ───────────────────────────────────────────────────
PER_CLAHE_CLIP = 2.0            # mild CLAHE (was 12.0, too aggressive)
PER_CLAHE_CONTRAST_THRESH = 40.0  # only apply CLAHE when L-channel std < this

# ── Perspective scaling confidence penalty ─────────────────────────────────
PERSP_DEVIATION_THRESH = 3.0    # ratio >3x or <1/3x triggers penalty
PERSP_CONF_PENALTY = 0.85       # multiply conf by this for perspective violations
PERSP_MIN_DETECTIONS = 3        # need ≥3 detections to estimate model
PERSP_MIN_Y_SPREAD = 0.15       # min y-spread as fraction of image height

# ── Pose FP filter + box refinement config ──────────────────────────────────
POSE_CONF_THRESH = 0.25         # Minimum confidence for pose detection
POSE_NMS_IOU = 0.65             # NMS IoU threshold for pose detections
POSE_MATCH_IOU = 0.30           # IoU threshold to match pose to person box
POSE_KP_CONF = 0.3              # Keypoint visibility threshold
POSE_FP_MAX_CONF = 0.65         # Max conf below which unmatched large boxes are suppressed
POSE_FP_MIN_AREA = 0.04         # Min area ratio (of image) for FP suppression to apply
POSE_REFINE_BLEND = 0.25        # Blend factor for keypoint box refinement (0=original, 1=keypoint)
POSE_KP_PAD = 0.10              # Padding around keypoint tight bbox

# ── Anatomical keypoint scoring ─────────────────────────────────────────────
# COCO keypoints: 0=nose 1=l_eye 2=r_eye 3=l_ear 4=r_ear
#   5=l_shoulder 6=r_shoulder 7=l_elbow 8=r_elbow 9=l_wrist 10=r_wrist
#   11=l_hip 12=r_hip 13=l_knee 14=r_knee 15=l_ankle 16=r_ankle
POSE_HEAD_KP = [0, 1, 2, 3, 4]                  # nose + eyes + ears
POSE_UPPER_KP = [5, 6, 7, 8, 9, 10]             # shoulders + elbows + wrists
POSE_LOWER_KP = [11, 12, 13, 14, 15, 16]        # hips + knees + ankles
# Per-keypoint weights (head > upper > lower). Sum of all = 1.0.
POSE_KP_WEIGHTS = np.array([
    0.12,  # 0  nose        — strongest single indicator
    0.08,  # 1  left_eye
    0.08,  # 2  right_eye
    0.05,  # 3  left_ear
    0.05,  # 4  right_ear
    0.07,  # 5  left_shoulder
    0.07,  # 6  right_shoulder
    0.05,  # 7  left_elbow
    0.05,  # 8  right_elbow
    0.04,  # 9  left_wrist
    0.04,  # 10 right_wrist
    0.05,  # 11 left_hip
    0.05,  # 12 right_hip
    0.04,  # 13 left_knee
    0.04,  # 14 right_knee
    0.03,  # 15 left_ankle
    0.04,  # 16 right_ankle
], dtype=np.float32)  # sums to 1.0
POSE_ANAT_REFINE_THRESH = 0.15  # Score above which we refine box with keypoints
POSE_ANAT_SUPPRESS_THRESH = 0.0 # Score at or below which suppression is considered

# ── TensorRT engine cache config ────────────────────────────────────────────
TRT_CACHE_PATH = "/tmp/trt_engine_cache"
TRT_FP16 = True
TRT_WORKSPACE_GB = 4

# ── Shared ──────────────────────────────────────────────────────────────────
WBF_SKIP_THR = 0.0001

# ── Speed config ────────────────────────────────────────────────────────────
ENABLE_TTA = True
ENABLE_PARALLEL = True

# ── Secondary HF repo for vehicle weights ───────────────────────────────────
VEHICLE_HF_REPO = "meaculpitt/ScoreVision-Vehicle"



def _wbf_multi(boxes_list, scores_list, labels_list, iou_thr=0.55, skip_thr=0.0001):
    """Weighted Boxes Fusion (multi-class). Boxes in [0,1] normalized coords."""
    if not boxes_list:
        return np.empty((0, 4)), np.empty(0), np.empty(0)

    all_b, all_s, all_l = [], [], []
    for bx, sc, lb in zip(boxes_list, scores_list, labels_list):
        for i in range(len(bx)):
            if sc[i] < skip_thr:
                continue
            all_b.append(bx[i])
            all_s.append(sc[i])
            all_l.append(int(lb[i]))

    if not all_b:
        return np.empty((0, 4)), np.empty(0), np.empty(0)

    all_b = np.array(all_b)
    all_s = np.array(all_s)
    all_l = np.array(all_l, dtype=int)

    fused_b, fused_s, fused_l = [], [], []
    for cls in np.unique(all_l):
        m = all_l == cls
        cb, cs = all_b[m], all_s[m]
        order = cs.argsort()[::-1]
        cb, cs = cb[order], cs[order]

        clusters, cboxes = [], []
        for i in range(len(cb)):
            matched, best_iou = -1, iou_thr
            for ci, cbox in enumerate(cboxes):
                xx1 = max(cb[i, 0], cbox[0])
                yy1 = max(cb[i, 1], cbox[1])
                xx2 = min(cb[i, 2], cbox[2])
                yy2 = min(cb[i, 3], cbox[3])
                inter = max(0, xx2 - xx1) * max(0, yy2 - yy1)
                a1 = (cb[i, 2] - cb[i, 0]) * (cb[i, 3] - cb[i, 1])
                a2 = (cbox[2] - cbox[0]) * (cbox[3] - cbox[1])
                iou = inter / (a1 + a2 - inter + 1e-9)
                if iou > best_iou:
                    best_iou = iou
                    matched = ci
            if matched >= 0:
                clusters[matched].append(i)
                idxs = clusters[matched]
                w = cs[idxs]
                cboxes[matched] = (cb[idxs] * w[:, None]).sum(0) / w.sum()
            else:
                clusters.append([i])
                cboxes.append(cb[i].copy())

        for ci, idxs in enumerate(clusters):
            fused_b.append(cboxes[ci])
            fused_s.append(cs[idxs].mean())
            fused_l.append(cls)

    if not fused_b:
        return np.empty((0, 4)), np.empty(0), np.empty(0)
    return np.array(fused_b), np.array(fused_s), np.array(fused_l)


def _wbf_single(boxes_list, scores_list, iou_thr=0.45, skip_thr=0.0001):
    """Weighted Boxes Fusion (single-class). Boxes in [0,1] normalized coords."""
    if not boxes_list:
        return np.empty((0, 4)), np.empty(0)

    all_b, all_s = [], []
    for bx, sc in zip(boxes_list, scores_list):
        for i in range(len(bx)):
            if sc[i] < skip_thr:
                continue
            all_b.append(bx[i])
            all_s.append(sc[i])

    if not all_b:
        return np.empty((0, 4)), np.empty(0)

    all_b = np.array(all_b)
    all_s = np.array(all_s)
    order = all_s.argsort()[::-1]
    all_b, all_s = all_b[order], all_s[order]

    clusters, cboxes = [], []
    for i in range(len(all_b)):
        matched, best_iou = -1, iou_thr
        for ci, cbox in enumerate(cboxes):
            xx1 = max(all_b[i, 0], cbox[0])
            yy1 = max(all_b[i, 1], cbox[1])
            xx2 = min(all_b[i, 2], cbox[2])
            yy2 = min(all_b[i, 3], cbox[3])
            inter = max(0, xx2 - xx1) * max(0, yy2 - yy1)
            a1 = (all_b[i, 2] - all_b[i, 0]) * (all_b[i, 3] - all_b[i, 1])
            a2 = (cbox[2] - cbox[0]) * (cbox[3] - cbox[1])
            iou = inter / (a1 + a2 - inter + 1e-9)
            if iou > best_iou:
                best_iou = iou
                matched = ci
        if matched >= 0:
            clusters[matched].append(i)
            idxs = clusters[matched]
            w = all_s[idxs]
            cboxes[matched] = (all_b[idxs] * w[:, None]).sum(0) / w.sum()
        else:
            clusters.append([i])
            cboxes.append(all_b[i].copy())

    fused_b, fused_s = [], []
    for ci, idxs in enumerate(clusters):
        fused_b.append(cboxes[ci])
        fused_s.append(all_s[idxs].mean())

    if not fused_b:
        return np.empty((0, 4)), np.empty(0)
    return np.array(fused_b), np.array(fused_s)


def _nms_per_class_boost(boxes, scores, labels, iou_thr=0.50):
    """Per-class hard NMS with max-score cluster boosting.
    Surviving box keeps its coordinates but gets the max confidence
    among all boxes in its overlap cluster."""
    if len(boxes) == 0:
        return np.empty((0, 4)), np.empty(0), np.empty(0, dtype=int)

    out_b, out_s, out_l = [], [], []
    for cls in np.unique(labels):
        m = labels == cls
        cb, cs = boxes[m], scores[m]
        order = cs.argsort()[::-1]
        cb, cs = cb[order], cs[order]

        suppressed = set()
        for i in range(len(cb)):
            if i in suppressed:
                continue
            max_score = float(cs[i])
            for j in range(i + 1, len(cb)):
                if j in suppressed:
                    continue
                xx1 = max(cb[i, 0], cb[j, 0])
                yy1 = max(cb[i, 1], cb[j, 1])
                xx2 = min(cb[i, 2], cb[j, 2])
                yy2 = min(cb[i, 3], cb[j, 3])
                inter = max(0, xx2 - xx1) * max(0, yy2 - yy1)
                a1 = (cb[i, 2] - cb[i, 0]) * (cb[i, 3] - cb[i, 1])
                a2 = (cb[j, 2] - cb[j, 0]) * (cb[j, 3] - cb[j, 1])
                iou = inter / (a1 + a2 - inter + 1e-9)
                if iou >= iou_thr:
                    max_score = max(max_score, float(cs[j]))
                    suppressed.add(j)
            out_b.append(cb[i])
            out_s.append(max_score)
            out_l.append(cls)

    if not out_b:
        return np.empty((0, 4)), np.empty(0), np.empty(0, dtype=int)
    return np.array(out_b), np.array(out_s), np.array(out_l, dtype=int)


class BoundingBox(BaseModel):
    x1: int
    y1: int
    x2: int
    y2: int
    cls_id: int
    conf: float


class TVFrameResult(BaseModel):
    frame_id: int
    boxes: list[BoundingBox]
    keypoints: list[tuple[int, int]]


class Miner:
    def __init__(self, path_hf_repo: Path) -> None:
        self.path_hf_repo = path_hf_repo

        # Vehicle model — download from secondary HF repo with safety guard
        t0 = time.monotonic()
        veh_path = None  # Path to secondary repo snapshot (also used for plate model)
        try:
            from huggingface_hub import snapshot_download as _sd
            veh_path = Path(_sd(VEHICLE_HF_REPO))
            veh_weights = str(veh_path / "vehicle_weights.onnx")
            logger.info(f"[init] Vehicle weights from {VEHICLE_HF_REPO} in {time.monotonic()-t0:.1f}s")
        except Exception as e:
            # Fallback: try loading from primary repo (backward compat)
            logger.warning(f"[init] Vehicle secondary repo failed ({e}), trying primary repo")
            veh_weights = str(path_hf_repo / "vehicle_weights.onnx")
            if not Path(veh_weights).exists():
                raise FileNotFoundError(f"vehicle_weights.onnx not found in primary or secondary repo") from e

        self.veh_session = ort.InferenceSession(
            veh_weights,
            providers=["CUDAExecutionProvider", "CPUExecutionProvider"],
        )
        veh_actual = self.veh_session.get_providers()
        logger.warning(f"[init] Vehicle session ACTIVE providers: {veh_actual}")
        if "CUDAExecutionProvider" not in veh_actual:
            logger.error("[init] ⚠ VEHICLE IS ON CPU — CUDA EP NOT ACTIVE")
        self.veh_input_name = self.veh_session.get_inputs()[0].name
        veh_shape = self.veh_session.get_inputs()[0].shape
        self.veh_h = int(veh_shape[2])
        self.veh_w = int(veh_shape[3])

        # FP32 fallback — lazy-loaded on first trigger to save ~300MB VRAM at startup
        self.veh_session_fp32 = None
        self._veh_fp32_path = None
        try:
            veh_fp32 = str(veh_path / "vehicle_weights_fp32.onnx") if veh_path else None
            if veh_fp32 and Path(veh_fp32).exists():
                self._veh_fp32_path = veh_fp32
                logger.info("[init] Vehicle FP32 fallback available (lazy-load)")
            else:
                logger.info("[init] Vehicle FP32 fallback not available")
        except Exception as e:
            logger.warning(f"[init] Vehicle FP32 fallback path check failed: {e}")

        # Person model — CUDA immediately, TRT engine builds in background
        per_onnx = str(path_hf_repo / "person_weights.onnx")
        self.per_session = ort.InferenceSession(
            per_onnx,
            providers=["CUDAExecutionProvider", "CPUExecutionProvider"],
        )
        self.per_input_name = self.per_session.get_inputs()[0].name
        per_shape = self.per_session.get_inputs()[0].shape
        self.per_h = int(per_shape[2])
        self.per_w = int(per_shape[3])
        self._trt_ready = False
        logger.info("[init] Person model: CUDA (TRT build starting in background)")

        # Launch background TRT engine build
        os.makedirs(TRT_CACHE_PATH, exist_ok=True)
        threading.Thread(
            target=self._build_trt_engine,
            args=(per_onnx,),
            daemon=True,
            name="trt-builder",
        ).start()

        # Pose model — for FP filtering + box refinement
        pose_path = path_hf_repo / "pose_weights.onnx"
        if pose_path.exists():
            self.pose_session = ort.InferenceSession(
                str(pose_path),
                providers=["CUDAExecutionProvider", "CPUExecutionProvider"],
            )
            self.pose_input_name = self.pose_session.get_inputs()[0].name
            pose_shape = self.pose_session.get_inputs()[0].shape
            self.pose_h = int(pose_shape[2])
            self.pose_w = int(pose_shape[3])
            logger.info(f"[init] Pose model loaded: {self.pose_h}x{self.pose_w}")
        else:
            self.pose_session = None
            logger.info("[init] No pose model found, FP filter disabled")

        # Face detector (SCRFD-500M) — confirms person boxes, prevents FP suppression
        face_path = path_hf_repo / "face_weights.onnx"
        if face_path.exists():
            self.face_session = ort.InferenceSession(
                str(face_path),
                providers=["CUDAExecutionProvider", "CPUExecutionProvider"],
            )
            self.face_input_name = self.face_session.get_inputs()[0].name
            logger.info("[init] Face model (SCRFD-500M) loaded")
        else:
            self.face_session = None
            logger.info("[init] No face model found")

        # License plate detector — loaded from secondary HF repo alongside vehicle weights
        plate_path = veh_path / "plate_weights.onnx" if veh_path else None
        if plate_path and plate_path.exists():
            self.plate_session = ort.InferenceSession(
                str(plate_path),
                providers=["CUDAExecutionProvider", "CPUExecutionProvider"],
            )
            self.plate_input_name = self.plate_session.get_inputs()[0].name
            plate_shape = self.plate_session.get_inputs()[0].shape
            self.plate_h = int(plate_shape[2]) if isinstance(plate_shape[2], int) else 640
            self.plate_w = int(plate_shape[3]) if isinstance(plate_shape[3], int) else 640
            logger.info(f"[init] Plate model loaded: {self.plate_h}x{self.plate_w}")
        else:
            self.plate_session = None
            logger.info("[init] No plate model found, plate confirmation disabled")


        # Pose cache — populated by _pose_filter_refine, read by vehicle parts
        self._cached_pose_data = None

        # Thread pool for parallel inference
        self._executor = ThreadPoolExecutor(max_workers=2)

        # Log provider info
        veh_prov = self.veh_session.get_providers()
        per_prov = self.per_session.get_providers()
        logger.info(f"Vehicle ORT providers: {veh_prov}")
        logger.info(f"Person ORT providers: {per_prov} (TRT building in background)")
        logger.info(f"TTA={ENABLE_TTA} PARALLEL={ENABLE_PARALLEL}")

    def _build_trt_engine(self, per_onnx):
        """Build TRT FP16 engine in background, swap person session when ready.

        On fresh nodes: ~18 min to compile. Cached engine loads in <1s.
        During build, inference uses CUDAExecutionProvider (passes RTF at ~78ms).
        After build, atomically swaps to TRT session (~29ms pipeline).
        """
        try:
            trt_opts = {
                "trt_fp16_enable": str(TRT_FP16).lower(),
                "trt_max_workspace_size": str(TRT_WORKSPACE_GB << 30),
                "trt_engine_cache_enable": "true",
                "trt_engine_cache_path": TRT_CACHE_PATH,
            }
            t0 = time.monotonic()
            logger.info("[trt-build] Creating TRT session (may take ~18min on fresh node)...")
            trt_session = ort.InferenceSession(
                per_onnx,
                providers=[
                    ("TensorrtExecutionProvider", trt_opts),
                    "CUDAExecutionProvider",
                    "CPUExecutionProvider",
                ],
            )

            provs = trt_session.get_providers()
            if "TensorrtExecutionProvider" not in provs:
                logger.warning("[trt-build] TRT provider not active (%s), keeping CUDA", provs)
                return

            # Run dummy inference to fully materialize the engine
            inp_name = trt_session.get_inputs()[0].name
            inp_shape = trt_session.get_inputs()[0].shape
            dummy = np.zeros((1, 3, int(inp_shape[2]), int(inp_shape[3])), dtype=np.float32)
            trt_session.run(None, {inp_name: dummy})

            dt = time.monotonic() - t0
            logger.info("[trt-build] TRT engine ready in %.1fs — swapping person session", dt)

            # Atomic swap — Python GIL makes single attribute assignment safe.
            # Any in-flight inference holds a reference to the old session, which
            # stays alive until that inference completes.
            self.per_session = trt_session
            self._trt_ready = True

            logger.info("[trt-build] Person model now using TensorRT FP16")
        except Exception as e:
            logger.warning("[trt-build] TRT build failed (%s), keeping CUDA", e)

    def __repr__(self) -> str:
        trt_status = "TRT" if self._trt_ready else "CUDA (TRT building)"
        return f"Unified Miner v3.16 — person={trt_status}, background TRT engine build"

    # ── Vehicle preprocessing (letterbox) ───────────────────────────────────

    def _veh_letterbox(self, img):
        h, w = img.shape[:2]
        r = min(self.veh_h / h, self.veh_w / w)
        nw, nh = int(round(w * r)), int(round(h * r))
        img_r = cv2.resize(img, (nw, nh), interpolation=cv2.INTER_LINEAR)
        dw, dh = self.veh_w - nw, self.veh_h - nh
        pl, pt = dw // 2, dh // 2
        img_p = cv2.copyMakeBorder(
            img_r, pt, dh - pt, pl, dw - pl,
            cv2.BORDER_CONSTANT, value=(114, 114, 114),
        )
        return img_p, r, pl, pt

    def _veh_preprocess(self, image_bgr):
        img_p, ratio, pl, pt = self._veh_letterbox(image_bgr)
        rgb = cv2.cvtColor(img_p, cv2.COLOR_BGR2RGB)
        inp = rgb.astype(np.float32) / 255.0
        inp = np.ascontiguousarray(inp.transpose(2, 0, 1)[np.newaxis])
        return inp, ratio, pl, pt

    def _veh_decode(self, raw, ratio, pl, pt, ow, oh, conf_thresh):
        pred = raw[0]
        if pred.shape[0] < pred.shape[1]:
            pred = pred.T
        cls_scores = pred[:, 4:]
        cls_ids = np.argmax(cls_scores, axis=1)
        confs = np.max(cls_scores, axis=1)
        mask = confs >= conf_thresh
        if not mask.any():
            return np.empty((0, 4)), np.empty(0), np.empty(0, dtype=int)
        bx, confs, cls_ids = pred[mask, :4], confs[mask], cls_ids[mask]
        cx, cy, bw, bh = bx[:, 0], bx[:, 1], bx[:, 2], bx[:, 3]
        x1 = np.clip((cx - bw / 2 - pl) / ratio, 0, ow)
        y1 = np.clip((cy - bh / 2 - pt) / ratio, 0, oh)
        x2 = np.clip((cx + bw / 2 - pl) / ratio, 0, ow)
        y2 = np.clip((cy + bh / 2 - pt) / ratio, 0, oh)
        return np.stack([x1, y1, x2, y2], axis=1), confs, cls_ids

    def _veh_run_pass(self, image_bgr, conf_thresh, session=None):
        if session is None:
            session = self.veh_session
        oh, ow = image_bgr.shape[:2]
        inp, ratio, pl, pt = self._veh_preprocess(image_bgr)
        raw = session.run(None, {self.veh_input_name: inp})[0]
        return self._veh_decode(raw, ratio, pl, pt, ow, oh, conf_thresh)

    def _infer_vehicle_core(self, image_bgr, session=None):
        """Core vehicle detection pipeline. session param allows FP32 fallback."""
        oh, ow = image_bgr.shape[:2]

        # Primary pass
        boxes, confs, cls_ids = self._veh_run_pass(image_bgr, VEH_CONF_THRES, session)

        # Flip TTA pass — horizontal flip, mirror boxes back
        if ENABLE_TTA:
            flipped = cv2.flip(image_bgr, 1)
            f_boxes, f_confs, f_cls = self._veh_run_pass(flipped, VEH_TTA_CONF, session)
            if len(f_boxes) > 0:
                # Mirror x-coords: x1'=ow-x2, x2'=ow-x1
                f_boxes[:, 0], f_boxes[:, 2] = ow - f_boxes[:, 2], ow - f_boxes[:, 0]
                if len(boxes) > 0:
                    boxes = np.concatenate([boxes, f_boxes])
                    confs = np.concatenate([confs, f_confs])
                    cls_ids = np.concatenate([cls_ids, f_cls])
                else:
                    boxes, confs, cls_ids = f_boxes, f_confs, f_cls

        if len(boxes) == 0:
            return []

        # Remap model classes to output classes
        out_cls = np.array([VEH_MODEL_TO_OUT[int(c)] for c in cls_ids])

        # Per-class hard NMS with max-score cluster boosting
        boxes, confs, out_cls = _nms_per_class_boost(
            boxes, confs, out_cls, iou_thr=VEH_NMS_IOU)

        if len(boxes) == 0:
            return []

        # Per-class confidence filter + aspect ratio filter + bus suppression
        img_area = float(oh * ow)
        sane = []
        for i in range(len(boxes)):
            cls = int(out_cls[i])

            # Skip bus entirely (not scored by validator, just generates FP)
            if cls in VEH_SKIP_CLS:
                continue

            # Per-class confidence threshold
            min_conf = VEH_CLASS_CONF.get(cls, VEH_CONF_THRES)
            if confs[i] < min_conf:
                continue

            bw = boxes[i, 2] - boxes[i, 0]
            bh = boxes[i, 3] - boxes[i, 1]

            # Minimum dimension
            if bw < VEH_MIN_WH or bh < VEH_MIN_WH:
                continue

            area = bw * bh

            # Per-class minimum area
            min_area = VEH_CLASS_MIN_AREA.get(cls, VEH_MIN_AREA)
            if area < min_area:
                continue

            # Per-class aspect ratio filter
            aspect = max(bw, bh) / max(min(bw, bh), 1e-6)
            max_aspect = VEH_CLASS_ASPECT.get(cls, VEH_MAX_ASPECT)
            if aspect > max_aspect:
                continue

            # Max area ratio (covers entire image — likely FP)
            if area / img_area > VEH_MAX_AREA_RATIO:
                continue

            sane.append(i)

        if not sane:
            return []
        boxes, confs, out_cls = boxes[sane], confs[sane], out_cls[sane]

        # Limit max detections
        if len(boxes) > VEH_MAX_DET:
            top_k = np.argsort(confs)[::-1][:VEH_MAX_DET]
            boxes, confs, out_cls = boxes[top_k], confs[top_k], out_cls[top_k]

        out = []
        for i in range(len(boxes)):
            b = boxes[i]
            out.append(BoundingBox(
                x1=max(0, min(ow, math.floor(b[0]))),
                y1=max(0, min(oh, math.floor(b[1]))),
                x2=max(0, min(ow, math.ceil(b[2]))),
                y2=max(0, min(oh, math.ceil(b[3]))),
                cls_id=int(out_cls[i]),
                conf=max(0.0, min(1.0, float(confs[i]))),
            ))
        return out

    def _infer_vehicle(self, image_bgr):
        """Vehicle detection with FP32 fallback on catastrophic INT8 failure.

        Runs INT8 model first. If it returns 0 boxes (true catastrophic failure,
        see block 7905900), retries with FP32 model. Single-box results are
        kept as-is — likely real sparse scenes, not INT8 degradation.
        """
        if not hasattr(self, '_veh_providers_logged'):
            provs = self.veh_session.get_providers()
            logger.warning(f"[vehicle] First inference — active providers: {provs}")
            self._veh_providers_logged = True
        boxes = self._infer_vehicle_core(image_bgr, self.veh_session)

        if len(boxes) == 0 and (self.veh_session_fp32 or self._veh_fp32_path):
            # Lazy-load FP32 session on first trigger
            if self.veh_session_fp32 is None and self._veh_fp32_path:
                try:
                    self.veh_session_fp32 = ort.InferenceSession(
                        self._veh_fp32_path,
                        providers=["CUDAExecutionProvider", "CPUExecutionProvider"],
                    )
                    logger.info("[vehicle] FP32 fallback lazy-loaded")
                except Exception as e:
                    logger.warning(f"[vehicle] FP32 lazy-load failed: {e}")
                    self._veh_fp32_path = None
            if self.veh_session_fp32:
                boxes_fp32 = self._infer_vehicle_core(image_bgr, self.veh_session_fp32)
                if len(boxes_fp32) > len(boxes):
                    logger.warning(
                        f"[vehicle] INT8 degraded ({len(boxes)} boxes), "
                        f"FP32 fallback recovered ({len(boxes_fp32)} boxes)"
                    )
                    return boxes_fp32

        return boxes

    # ── Vehicle parts confirmation ───────────────────────────────────────

    @staticmethod
    def _veh_check_driver(vb, person_boxes):
        """Check if any person detection overlaps the driver/passenger region.

        Driver region: upper 55% height, center 70% width of vehicle box.
        A person's center inside this region → vehicle confirmed.
        """
        if not person_boxes:
            return False
        vw = vb.x2 - vb.x1
        vh = vb.y2 - vb.y1
        dr_x1 = vb.x1 + vw * 0.15
        dr_y1 = vb.y1
        dr_x2 = vb.x2 - vw * 0.15
        dr_y2 = vb.y1 + vh * 0.55
        for pb in person_boxes:
            pcx = (pb.x1 + pb.x2) / 2
            pcy = (pb.y1 + pb.y2) / 2
            if dr_x1 <= pcx <= dr_x2 and dr_y1 <= pcy <= dr_y2:
                return True
        return False

    def _veh_check_rider(self, moto_box, person_boxes):
        """Check if motorcycle has a rider, optionally with forward-lean pose.

        Returns (has_overlap, has_lean_pose).
        Uses cached pose keypoints from person pipeline to check torso angle.
        Motorcycle riders lean forward (torso > 15° from vertical).
        """
        if not person_boxes:
            return False, False
        mw = moto_box.x2 - moto_box.x1
        mh = moto_box.y2 - moto_box.y1
        mx = mw * 0.1
        my = mh * 0.1
        has_overlap = False
        for pb in person_boxes:
            pcx = (pb.x1 + pb.x2) / 2
            pcy = (pb.y1 + pb.y2) / 2
            if (moto_box.x1 - mx <= pcx <= moto_box.x2 + mx and
                    moto_box.y1 - my <= pcy <= moto_box.y2 + my):
                has_overlap = True
                break
        if not has_overlap:
            return False, False

        # Check forward-lean pose using cached pose data
        if self._cached_pose_data is None:
            return True, False
        pose_boxes, pose_kps = self._cached_pose_data
        if len(pose_boxes) == 0:
            return True, False

        for j in range(len(pose_boxes)):
            pb = pose_boxes[j]
            pcx = (pb[0] + pb[2]) / 2
            pcy = (pb[1] + pb[3]) / 2
            if not (moto_box.x1 - mx <= pcx <= moto_box.x2 + mx and
                    moto_box.y1 - my <= pcy <= moto_box.y2 + my):
                continue
            kps = pose_kps[j]
            # Need at least one shoulder + one hip visible
            l_sh, r_sh = kps[5], kps[6]
            l_hip, r_hip = kps[11], kps[12]
            sh_vis = [k[:2] for k in [l_sh, r_sh] if k[2] >= POSE_KP_CONF]
            hip_vis = [k[:2] for k in [l_hip, r_hip] if k[2] >= POSE_KP_CONF]
            if not sh_vis or not hip_vis:
                continue
            sh_mid = np.mean(sh_vis, axis=0)
            hip_mid = np.mean(hip_vis, axis=0)
            dx = sh_mid[0] - hip_mid[0]
            dy = hip_mid[1] - sh_mid[1]  # positive = shoulder above hip
            if dy <= 0:
                continue
            angle = math.degrees(math.atan2(abs(dx), dy))
            if angle >= VEH_PARTS_RIDER_LEAN_DEG:
                return True, True
        return True, False

    def _veh_check_headlights(self, vb, image_bgr):
        """Detect bright symmetric pair in lower portion of vehicle box.

        Requires two bright blobs at similar y, on opposite sides of center,
        with similar area. Only checks vehicles wider than VEH_PARTS_HL_MIN_PX.
        """
        bw = vb.x2 - vb.x1
        bh = vb.y2 - vb.y1
        if bw < VEH_PARTS_HL_MIN_PX or bh < 30:
            return False

        oh, ow = image_bgr.shape[:2]
        y1 = max(0, min(oh, int(vb.y1 + bh * 0.65)))
        y2 = max(0, min(oh, int(vb.y2)))
        x1 = max(0, min(ow, int(vb.x1)))
        x2 = max(0, min(ow, int(vb.x2)))
        if y2 - y1 < 5 or x2 - x1 < 10:
            return False

        roi = image_bgr[y1:y2, x1:x2]
        gray = cv2.cvtColor(roi, cv2.COLOR_BGR2GRAY)
        _, bright = cv2.threshold(gray, VEH_PARTS_HL_BRIGHT, 255, cv2.THRESH_BINARY)
        contours, _ = cv2.findContours(bright, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)

        blobs = []
        for c in contours:
            area = cv2.contourArea(c)
            if area < VEH_PARTS_HL_MIN_BLOB:
                continue
            M = cv2.moments(c)
            if M["m00"] < 1:
                continue
            blobs.append((M["m10"] / M["m00"], M["m01"] / M["m00"], area))

        if len(blobs) < 2:
            return False

        roi_mid = (x2 - x1) / 2.0
        roi_h = y2 - y1
        for i in range(len(blobs)):
            for j in range(i + 1, len(blobs)):
                b1, b2 = blobs[i], blobs[j]
                if abs(b1[1] - b2[1]) > roi_h * 0.4:
                    continue
                if max(b1[2], b2[2]) / max(min(b1[2], b2[2]), 1) > 3.0:
                    continue
                if (b1[0] - roi_mid) * (b2[0] - roi_mid) < 0:
                    return True
        return False

    def _veh_check_windows(self, vb, image_bgr):
        """Detect repeated window pattern (bus/coach signature) using vertical edge periodicity.

        Extracts middle horizontal band, applies vertical Sobel, projects vertically,
        and checks for 3+ regularly-spaced peaks (window frame edges).
        Only for large vehicles (truck cls_id=2).
        """
        bw = vb.x2 - vb.x1
        bh = vb.y2 - vb.y1
        if bw < VEH_PARTS_WINDOW_MIN_PX or bh < 40:
            return False

        oh, ow = image_bgr.shape[:2]
        # Middle 40% of height (window band on a bus/coach)
        y1 = max(0, min(oh, int(vb.y1 + bh * 0.30)))
        y2 = max(0, min(oh, int(vb.y1 + bh * 0.70)))
        x1 = max(0, min(ow, int(vb.x1)))
        x2 = max(0, min(ow, int(vb.x2)))
        if y2 - y1 < 10 or x2 - x1 < 30:
            return False

        roi = image_bgr[y1:y2, x1:x2]
        gray = cv2.cvtColor(roi, cv2.COLOR_BGR2GRAY)

        # Vertical edge detection (window frames are vertical edges)
        sobel_v = cv2.Sobel(gray, cv2.CV_64F, 1, 0, ksize=3)
        abs_sobel = np.abs(sobel_v)

        # Project vertically: mean per column
        projection = abs_sobel.mean(axis=0)
        if len(projection) < 10:
            return False

        # Smooth projection
        ks = max(3, int(len(projection) * 0.02) | 1)
        projection = np.convolve(projection, np.ones(ks) / ks, mode='same')

        # Find peaks above mean + 1 std
        thresh = projection.mean() + projection.std()
        peaks = []
        in_peak = False
        pk_start = 0
        for i in range(len(projection)):
            if projection[i] > thresh:
                if not in_peak:
                    pk_start = i
                    in_peak = True
            else:
                if in_peak:
                    peaks.append((pk_start + i) // 2)
                    in_peak = False
        if in_peak:
            peaks.append((pk_start + len(projection) - 1) // 2)

        if len(peaks) < VEH_PARTS_WINDOW_MIN_PEAKS:
            return False

        # Check regular spacing: gaps within 40% of median
        gaps = [peaks[i + 1] - peaks[i] for i in range(len(peaks) - 1)]
        if not gaps:
            return False
        med = sorted(gaps)[len(gaps) // 2]
        if med < 5:
            return False
        regular = sum(1 for g in gaps if abs(g - med) / max(med, 1) < 0.4)
        return regular >= len(gaps) * 0.6

    def _veh_check_plate(self, vb, image_bgr):
        """Run license plate detector on a vehicle crop. Returns True if plate found."""
        if self.plate_session is None:
            return False
        bw = vb.x2 - vb.x1
        if bw < VEH_PARTS_PLATE_MIN_PX:
            return False

        oh, ow = image_bgr.shape[:2]
        # Crop vehicle region with 5% padding
        pad_x = int(bw * 0.05)
        pad_y = int((vb.y2 - vb.y1) * 0.05)
        cx1 = max(0, int(vb.x1) - pad_x)
        cy1 = max(0, int(vb.y1) - pad_y)
        cx2 = min(ow, int(vb.x2) + pad_x)
        cy2 = min(oh, int(vb.y2) + pad_y)
        crop = image_bgr[cy1:cy2, cx1:cx2]
        if crop.size == 0:
            return False

        # Letterbox to plate model input
        ch, cw = crop.shape[:2]
        r = min(self.plate_h / ch, self.plate_w / cw)
        nw, nh = int(round(cw * r)), int(round(ch * r))
        img_r = cv2.resize(crop, (nw, nh), interpolation=cv2.INTER_LINEAR)
        dw, dh = self.plate_w - nw, self.plate_h - nh
        pl, pt = dw // 2, dh // 2
        img_p = cv2.copyMakeBorder(
            img_r, pt, dh - pt, pl, dw - pl,
            cv2.BORDER_CONSTANT, value=(114, 114, 114),
        )
        rgb = cv2.cvtColor(img_p, cv2.COLOR_BGR2RGB)
        inp = rgb.astype(np.float32) / 255.0
        inp = np.ascontiguousarray(inp.transpose(2, 0, 1)[np.newaxis])

        raw = self.plate_session.run(None, {self.plate_input_name: inp})[0]
        pred = raw[0] if raw.ndim == 3 else raw

        # Handle both [N,6] end2end (post-NMS) and [N, 5+nc] raw formats
        if pred.shape[0] < pred.shape[1]:
            pred = pred.T  # transpose [5+nc, N] -> [N, 5+nc]
        if pred.shape[1] < 5:
            return False
        # End2end post-NMS: few detections (< 500), col4=conf already final
        if pred.shape[0] < 500 and pred.shape[1] == 6:
            confs = pred[:, 4]
        elif pred.shape[1] == 5:
            confs = pred[:, 4]  # single objectness score
        else:
            # Raw: x,y,w,h,objectness,cls_scores... → conf = obj * max(cls)
            confs = pred[:, 4] * np.max(pred[:, 5:], axis=1)
        return bool((confs >= VEH_PARTS_PLATE_CONF).any())

    def _vehicle_parts_confirm(self, vehicle_boxes, person_boxes, image_bgr):
        """Parts-based confidence scoring for vehicle detections.

        Scoring hierarchy (confidence boosts are additive):
        1. License plate detected → +0.12 (strong, never suppress)
        2. Person (driver/rider) inside vehicle → +0.08-0.10
        3. Headlight pair detected → +0.05
        4. Bus window pattern on truck → +0.06
        5. No parts but small/distant or high-conf → keep original
        6. Large + low-conf + no parts → suppress as FP

        Small/distant vehicles (area < 0.4% of image) are always exempt.
        Bus (cls_id=4) suppressed in _infer_vehicle — window check applies to trucks.
        """
        if not vehicle_boxes or not VEH_PARTS_ENABLED:
            return vehicle_boxes

        oh, ow = image_bgr.shape[:2]
        img_area = float(oh * ow)
        has_plate_model = self.plate_session is not None
        # Skip plate checks on crowded scenes (aerial/drone, plates invisible)
        skip_plate = len(vehicle_boxes) > 20

        result = []
        n_driver = 0
        n_rider = 0
        n_rider_lean = 0
        n_headlight = 0
        n_window = 0
        n_plate = 0
        n_suppressed = 0

        for vb in vehicle_boxes:
            bw = vb.x2 - vb.x1
            bh = vb.y2 - vb.y1
            area_ratio = (bw * bh) / img_area

            # Small/distant: exempt from parts check
            if area_ratio < VEH_PARTS_SMALL_AREA:
                result.append(vb)
                continue

            boost = 0.0
            confirmed = False

            # Check 1: License plate (strongest signal)
            if has_plate_model and not skip_plate and bw >= VEH_PARTS_PLATE_MIN_PX:
                try:
                    if self._veh_check_plate(vb, image_bgr):
                        boost += VEH_PARTS_BOOST_PLATE
                        confirmed = True
                        n_plate += 1
                except Exception:
                    pass

            # Check 2: Driver/passenger inside car or truck
            if vb.cls_id in (1, 2):
                if self._veh_check_driver(vb, person_boxes):
                    boost += VEH_PARTS_BOOST_DRIVER
                    confirmed = True
                    n_driver += 1

            # Check 3: Motorcycle rider (overlap + optional lean pose)
            if vb.cls_id == 3:
                has_overlap, has_lean = self._veh_check_rider(vb, person_boxes)
                if has_overlap:
                    boost += VEH_PARTS_BOOST_RIDER
                    if has_lean:
                        boost += 0.05  # Extra for confirmed lean pose
                        n_rider_lean += 1
                    confirmed = True
                    n_rider += 1

            # Check 4: Headlight pair
            if bw >= VEH_PARTS_HL_MIN_PX:
                try:
                    if self._veh_check_headlights(vb, image_bgr):
                        boost += VEH_PARTS_BOOST_HL
                        confirmed = True
                        n_headlight += 1
                except Exception:
                    pass

            # Check 5: Window pattern (large trucks that might be buses)
            if vb.cls_id == 2 and bw >= VEH_PARTS_WINDOW_MIN_PX:
                try:
                    if self._veh_check_windows(vb, image_bgr):
                        boost += VEH_PARTS_BOOST_WINDOW
                        n_window += 1
                except Exception:
                    pass

            # Apply boost and decide
            new_conf = min(1.0, vb.conf + boost)

            if confirmed:
                result.append(BoundingBox(
                    x1=vb.x1, y1=vb.y1, x2=vb.x2, y2=vb.y2,
                    cls_id=vb.cls_id, conf=new_conf,
                ))
            elif area_ratio > VEH_PARTS_FP_AREA:
                # Large vehicle — use stricter threshold if plate model loaded
                fp_thresh = VEH_PARTS_FP_CONF_STRICT if (has_plate_model and not skip_plate) else VEH_PARTS_FP_CONF
                if vb.conf < fp_thresh:
                    n_suppressed += 1
                else:
                    result.append(vb)
            else:
                result.append(vb)

        if n_driver or n_rider or n_headlight or n_window or n_plate or n_suppressed:
            logger.info(f"[veh-parts] plate={n_plate} driver={n_driver} rider={n_rider}"
                        f"(lean={n_rider_lean}) hl={n_headlight} win={n_window} "
                        f"suppress={n_suppressed}, kept {len(result)}/{len(vehicle_boxes)}")
        return result

    # ── Person preprocessing (letterbox) ──────────────────────────────────

    def _per_letterbox(self, img):
        h, w = img.shape[:2]
        r = min(self.per_h / h, self.per_w / w)
        nw, nh = int(round(w * r)), int(round(h * r))
        interp = cv2.INTER_CUBIC if r > 1.0 else cv2.INTER_LINEAR
        img_r = cv2.resize(img, (nw, nh), interpolation=interp)
        dw, dh = self.per_w - nw, self.per_h - nh
        pl, pt = dw // 2, dh // 2
        img_p = cv2.copyMakeBorder(
            img_r, pt, dh - pt, pl, dw - pl,
            cv2.BORDER_CONSTANT, value=(114, 114, 114),
        )
        return img_p, r, pl, pt

    def _per_preprocess(self, image_bgr):
        img_p, ratio, pl, pt = self._per_letterbox(image_bgr)
        rgb = cv2.cvtColor(img_p, cv2.COLOR_BGR2RGB)
        inp = rgb.astype(np.float32) / 255.0
        inp = np.ascontiguousarray(inp.transpose(2, 0, 1)[np.newaxis])
        return inp, ratio, pl, pt

    def _per_enhance(self, img_bgr):
        """Adaptive CLAHE: only apply to low-contrast frames, mild clip=2.0."""
        lab = cv2.cvtColor(img_bgr, cv2.COLOR_BGR2LAB)
        l, a, b = cv2.split(lab)
        if float(l.std()) < PER_CLAHE_CONTRAST_THRESH:
            clahe = cv2.createCLAHE(clipLimit=PER_CLAHE_CLIP, tileGridSize=(8, 8))
            l = clahe.apply(l)
            return cv2.cvtColor(cv2.merge([l, a, b]), cv2.COLOR_LAB2BGR)
        return img_bgr  # skip CLAHE on normal-contrast images

    @staticmethod
    def _frame_blur_score(img_bgr):
        """Laplacian variance blur metric. Lower = blurrier."""
        gray = cv2.cvtColor(img_bgr, cv2.COLOR_BGR2GRAY)
        return cv2.Laplacian(gray, cv2.CV_64F).var()

    @staticmethod
    def _perspective_penalty(boxes, confs, image_h):
        """Apply confidence penalty to perspective-anomalous person detections.

        Model: expected_height(y) = alpha * (y_foot - y_vp), where y_vp = image_h / 3.
        Alpha is estimated from the median height/distance ratio across detections.
        Detections deviating >3x from expected get conf *= 0.85.
        Fails open (returns confs unchanged) when model can't be estimated.
        """
        n = len(boxes)
        if n < PERSP_MIN_DETECTIONS:
            return confs

        y_vp = image_h / 3.0
        y_feet = boxes[:, 3]
        heights = boxes[:, 3] - boxes[:, 1]

        valid = y_feet > (y_vp + 10)
        if valid.sum() < PERSP_MIN_DETECTIONS:
            return confs

        valid_y = y_feet[valid]
        valid_h = heights[valid]

        y_spread = (valid_y.max() - valid_y.min()) / image_h
        if y_spread < PERSP_MIN_Y_SPREAD:
            return confs

        alpha = float(np.median(valid_h / (valid_y - y_vp)))
        if alpha <= 0.01:
            return confs

        new_confs = confs.copy()
        for i in range(n):
            if y_feet[i] <= y_vp:
                continue
            expected_h = alpha * (y_feet[i] - y_vp)
            if expected_h <= 0:
                continue
            ratio = heights[i] / expected_h
            if ratio > PERSP_DEVIATION_THRESH or ratio < (1.0 / PERSP_DEVIATION_THRESH):
                new_confs[i] *= PERSP_CONF_PENALTY

        return new_confs

    def _per_decode(self, raw, ratio, pl, pt, oh, ow, conf_thresh):
        pred = raw[0]
        if pred.ndim != 2:
            return np.empty((0, 4)), np.empty(0)

        # Auto-detect output format
        if pred.shape[-1] == 6 and pred.shape[0] > pred.shape[1]:
            # YOLO26 end2end: [N, 6] = [x1, y1, x2, y2, conf, class_id]
            confs = pred[:, 4]
            keep = confs >= conf_thresh
            boxes, confs = pred[keep, :4], confs[keep]
            if len(boxes) == 0:
                return np.empty((0, 4)), np.empty(0)
            boxes[:, 0] = np.floor((boxes[:, 0] - pl) / ratio)
            boxes[:, 1] = np.floor((boxes[:, 1] - pt) / ratio)
            boxes[:, 2] = np.ceil((boxes[:, 2] - pl) / ratio)
            boxes[:, 3] = np.ceil((boxes[:, 3] - pt) / ratio)
            boxes = np.clip(boxes, 0, [[ow, oh, ow, oh]])
            return boxes, confs

        # YOLO11 raw format: [5+nc, N] or [N, 5+nc]
        if pred.shape[0] < pred.shape[1]:
            pred = pred.T
        if pred.shape[1] < 5:
            return np.empty((0, 4)), np.empty(0)
        cls_scores = pred[:, 4:]
        confs = np.max(cls_scores, axis=1)
        keep = confs >= conf_thresh
        boxes, confs = pred[keep, :4], confs[keep]
        if len(boxes) == 0:
            return np.empty((0, 4)), np.empty(0)
        cx, cy, bw, bh = boxes[:, 0], boxes[:, 1], boxes[:, 2], boxes[:, 3]
        x1 = np.clip(np.floor((cx - bw / 2 - pl) / ratio), 0, ow)
        y1 = np.clip(np.floor((cy - bh / 2 - pt) / ratio), 0, oh)
        x2 = np.clip(np.ceil((cx + bw / 2 - pl) / ratio), 0, ow)
        y2 = np.clip(np.ceil((cy + bh / 2 - pt) / ratio), 0, oh)
        return np.stack([x1, y1, x2, y2], axis=1), confs

    def _per_run_pass(self, image_bgr, conf_thresh):
        oh, ow = image_bgr.shape[:2]
        inp, ratio, pl, pt = self._per_preprocess(image_bgr)
        raw = self.per_session.run(None, {self.per_input_name: inp})[0]
        return self._per_decode(raw, ratio, pl, pt, oh, ow, conf_thresh)

    def _generate_tiles(self, h, w):
        """SAHI-inspired tile generation.

        Smart 2-tile split: horizontal for landscape, vertical for portrait.
        Edge-aware: for landscape, split in upper portion to avoid cutting
        through people standing in bottom third.
        Returns: [(x1,y1,x2,y2), ...] — always starts with full image.
        """
        tiles = [(0, 0, w, h)]  # full image always first

        # Only tile if image significantly exceeds model input
        if max(h, w) <= max(self.per_h, self.per_w) * PER_TILE_MIN_DIM_RATIO:
            return tiles

        overlap_px_x = int(w * PER_TILE_OVERLAP)
        overlap_px_y = int(h * PER_TILE_OVERLAP)

        if w >= h:
            # Landscape: 2 horizontal tiles (left + right)
            mid = w // 2
            tiles.append((0, 0, mid + overlap_px_x, h))
            tiles.append((mid - overlap_px_x, 0, w, h))
        else:
            # Portrait: 2 vertical tiles (top + bottom)
            # Edge-aware: bias split toward upper portion (people stand at bottom)
            mid = int(h * 0.45)  # split at 45% height, not 50%
            tiles.append((0, 0, w, mid + overlap_px_y))
            tiles.append((0, mid - overlap_px_y, w, h))

        return tiles

    def _per_run_tile(self, image_bgr, tile_region, conf_thresh):
        """Run person model on a tile crop, return boxes in original coords."""
        x1t, y1t, x2t, y2t = tile_region
        crop = image_bgr[y1t:y2t, x1t:x2t]
        boxes, confs = self._per_run_pass(crop, conf_thresh)
        if len(boxes) == 0:
            return np.empty((0, 4)), np.empty(0)
        # Shift back to original image coordinates
        boxes[:, 0] += x1t
        boxes[:, 1] += y1t
        boxes[:, 2] += x1t
        boxes[:, 3] += y1t
        return boxes, confs

    @staticmethod
    @staticmethod
    def _nms_max_conf(boxes, scores, iou_thr, sigma=0.5, min_conf=0.20):
        """Soft-NMS with Gaussian decay (replaces hard NMS).

        Instead of suppressing overlapping boxes entirely, decays their
        confidence: score_j *= exp(-(iou^2) / sigma). This preserves
        partially-occluded detections in crowds while still penalising
        duplicates. Boxes whose confidence decays below min_conf are
        removed.
        """
        if len(boxes) == 0:
            return np.empty((0, 4)), np.empty(0)

        b = boxes.copy().astype(np.float64)
        s = scores.copy().astype(np.float64)
        n = len(s)
        indices = list(range(n))

        for i in range(n):
            # Find current max-confidence box
            max_idx = i
            for j in range(i + 1, n):
                if s[indices[j]] > s[indices[max_idx]]:
                    max_idx = j
            # Swap to front
            indices[i], indices[max_idx] = indices[max_idx], indices[i]

            ix = indices[i]
            # Decay overlapping boxes
            for j in range(i + 1, n):
                jx = indices[j]
                xx1 = max(b[ix, 0], b[jx, 0])
                yy1 = max(b[ix, 1], b[jx, 1])
                xx2 = min(b[ix, 2], b[jx, 2])
                yy2 = min(b[ix, 3], b[jx, 3])
                inter = max(0.0, xx2 - xx1) * max(0.0, yy2 - yy1)
                a1 = (b[ix, 2] - b[ix, 0]) * (b[ix, 3] - b[ix, 1])
                a2 = (b[jx, 2] - b[jx, 0]) * (b[jx, 3] - b[jx, 1])
                iou = inter / (a1 + a2 - inter + 1e-9)
                if iou > 0:
                    s[jx] *= np.exp(-(iou * iou) / sigma)

        # Keep boxes above min_conf
        keep = [indices[i] for i in range(n) if s[indices[i]] >= min_conf]
        if not keep:
            return np.empty((0, 4)), np.empty(0)
        return b[keep], s[keep]

    # ── Pose FP filter + box refinement ──────────────────────────────────

    def _pose_run(self, image_bgr):
        """Run pose model on full image, return (boxes [N,4], confs [N], keypoints [N,17,3]) in original coords."""
        if self.pose_session is None:
            return np.empty((0, 4)), np.empty(0), np.empty((0, 17, 3))

        oh, ow = image_bgr.shape[:2]

        # Letterbox to pose model input size
        r = min(self.pose_h / oh, self.pose_w / ow)
        nw, nh = int(round(ow * r)), int(round(oh * r))
        img_r = cv2.resize(image_bgr, (nw, nh), interpolation=cv2.INTER_LINEAR)
        dw, dh = self.pose_w - nw, self.pose_h - nh
        pl, pt = dw // 2, dh // 2
        img_p = cv2.copyMakeBorder(
            img_r, pt, dh - pt, pl, dw - pl,
            cv2.BORDER_CONSTANT, value=(114, 114, 114),
        )

        rgb = cv2.cvtColor(img_p, cv2.COLOR_BGR2RGB)
        inp = rgb.astype(np.float32) / 255.0
        inp = np.ascontiguousarray(inp.transpose(2, 0, 1)[np.newaxis])

        raw = self.pose_session.run(None, {self.pose_input_name: inp})[0]

        # raw shape: [1, 56, 8400] -> transpose to [8400, 56]
        pred = raw[0] if raw.ndim == 3 else raw
        if pred.shape[0] < pred.shape[1]:
            pred = pred.T

        # Decode: cols 0-3=xywh, col 4=conf, cols 5-55=17*3 keypoints
        confs = pred[:, 4]
        keep = confs >= POSE_CONF_THRESH
        if not keep.any():
            return np.empty((0, 4)), np.empty(0), np.empty((0, 17, 3))

        pred = pred[keep]
        confs = pred[:, 4]

        # Convert xywh to x1y1x2y2 in original coords
        cx, cy, bw, bh = pred[:, 0], pred[:, 1], pred[:, 2], pred[:, 3]
        x1 = np.clip((cx - bw / 2 - pl) / r, 0, ow)
        y1 = np.clip((cy - bh / 2 - pt) / r, 0, oh)
        x2 = np.clip((cx + bw / 2 - pl) / r, 0, ow)
        y2 = np.clip((cy + bh / 2 - pt) / r, 0, oh)
        boxes = np.stack([x1, y1, x2, y2], axis=1)

        # Decode keypoints: [N, 51] -> [N, 17, 3]
        kp_raw = pred[:, 5:].reshape(-1, 17, 3).copy()
        kp_raw[:, :, 0] = (kp_raw[:, :, 0] - pl) / r  # x
        kp_raw[:, :, 1] = (kp_raw[:, :, 1] - pt) / r  # y
        kp_raw[:, :, 0] = np.clip(kp_raw[:, :, 0], 0, ow)
        kp_raw[:, :, 1] = np.clip(kp_raw[:, :, 1], 0, oh)

        # NMS on pose detections
        order = np.argsort(-confs)
        boxes = boxes[order]
        confs = confs[order]
        kp_raw = kp_raw[order]

        keep_idx = []
        suppressed = set()
        for i in range(len(boxes)):
            if i in suppressed:
                continue
            keep_idx.append(i)
            for j in range(i + 1, len(boxes)):
                if j in suppressed:
                    continue
                xx1 = max(boxes[i, 0], boxes[j, 0])
                yy1 = max(boxes[i, 1], boxes[j, 1])
                xx2 = min(boxes[i, 2], boxes[j, 2])
                yy2 = min(boxes[i, 3], boxes[j, 3])
                inter = max(0, xx2 - xx1) * max(0, yy2 - yy1)
                a1 = (boxes[i, 2] - boxes[i, 0]) * (boxes[i, 3] - boxes[i, 1])
                a2 = (boxes[j, 2] - boxes[j, 0]) * (boxes[j, 3] - boxes[j, 1])
                iou_val = inter / (a1 + a2 - inter + 1e-9)
                if iou_val >= POSE_NMS_IOU:
                    suppressed.add(j)

        if not keep_idx:
            return np.empty((0, 4)), np.empty(0), np.empty((0, 17, 3))
        keep_idx = np.array(keep_idx)
        return boxes[keep_idx], confs[keep_idx], kp_raw[keep_idx]

    _FACE_SIZE = 640
    _FACE_STRIDES = (8, 16, 32)
    _FACE_NUM_ANCHORS = 2
    _FACE_THRESH = 0.5
    _FACE_NMS_THRESH = 0.4

    def _face_run(self, image_bgr):
        """Run SCRFD-500M face detector. Returns (face_boxes [N,4], face_confs [N])."""
        if self.face_session is None:
            return np.empty((0, 4)), np.empty(0)

        oh, ow = image_bgr.shape[:2]
        sz = self._FACE_SIZE

        # Letterbox resize preserving aspect ratio (top-left aligned)
        scale = min(sz / oh, sz / ow)
        nw, nh = int(round(ow * scale)), int(round(oh * scale))
        resized = cv2.resize(image_bgr, (nw, nh), interpolation=cv2.INTER_LINEAR)
        det_img = np.zeros((sz, sz, 3), dtype=np.uint8)
        det_img[:nh, :nw, :] = resized

        # Preprocess: BGR→RGB, (pixel - 127.5) / 128.0
        blob = cv2.dnn.blobFromImage(
            det_img, 1.0 / 128.0, (sz, sz), (127.5, 127.5, 127.5), swapRB=True,
        )

        outputs = self.face_session.run(None, {self.face_input_name: blob})

        # Decode 3 stride levels: outputs[0:3]=scores, [3:6]=bboxes, [6:9]=kps
        all_scores, all_boxes = [], []
        for idx, stride in enumerate(self._FACE_STRIDES):
            scores = outputs[idx][:, 0]  # (N,)
            bbox_d = outputs[idx + 3]    # (N, 4) distances
            keep = scores >= self._FACE_THRESH
            if not keep.any():
                continue
            scores = scores[keep]
            bbox_d = bbox_d[keep]

            # Generate anchor centers for kept positions
            fh, fw = sz // stride, sz // stride
            grid_y, grid_x = np.mgrid[:fh, :fw]
            centers = np.stack([grid_x, grid_y], axis=-1).astype(np.float32).reshape(-1, 2)
            centers = np.tile(centers, (1, self._FACE_NUM_ANCHORS)).reshape(-1, 2) * stride
            centers = centers[keep]

            # distance → bbox: [x1, y1, x2, y2]
            x1 = centers[:, 0] - bbox_d[:, 0] * stride
            y1 = centers[:, 1] - bbox_d[:, 1] * stride
            x2 = centers[:, 0] + bbox_d[:, 2] * stride
            y2 = centers[:, 1] + bbox_d[:, 3] * stride
            boxes = np.stack([x1, y1, x2, y2], axis=-1) / scale

            all_scores.append(scores)
            all_boxes.append(boxes)

        if not all_scores:
            return np.empty((0, 4)), np.empty(0)

        scores = np.concatenate(all_scores)
        boxes = np.concatenate(all_boxes)

        # NMS
        order = scores.argsort()[::-1]
        scores, boxes = scores[order], boxes[order]
        keep = []
        x1, y1, x2, y2 = boxes[:, 0], boxes[:, 1], boxes[:, 2], boxes[:, 3]
        areas = (x2 - x1) * (y2 - y1)
        suppressed = np.zeros(len(scores), dtype=bool)
        for i in range(len(scores)):
            if suppressed[i]:
                continue
            keep.append(i)
            xx1 = np.maximum(x1[i], x1[i + 1:])
            yy1 = np.maximum(y1[i], y1[i + 1:])
            xx2 = np.minimum(x2[i], x2[i + 1:])
            yy2 = np.minimum(y2[i], y2[i + 1:])
            inter = np.maximum(0, xx2 - xx1) * np.maximum(0, yy2 - yy1)
            ovr = inter / (areas[i] + areas[i + 1:] - inter + 1e-6)
            suppressed[i + 1:] |= ovr > self._FACE_NMS_THRESH

        return boxes[keep], scores[keep]

    @staticmethod
    def _anatomical_score(kps, kp_conf_thresh=POSE_KP_CONF):
        """Compute weighted anatomical score from keypoints [17, 3].

        Returns (score, has_head, n_visible):
          score: weighted sum of visible keypoints (0.0-1.0)
          has_head: True if any head keypoint (nose/eyes/ears) is visible
          n_visible: number of visible keypoints
        """
        visible = kps[:, 2] >= kp_conf_thresh
        n_visible = int(visible.sum())
        score = float((visible.astype(np.float32) * POSE_KP_WEIGHTS).sum())
        has_head = bool(visible[POSE_HEAD_KP].any())
        return score, has_head, n_visible

    def _refine_box_with_keypoints(self, pb, kps, ow, oh):
        """Blend person box with tight keypoint bbox."""
        visible = kps[:, 2] >= POSE_KP_CONF
        if not visible.any():
            return pb
        vis_kps = kps[visible]
        kp_x1 = float(vis_kps[:, 0].min())
        kp_y1 = float(vis_kps[:, 1].min())
        kp_x2 = float(vis_kps[:, 0].max())
        kp_y2 = float(vis_kps[:, 1].max())

        # Pad around keypoint bbox
        kp_w = kp_x2 - kp_x1
        kp_h = kp_y2 - kp_y1
        pad_x = kp_w * POSE_KP_PAD
        pad_y = kp_h * POSE_KP_PAD
        kp_x1 = max(0, kp_x1 - pad_x)
        kp_y1 = max(0, kp_y1 - pad_y)
        kp_x2 = min(ow, kp_x2 + pad_x)
        kp_y2 = min(oh, kp_y2 + pad_y)

        a = POSE_REFINE_BLEND
        return BoundingBox(
            x1=max(0, min(ow, int(pb.x1 * (1 - a) + kp_x1 * a))),
            y1=max(0, min(oh, int(pb.y1 * (1 - a) + kp_y1 * a))),
            x2=max(0, min(ow, int(pb.x2 * (1 - a) + kp_x2 * a))),
            y2=max(0, min(oh, int(pb.y2 * (1 - a) + kp_y2 * a))),
            cls_id=0,
            conf=pb.conf,
        )

    def _pose_filter_refine(self, person_boxes, image_bgr):
        """Filter FP detections and refine boxes using anatomical keypoint scoring.

        Anatomical scoring: weighted sum of visible keypoints where head/face
        keypoints (nose, eyes, ears) contribute most, upper body (shoulders,
        elbows, wrists) next, lower body (hips, knees, ankles) least.

        Decision logic:
        1. Run pose model once on full image.
        2. Run face detector (if available) for additional confirmation.
        3. Match each person detection to best-overlapping pose detection.
        4. For matched boxes:
           a. Head keypoints visible OR face detected → KEEP + refine (never suppress)
           b. Anatomical score >= REFINE threshold → KEEP + refine
           c. Anatomical score > 0 → KEEP as-is (partially visible person)
           d. Anatomical score == 0 + large + low-conf → SUPPRESS (FP candidate)
        5. For unmatched boxes:
           a. Face detected inside box → KEEP
           b. Large + low-conf → SUPPRESS
           c. Small or high-conf → KEEP (SAHI-detected or confident)
        """
        if not person_boxes or self.pose_session is None:
            return person_boxes

        oh, ow = image_bgr.shape[:2]
        img_area = float(oh * ow)

        # Run pose model
        t_pose = time.monotonic()
        pose_boxes, pose_confs, pose_kps = self._pose_run(image_bgr)
        dt_pose = (time.monotonic() - t_pose) * 1000

        # Cache pose data for motorcycle rider check in vehicle parts confirmation
        self._cached_pose_data = (pose_boxes, pose_kps)

        # Run face detector if available
        face_boxes = np.empty((0, 4))
        if self.face_session is not None:
            t_face = time.monotonic()
            face_boxes, _ = self._face_run(image_bgr)
            dt_face = (time.monotonic() - t_face) * 1000
            logger.info(f"[pose] {len(pose_boxes)} pose, {len(face_boxes)} faces "
                        f"in {dt_pose:.0f}+{dt_face:.0f}ms")
        else:
            logger.info(f"[pose] {len(pose_boxes)} pose detections in {dt_pose:.0f}ms")

        # Helper: check if any face detection is inside a person box
        def has_face_inside(pb):
            if len(face_boxes) == 0:
                return False
            for fb in face_boxes:
                # Face center must be inside person box
                fcx = (fb[0] + fb[2]) / 2
                fcy = (fb[1] + fb[3]) / 2
                if pb.x1 <= fcx <= pb.x2 and pb.y1 <= fcy <= pb.y2:
                    return True
            return False

        if len(pose_boxes) == 0:
            # No pose detections — use face detector or size/conf heuristic
            result = []
            n_suppressed = 0
            for pb in person_boxes:
                if has_face_inside(pb):
                    result.append(pb)
                    continue
                bw = pb.x2 - pb.x1
                bh = pb.y2 - pb.y1
                area_ratio = (bw * bh) / img_area
                if area_ratio > POSE_FP_MIN_AREA and pb.conf < POSE_FP_MAX_CONF:
                    n_suppressed += 1
                    continue
                result.append(pb)
            if n_suppressed:
                logger.info(f"[pose] Suppressed {n_suppressed} FP (no pose detections)")
            return result

        # Match person detections to pose detections via IoU
        result = []
        n_refined = 0
        n_suppressed = 0
        n_face_saved = 0

        for pb in person_boxes:
            pb_arr = np.array([pb.x1, pb.y1, pb.x2, pb.y2], dtype=float)
            best_iou = 0.0
            best_idx = -1

            for j in range(len(pose_boxes)):
                xx1 = max(pb_arr[0], pose_boxes[j, 0])
                yy1 = max(pb_arr[1], pose_boxes[j, 1])
                xx2 = min(pb_arr[2], pose_boxes[j, 2])
                yy2 = min(pb_arr[3], pose_boxes[j, 3])
                inter = max(0, xx2 - xx1) * max(0, yy2 - yy1)
                a1 = (pb_arr[2] - pb_arr[0]) * (pb_arr[3] - pb_arr[1])
                a2 = (pose_boxes[j, 2] - pose_boxes[j, 0]) * (pose_boxes[j, 3] - pose_boxes[j, 1])
                iou_val = inter / (a1 + a2 - inter + 1e-9)
                if iou_val > best_iou:
                    best_iou = iou_val
                    best_idx = j

            if best_iou >= POSE_MATCH_IOU and best_idx >= 0:
                # Matched to a pose detection — compute anatomical score
                kps = pose_kps[best_idx]  # [17, 3]
                anat_score, has_head, n_vis = self._anatomical_score(kps)

                if has_head or has_face_inside(pb):
                    # Head/face visible → definitely a person, refine box
                    result.append(self._refine_box_with_keypoints(pb, kps, ow, oh))
                    n_refined += 1
                elif anat_score >= POSE_ANAT_REFINE_THRESH:
                    # Good anatomical score → person confirmed, refine
                    result.append(self._refine_box_with_keypoints(pb, kps, ow, oh))
                    n_refined += 1
                elif anat_score > POSE_ANAT_SUPPRESS_THRESH:
                    # Some keypoints visible but low score — keep as-is
                    result.append(pb)
                else:
                    # Matched to pose bbox but ZERO keypoints visible
                    # Only suppress if also large and low confidence
                    bw = pb.x2 - pb.x1
                    bh = pb.y2 - pb.y1
                    area_ratio = (bw * bh) / img_area
                    if area_ratio > POSE_FP_MIN_AREA and pb.conf < POSE_FP_MAX_CONF:
                        n_suppressed += 1
                        continue
                    result.append(pb)
            else:
                # Not matched to any pose detection
                if has_face_inside(pb):
                    # Face detector confirms a person
                    result.append(pb)
                    n_face_saved += 1
                    continue

                bw = pb.x2 - pb.x1
                bh = pb.y2 - pb.y1
                area_ratio = (bw * bh) / img_area

                if area_ratio > POSE_FP_MIN_AREA and pb.conf < POSE_FP_MAX_CONF:
                    # Large unmatched low-conf box — likely FP
                    n_suppressed += 1
                    continue
                else:
                    # Small box or high conf — keep
                    result.append(pb)

        if n_refined or n_suppressed or n_face_saved:
            logger.info(f"[pose] Refined {n_refined}, suppressed {n_suppressed} FP, "
                        f"face-saved {n_face_saved}, "
                        f"kept {len(result)}/{len(person_boxes)}")
        return result

    # ── Person inference with SAHI tiling ────────────────────────────────

    @staticmethod
    def _match_boxes_iou(boxes_a, boxes_b, iou_thr):
        """Match boxes from two sets by IoU. Returns (matched_pairs, unmatched_a, unmatched_b).

        matched_pairs: list of (idx_a, idx_b, iou) tuples
        unmatched_a: list of indices in boxes_a with no match
        unmatched_b: list of indices in boxes_b with no match
        """
        if len(boxes_a) == 0:
            return [], [], list(range(len(boxes_b)))
        if len(boxes_b) == 0:
            return [], list(range(len(boxes_a))), []

        matched_pairs = []
        used_b = set()

        for i in range(len(boxes_a)):
            best_iou = 0
            best_j = -1
            for j in range(len(boxes_b)):
                if j in used_b:
                    continue
                xx1 = max(boxes_a[i, 0], boxes_b[j, 0])
                yy1 = max(boxes_a[i, 1], boxes_b[j, 1])
                xx2 = min(boxes_a[i, 2], boxes_b[j, 2])
                yy2 = min(boxes_a[i, 3], boxes_b[j, 3])
                inter = max(0.0, xx2 - xx1) * max(0.0, yy2 - yy1)
                a1 = (boxes_a[i, 2] - boxes_a[i, 0]) * (boxes_a[i, 3] - boxes_a[i, 1])
                a2 = (boxes_b[j, 2] - boxes_b[j, 0]) * (boxes_b[j, 3] - boxes_b[j, 1])
                iou = inter / (a1 + a2 - inter + 1e-9)
                if iou > best_iou:
                    best_iou = iou
                    best_j = j
            if best_iou >= iou_thr:
                matched_pairs.append((i, best_j, best_iou))
                used_b.add(best_j)

        matched_a = {p[0] for p in matched_pairs}
        unmatched_a = [i for i in range(len(boxes_a)) if i not in matched_a]
        unmatched_b = [j for j in range(len(boxes_b)) if j not in used_b]

        return matched_pairs, unmatched_a, unmatched_b

    def _infer_person(self, image_bgr):
        """Person detection with TTA consensus merging.

        Pipeline (v3.23 — replaces concatenate+soft-NMS with consensus merging):
        1. Original pass at native 960px
        2. Flip TTA pass
        3. Match boxes across views (IoU >= PER_TTA_MATCH_IOU)
        4. Graduated confidence thresholds:
           - Confirmed by both views: keep at PER_TTA_CONF_BOTH (0.50)
           - Original-only: keep at PER_TTA_CONF_ORIG (0.60)
           - Flip-only: keep at PER_TTA_CONF_FLIP (0.75)
        5. Hard NMS on merged result
        6. Sanity filters + safety ceiling
        7. Pose FP filter + box refinement (if time allows)
        """
        oh, ow = image_bgr.shape[:2]
        t_start = time.monotonic()

        # Frame quality gating
        blur_score = self._frame_blur_score(image_bgr)
        is_blurry = blur_score < PER_BLUR_THRESHOLD

        # Pass 1: original image
        boxes_orig, confs_orig = self._per_run_pass(image_bgr, PER_TTA_CONF_BOTH)

        # Pass 2: horizontal flip
        flipped = cv2.flip(image_bgr, 1)
        boxes_flip, confs_flip = self._per_run_pass(flipped, PER_TTA_CONF_BOTH)
        if len(boxes_flip) > 0:
            boxes_flip[:, 0], boxes_flip[:, 2] = (
                ow - boxes_flip[:, 2], ow - boxes_flip[:, 0])

        if len(boxes_orig) == 0 and len(boxes_flip) == 0:
            return []

        # TTA consensus: match boxes across views
        matched, unmatched_o, unmatched_f = self._match_boxes_iou(
            boxes_orig, boxes_flip, PER_TTA_MATCH_IOU)

        # Build merged result with graduated thresholds
        merged_b = []
        merged_s = []

        # Confirmed by both views: keep original box, use max confidence, threshold=0.50
        for i_o, i_f, iou in matched:
            conf = max(float(confs_orig[i_o]), float(confs_flip[i_f]))
            if conf >= PER_TTA_CONF_BOTH:
                merged_b.append(boxes_orig[i_o])
                merged_s.append(conf)

        # Original-only: need higher confidence (0.60)
        for i_o in unmatched_o:
            if confs_orig[i_o] >= PER_TTA_CONF_ORIG:
                merged_b.append(boxes_orig[i_o])
                merged_s.append(float(confs_orig[i_o]))

        # Flip-only: strict threshold (0.75) — flip-only detections are likely FP
        for i_f in unmatched_f:
            if confs_flip[i_f] >= PER_TTA_CONF_FLIP:
                merged_b.append(boxes_flip[i_f])
                merged_s.append(float(confs_flip[i_f]))

        if not merged_b:
            return []

        merged_b = np.array(merged_b)
        merged_s = np.array(merged_s)

        # Hard NMS on merged result (no soft-NMS — no confidence decay)
        keep = _nms_per_class_boost(
            merged_b, merged_s,
            np.zeros(len(merged_s), dtype=int),  # single class
            iou_thr=PER_NMS_IOU)
        merged_b, merged_s = keep[0], keep[1]

        # Safety ceiling
        if len(merged_s) > PER_MAX_DET:
            top_idx = np.argsort(merged_s)[-PER_MAX_DET:]
            merged_b = merged_b[top_idx]
            merged_s = merged_s[top_idx]

        if len(merged_b) == 0:
            return []

        # Blur confidence penalty
        if is_blurry:
            merged_s = merged_s * PER_BLUR_CONF_PENALTY

        # Perspective scaling penalty
        merged_s = self._perspective_penalty(merged_b, merged_s, oh)

        # Final confidence floor (catches blur/perspective decay edge cases)
        keep_mask = merged_s >= PER_TTA_CONF_BOTH
        merged_b = merged_b[keep_mask]
        merged_s = merged_s[keep_mask]

        # Sanity filters
        img_area = float(oh * ow)
        out = []
        for i in range(len(merged_b)):
            bw = merged_b[i, 2] - merged_b[i, 0]
            bh = merged_b[i, 3] - merged_b[i, 1]
            if bw < PER_MIN_WH or bh < PER_MIN_WH:
                continue
            area = bw * bh
            if area < PER_MIN_AREA:
                continue
            if max(bw, bh) / max(min(bw, bh), 1e-6) > PER_MAX_ASPECT:
                continue
            if area / img_area > PER_MAX_AREA_RATIO:
                continue
            b = merged_b[i]
            # Shrink box 15% toward center to tighten fit (our boxes avg 57% larger than top miners')
            cx = (b[0] + b[2]) / 2.0
            cy = (b[1] + b[3]) / 2.0
            bw2 = (b[2] - b[0]) * 0.85 / 2.0
            bh2 = (b[3] - b[1]) * 0.85 / 2.0
            out.append(BoundingBox(
                x1=max(0, min(ow, int(cx - bw2))),
                y1=max(0, min(oh, int(cy - bh2))),
                x2=max(0, min(ow, int(cx + bw2))),
                y2=max(0, min(oh, int(cy + bh2))),
                cls_id=0,
                conf=max(0.0, min(1.0, float(merged_s[i]))),
            ))

        # Pose FP filter + box refinement (only if time budget allows)
        if time.monotonic() - t_start < PER_RTF_BUDGET * 0.85:
            out = self._pose_filter_refine(out, image_bgr)

        return out

    # ── Element detection (stack frame inspection) ──────────────────────────
    _CHALLENGE_TYPE_MAP = {2: 'person', 12: 'vehicle'}

    def _detect_element_hint(self) -> str:
        """Detect whether this request is for person or vehicle.

        Reads challenge_type_id from the chute template predict() metadata
        via stack frame inspection. Returns 'person', 'vehicle', or 'both'.
        """
        frame = None
        try:
            frame = inspect.currentframe()
            for _ in range(10):
                frame = frame.f_back
                if frame is None:
                    break
                meta = frame.f_locals.get('metadata')
                if isinstance(meta, dict) and 'challenge_type_id' in meta:
                    ct_id = meta['challenge_type_id']
                    hint = self._CHALLENGE_TYPE_MAP.get(ct_id)
                    if hint:
                        return hint
                    return 'both'
        except Exception:
            pass
        finally:
            del frame
        return 'both'

    # ── Unified inference ───────────────────────────────────────────────────

    def _infer_single(self, image_bgr: ndarray, element_hint: str = 'both') -> list[BoundingBox]:
        self._cached_pose_data = None  # reset before each frame

        if element_hint == 'person':
            return self._infer_person(image_bgr)

        if element_hint == 'vehicle':
            # Run vehicle detection + parts confirmation with empty person_boxes.
            # Plate/headlight/window checks fire normally; driver/rider overlap
            # check finds no matches (boost=0) but doesn't suppress.
            vehicle_boxes = self._infer_vehicle(image_bgr)
            return self._vehicle_parts_confirm(vehicle_boxes, [], image_bgr)

        # Fallback: run both (original behavior)
        if ENABLE_PARALLEL:
            veh_future = self._executor.submit(self._infer_vehicle, image_bgr)
            per_future = self._executor.submit(self._infer_person, image_bgr)
            vehicle_boxes = veh_future.result()
            person_boxes = per_future.result()
        else:
            vehicle_boxes = self._infer_vehicle(image_bgr)
            person_boxes = self._infer_person(image_bgr)

        # Vehicle parts confirmation: cross-reference with person detections
        vehicle_boxes = self._vehicle_parts_confirm(
            vehicle_boxes, person_boxes, image_bgr)

        return vehicle_boxes + person_boxes


    # -- Replay buffer -------------------------------------------------------
    REPLAY_DIR = Path("/home/miner/replay_buffer")
    REPLAY_MAX = 100

    def _replay_save(self, batch_images, results):
        try:
            ts = datetime.now(timezone.utc).strftime("%Y%m%d_%H%M%S_%f")
            query_dir = self.REPLAY_DIR / ts
            query_dir.mkdir(parents=True, exist_ok=True)

            for i, img in enumerate(batch_images):
                cv2.imwrite(str(query_dir / f"img_{i:03d}.jpg"), img,
                            [cv2.IMWRITE_JPEG_QUALITY, 95])

            preds = []
            for r in results:
                preds.append({
                    "frame_id": r.frame_id,
                    "boxes": [b.model_dump() for b in r.boxes],
                })
            meta = {
                "timestamp": ts,
                "num_images": len(batch_images),
                "image_shapes": [list(img.shape) for img in batch_images],
                "predictions": preds,
            }
            (query_dir / "meta.json").write_text(json.dumps(meta, indent=2))
            self._replay_prune()
        except Exception:
            pass

    def _replay_prune(self):
        try:
            dirs = sorted(
                [d for d in self.REPLAY_DIR.iterdir() if d.is_dir()],
                key=lambda d: d.name,
            )
            if len(dirs) > self.REPLAY_MAX:
                import shutil
                for old in dirs[: len(dirs) - self.REPLAY_MAX]:
                    shutil.rmtree(old, ignore_errors=True)
        except Exception:
            pass

    def predict_batch(
        self,
        batch_images: list[ndarray],
        offset: int,
        n_keypoints: int,
    ) -> list[TVFrameResult]:
        t_start = time.perf_counter()

        # Detect element type from caller metadata
        element_hint = self._detect_element_hint()
        t_setup = time.perf_counter()
        dt_setup = (t_setup - t_start) * 1000

        _lat_logger.info(
            "REQUEST batch=%d hint=%s setup=%.1fms",
            len(batch_images), element_hint, dt_setup,
        )

        results: list[TVFrameResult] = []
        for idx, image in enumerate(batch_images):
            t_img = time.perf_counter()
            boxes = self._infer_single(image, element_hint=element_hint)
            t_post = time.perf_counter()
            dt_infer = (t_post - t_img) * 1000

            keypoints = [(0, 0) for _ in range(max(0, int(n_keypoints)))]
            results.append(TVFrameResult(
                frame_id=offset + idx, boxes=boxes, keypoints=keypoints,
            ))
            dt_post = (time.perf_counter() - t_post) * 1000

            if idx < 3 or idx == len(batch_images) - 1:
                _lat_logger.info(
                    "  IMG %d/%d boxes=%d infer=%.1fms post=%.1fms shape=%s",
                    idx, len(batch_images), len(boxes), dt_infer, dt_post,
                    image.shape,
                )

        t_done = time.perf_counter()
        dt_total = (t_done - t_start) * 1000
        total_boxes = sum(len(r.boxes) for r in results)

        _lat_logger.info(
            "DONE batch=%d boxes=%d total=%.1fms setup=%.1fms hint=%s",
            len(batch_images), total_boxes, dt_total, dt_setup, element_hint,
        )
        logger.info(f"[miner] predict_batch: {len(batch_images)} images, "
                    f"{total_boxes} total boxes, {dt_total:.0f}ms (hint={element_hint})")

        threading.Thread(
            target=self._replay_save,
            args=(batch_images, results),
            daemon=True,
        ).start()

        return results
# Miner v3.19 — 1-pass vehicle + CLAHE pass + parts_confirm fix — element detection + per-step timing — background TRT engine build + CUDA-first fallback 20260402