Blame for code/case1/extract/extract_metrics.py

#!/usr/bin/env python3
"""Per-episode metric extractor for the Case 1 (ART) simulation logs.

Reads every episode directory RUN_DIR/episodes/<episode_id>/ written by the lock-step runner (episode_config.json,
summary.json, ticks.csv.gz, state100.csv.gz; column definitions in runner/schemas/log_schema.md) and writes ONE
compact row per episode:

    OUTDIR/<batch>/recovery_part<k>.csv.gz   fixed_horizon (recovery / TR) episodes
    OUTDIR/<batch>/fullpath_part<k>.csv.gz   path_completion (full-path) episodes
    OUTDIR/<batch>/errors_part<k>.jsonl      episodes that could not be extracted
    OUTDIR/<batch>/MANIFEST.json             inputs, counts, shard SHA256, extractor SHA256

Python >= 3.8 with numpy only. Nothing here reads physical data.

Streaming / multiprocess / resumable: episodes are processed by a worker pool and written in shards of at most
--shard-size rows; each shard appears atomically (tmp + rename). On a rerun the episode ids already present in
complete shards are skipped; if such an episode's summary.json changed (different SHA256), the run is refused
(outputs of different runner outputs are never mixed). Shards are never rewritten.

Metric definitions. All times are measured from control start c0 (runner semantics, t = step * 1e-3 - c0):
  * Primary error: released error_state(0.0) lateral e[1] and heading e[2] of the TRUE COG state on the 100 Hz log
    (state100 e_lat, e_head; recomputed here and checked). Tube cell (L, Hd): |e_lat| < L and |e_head| < Hd.
    Grid L in {0.05, 0.10, 0.20} m x Hd in {0.05, 0.10, 0.20} rad; primary cell 0.10 / 0.10.
  * first entry = t of the first 100 Hz sample with t >= 0 inside the cell.
  * sustained residence D in {0.5, 1.0} s = t of the first sample k such that every sample in [t_k, t_k + D]
    is inside (closed interval: 51 / 101 consecutive 100 Hz samples); the residence must be observed in the log.
  * started_inside = the first sample with t >= 0 is inside.
  * exits after first entry within H = number of inside->outside transitions among samples t_entry <= t <= H.
  * recovery distance to first entry: along-track progress (projection of the COG displacement on the straight
    reference direction) and odometer (sum of 100 Hz COG step lengths), both from the first sample with t >= 0.
  * bounded score at H in {15, 25, 40} s: T = entry if entry (+ D for residence variants) <= H and no controller
    failure (solver failure, exception, non-finite output) happened at or before entry; T = H otherwise if the log
    covers H or ended early by a terminal rule (divergence stop or terminate-on-failure); else censored (empty).
  * Variants: the same cells on the 10 Hz published ticks with the tracking error et1, et2 ("trk10") and
    the controller's 1 m lookahead error e1, e2 ("la10"); released error_state(1.0) of the true 100 Hz COG
    ("la100"); geometric perpendicular distance to the straight line and wrapped yaw error ("geo100", primary
    cell only).
  * Full paths: run MAE with the same definitions as the physical runs (physical/physical_runs.py: nearest
    waypoint, zero lookahead, signed body-frame lateral, wrapped heading) on the pose the controller used at each
    published 10 Hz tick (obs_x, obs_y, obs_theta), window from the first published tick to the completion index
    of the extended-lap forward tracker (track(window 30) + completion_index, one extra lap for closed paths,
    including the Path-3 loop-closure rule and a 100 Hz completion fallback). Evaluation outcome:
    completed / end_departure (no offline completion, max tracked index >= N - 10; window to the first sample at
    the maximum index) / failed (window = all samples); post_completion_departure = offline completed but the
    online runner outcome is not `completed`.

Usage:
    python3 extract_metrics.py --out OUTDIR --batch NAME [--workers N] [--shard-size 2000]
                               [--waypoints-dir DIR ...] [--run-label-root DIR] RUN_DIR [RUN_DIR ...]
    python3 extract_metrics.py --selftest            (synthetic checks of the metric kernels)
"""
import argparse
import glob
import gzip
import hashlib
import io
import json
import math
import multiprocessing as mp
import os
import platform
import sys
import time

import numpy as np

EXTRACTOR_VERSION = 'tr-c1-extract/1'
DT = 1e-3
LAT_GRID = (0.05, 0.10, 0.20)
HEAD_GRID = (0.05, 0.10, 0.20)
PRIMARY_CELL = (0.10, 0.10)
RESIDENCE_S = (0.5, 1.0)
HORIZONS = (15.0, 25.0, 40.0)
DIAG_H = 25.0                      # horizon for behaviour diagnostics and horizon means (primary H)
STEER_REVERSAL_BAND = 0.05         # deadband for steering reversals (command units)
SAT_EPS = 1e-12
EPS_T = 1e-9
CTRL_FAIL = ('controller_solver_failure', 'controller_exception', 'controller_nonfinite_output')
CLOSED_PATHS = {'path_iros_1': True, 'path_iros_2': False, 'lot17_sinsquare': True, 'circ_r2_ccw': True,
                'circ_r2_cw': True, 'circ_r5_ccw': True, 'circ_r5_cw': True, 'straight_line': False}
PATH_KEY = {'path_iros_1': 'P1', 'path_iros_2': 'P2', 'lot17_sinsquare': 'P3'}
SPIN = {'yaw_rate_min': 1.5, 'factor': 2.0, 'r_min': 1.02, 'n_consecutive': 3}   # runner metrics.SPIN_CRITERION
DEFAULT_WAYPOINT_DIRS = ['/sbel/Desktop/waypoints_paths',
                         os.path.join(os.path.dirname(os.path.abspath(__file__)), '..', 'data', 'waypoints')]


def cell_name(L, H):
    return f'L{int(round(L * 100)):03d}H{int(round(H * 100)):03d}'


CELLS = [(L, H) for L in LAT_GRID for H in HEAD_GRID]
PCELL = cell_name(*PRIMARY_CELL)


def res_name(D):
    return f'sus{int(round(D * 100)):03d}'


# ----------------------------------------------------------------------------------------------- column lists
# episode tags written by runner/batches/generate_lists.py; each becomes a 'tag_<key>' column (any other tag
# stays available in tags_json)
TAG_KEYS = ['batch', 'family', 'condition', 'row', 'rep', 'rf_arm', 'path', 'phase_ms', 'nominal_role',
            'sensitivity', 'repeat', 'placement_index', 'crn_source', 'tuning_family', 'tuning_config_index',
            'reference_name']
ID_COLUMNS = [
    'batch', 'run_dir', 'episode_id', 'scenario_id', 'kind', 'controller', 'ctrl_label', 'base_controller',
    'steering_scale', 'nnhd_gain', 'controller_params_json', 'ref_type', 'ref_name', 'path_key',
    'latency_model', 'controller_period_s', 'controller_phase_s', 'compute_delay_s', 'state_delay_s',
    'control_start_s', 'horizon_s', 'init_x', 'init_y', 'init_yaw', 'init_frame',
    'placement_heading', 'placement_x0', 'placement_y0', 'e_m', 'phi_rad', 'e_geom_m', 'phi_geom_rad',
    'steering_bias', 'actuation_delay_s', 'friction', 'noise_std_x', 'noise_std_y', 'noise_std_yaw', 'noise_seed',
    'tube_lat_m', 'tube_head_rad', 'divergence_dist_m', 'divergence_hold_s', 'stop_on_divergence',
    'on_controller_failure', 'osqp_rho_interval',
] + ['tag_' + k for k in TAG_KEYS] + [
    'tags_json', 'default_config_sha256',
    'config_hash', 'config_hash_ok', 'code_digest', 'host_fingerprint', 'run_id', 'attempt',
    'summary_sha256', 'ticks_sha256', 'state_sha256', 'files_hash_ok', 'reference_sha256', 'ref_sha_ok',
    'status', 'outcome', 'failure_category', 'failure_step', 'failure_t_ctrl', 'final_step', 'sim_s',
    'log_end_t', 'terminal_stop', 'completion_step', 'summary_divergence_step',
    'n_ticks', 'n_published', 'n_rate_limited', 'n_state_rows',
    'n_tick_ok', 'n_tick_no_state', 'n_tick_controller_dead', 'n_tick_solver_failure', 'n_tick_exception',
    'n_tick_nonfinite',
    'wall_total_s', 'wall_init_s', 'wall_stepping_s', 'wall_controller_s', 'wall_other_s',
    'server_spawn_to_ready_s', 'physics_pure_step_wall_s',
    'chassis_touchdown_step', 'first_command_apply_step', 'cog_z_at_first_command_apply', 'first_cmd_airborne',
]

DIAG_COLUMNS = [
    'diag_H', 'e_lat0', 'e_head0', 'first_zero_cross_t', 'overshoot_m', 'overshoot_t', 'n_lat_zero_cross',
    'n_lat_sign_changes_after_entry', 'p2p_lat_last5s', 'max_abs_lat_after_entry', 'max_abs_head_after_entry',
    'lat_at_H', 'head_at_H', 'max_abs_lat_H', 'max_abs_head_H',
    'frac_steer_sat', 'frac_rate_limited', 'n_steer_reversals', 'n_steer_sign_changes_raw',
    'thr_mean', 'thr_sd', 'frac_thr_one', 'frac_thr_zero', 'raw_thr_mean', 'raw_thr_max', 'thr_mean_to_entry',
    'speed_mean_H', 'speed_max_H', 't_speed_09', 'speed_mean_to_entry', 'max_abs_yaw_rate_H',
    'spin_flag', 'spin_first_step', 'spin_n_flagged', 'max_abs_yaw_rate_all',
    'div_flag', 'div_onset_step', 'div_onset_t', 'div_recovered_tube', 'time_beyond_div_dist_s', 'max_dist_min',
]


def recovery_columns():
    cols = []
    for (L, H) in CELLS:
        c = cell_name(L, H)
        cols += [f'trk100_{c}_first', f'trk100_{c}_{res_name(0.5)}', f'trk100_{c}_{res_name(1.0)}',
                 f'trk100_{c}_inside0', f'trk100_{c}_exits_H25', f'trk100_{c}_fracin_H25',
                 f'trk100_{c}_dprog', f'trk100_{c}_dodo',
                 f'trk10_{c}_first', f'la10_{c}_first', f'la100_{c}_first']
        for Hh in HORIZONS:
            cols += [f'T{int(Hh)}_trk100_{c}_first']
    for Hh in HORIZONS:
        for D in RESIDENCE_S:
            cols.append(f'T{int(Hh)}_trk100_{PCELL}_{res_name(D)}')
        cols += [f'T{int(Hh)}_trk10_{PCELL}_first', f'T{int(Hh)}_la10_{PCELL}_first',
                 f'T{int(Hh)}_la100_{PCELL}_first', f'T{int(Hh)}_geo100_{PCELL}_first',
                 f'reason{int(Hh)}_trk100_{PCELL}_first',
                 f'exits_{PCELL}_H{int(Hh)}', f'fracin_{PCELL}_H{int(Hh)}',
                 f'hmean_abs_lat_H{int(Hh)}', f'hmean_abs_head_H{int(Hh)}', f'hmean_complete_H{int(Hh)}']
    cols += [f'geo100_{PCELL}_first', f'geo100_{PCELL}_inside0', f'trk100_{PCELL}_final_stay_t',
             'max_abs_lat_all', 'final_e_lat', 'final_e_head']
    cols += DIAG_COLUMNS
    cols += ['chk_tube_first', 'chk_tube_tick', 'chk_tube_tick_la', 'chk_tube_100hz_la', 'chk_fracin_all',
             'chk_exits_all', 'chk_final_stay', 'chk_max_abs_lat', 'chk_final_e', 'chk_spin', 'chk_divergence',
             'chk_err_recompute_maxdiff', 'ref_index_at_end'] + [f'ref_index_at_end_H{int(Hh)}' for Hh in HORIZONS] + ['chk_all']
    return cols


FULLPATH_COLUMNS_EXTRA = [
    'n_waypoints', 'closed_path', 'ref_length_m', 'eval_outcome', 'online_outcome', 'post_completion_departure',
    'completion_from_100hz_fallback', 'completion_index', 'completion_time_s', 'max_tracked_index',
    'eval_end_index', 'eval_samples', 'mae_lat_m', 'mae_head_rad', 'mae_lat_true_m', 'mae_head_true_rad',
    'max_abs_lat_window_m', 'max_abs_lat_100hz_to_end_m', 'sim_s_to_offline_completion',
    'common_window_end_index', 'mae_lat_common_m', 'mae_head_common_rad',
    'online_mae_lat_m', 'online_mae_head_rad', 'online_mae_samples',
    'fp_frac_steer_sat', 'fp_frac_rate_limited', 'fp_n_steer_reversals', 'fp_thr_mean', 'fp_frac_thr_one',
    'fp_frac_thr_zero', 'fp_raw_thr_mean', 'fp_speed_mean', 'fp_speed_max', 'fp_max_abs_yaw_rate',
    'fp_spin_flag', 'fp_max_dist_min',
    'chk_online_mae', 'chk_spin', 'chk_all',
]


def fullpath_columns():
    return list(FULLPATH_COLUMNS_EXTRA)


# ----------------------------------------------------------------------------------------------- small helpers
def sha256_bytes(b):
    return hashlib.sha256(b).hexdigest()


def sha256_file(p):
    h = hashlib.sha256()
    with open(p, 'rb') as f:
        for blk in iter(lambda: f.read(1 << 20), b''):
            h.update(blk)
    return h.hexdigest()


def read_csv_gz_bytes(raw):
    """Parse a runner CSV (gzip bytes) into {column: numpy array of str}. Empty field = ''."""
    text = gzip.decompress(raw).decode()
    lines = text.split('\n')
    head = lines[0].split(',')
    rows = [l.split(',') for l in lines[1:] if l]
    if rows and any(len(r) != len(head) for r in rows):
        raise ValueError('ragged CSV')
    if not rows:
        return {k: np.array([], dtype=str) for k in head}, 0
    arr = np.array(rows, dtype=object)
    return {k: arr[:, i].astype(str) for i, k in enumerate(head)}, len(rows)


def fcol(d, k):
    a = d[k]
    if a.size == 0:
        return np.zeros(0)
    a = np.where(a == '', 'nan', a)
    return a.astype(float)


def icol(d, k):
    return np.array([int(v) if v != '' else -999999 for v in d[k]], dtype=np.int64)


def wrap(a):
    return np.arctan2(np.sin(a), np.cos(a))


def fnone(v):
    return None if v is None else float(v)


def same_opt(a, b, tol=1e-9):
    """True if both None or |a - b| <= tol."""
    if a is None or b is None:
        return a is None and b is None
    return abs(float(a) - float(b)) <= tol


# ----------------------------------------------------------------- released error state (copy of runner metrics.py)
def wrap_released(theta):
    th = np.array(theta, dtype=float, copy=True)
    for _ in range(64):
        lo = th < -np.pi
        hi = th > np.pi
        if not (lo.any() or hi.any()):
            break
        th[lo] = th[lo] + 2 * np.pi
        th[hi] = th[hi] - 2 * np.pi
    return th


def heading_error_released(ref, act):
    ref = np.asarray(ref, dtype=float)
    act = np.asarray(act, dtype=float)
    out = np.empty(np.broadcast(ref, act).shape)
    same = ((ref > 0) & (act > 0)) | ((ref <= 0) & (act <= 0))
    c2 = (~same) & (ref <= 0) & (act > 0)
    c3 = (~same) & ~c2
    out[same] = (ref - act)[same]
    a2 = np.abs(ref - act) < np.abs(2 * np.pi + ref - act)
    out[c2 & a2] = (-np.abs(act - ref))[c2 & a2]
    out[c2 & ~a2] = np.abs(2 * np.pi + ref - act)[c2 & ~a2]
    a3 = np.abs(ref - act) < np.abs(2 * np.pi - ref + act)
    out[c3 & a3] = np.abs(act - ref)[c3 & a3]
    out[c3 & ~a3] = (-np.abs(2 * np.pi - ref + act))[c3 & ~a3]
    return out


def nearest_index(x, y, theta, ref, lookahead, chunk=2048):
    x = np.atleast_1d(np.asarray(x, dtype=float))
    y = np.atleast_1d(np.asarray(y, dtype=float))
    th = wrap_released(np.atleast_1d(np.asarray(theta, dtype=float)))
    px = x + np.cos(th) * lookahead
    py = y + np.sin(th) * lookahead
    idx = np.empty(len(x), dtype=np.int64)
    for i in range(0, len(x), chunk):
        d = (px[i:i + chunk, None] - ref[None, :, 0]) ** 2 + (py[i:i + chunk, None] - ref[None, :, 1]) ** 2
        idx[i:i + chunk] = d.argmin(axis=1)
    return idx


def error_state_vec(x, y, theta, v, ref, lookahead):
    x = np.atleast_1d(np.asarray(x, dtype=float))
    y = np.atleast_1d(np.asarray(y, dtype=float))
    v = np.atleast_1d(np.asarray(v, dtype=float))
    th = wrap_released(np.atleast_1d(np.asarray(theta, dtype=float)))
    idx = nearest_index(x, y, th, ref, lookahead)
    r = ref[idx]
    err_theta = heading_error_released(r[:, 2], th)
    c, s = np.cos(-th), np.sin(-th)
    dx, dy = r[:, 0] - x, r[:, 1] - y
    e0 = c * dx + (-s) * dy
    e1 = s * dx + c * dy
    return np.column_stack((e0, e1, err_theta, r[:, 3] - v)), idx


def audit_errors(x, y, yaw, ref, chunk=4096):
    """errors() of physical/physical_runs.py with zero lookahead (identical formula), chunked."""
    x, y, yaw = (np.asarray(a, dtype=float) for a in (x, y, yaw))
    idx = np.empty(len(x), dtype=np.int64)
    for i in range(0, len(x), chunk):
        idx[i:i + chunk] = ((x[i:i + chunk, None] - ref[None, :, 0]) ** 2 +
                            (y[i:i + chunk, None] - ref[None, :, 1]) ** 2).argmin(axis=1)
    dx, dy = ref[idx, 0] - x, ref[idx, 1] - y
    lat = -np.sin(yaw) * dx + np.cos(yaw) * dy
    head = wrap(ref[idx, 2] - yaw)
    return lat, head


# --------------------------------------------- full-path window helpers (forward tracker and completion rule)
def track(xy, ref, closed, window=30):
    R = np.vstack([ref[:, :2], ref[:, :2]]) if closed else ref[:, :2]
    idx, out = 0, np.zeros(len(xy), int)
    for k, p in enumerate(xy):
        hi = min(len(R), idx + window + 1)
        idx += int(np.argmin(((R[idx:hi] - p) ** 2).sum(1)))
        out[k] = idx
    return out


def completion_index(xy, tracked, ref):
    n = len(ref)
    d = ref[-1, :2] - ref[-2, :2]
    d = d / np.linalg.norm(d)
    ok = (tracked >= n) | ((tracked >= n - 1) & (((xy - ref[-1, :2]) @ d) >= 0))
    return int(np.argmax(ok)) if ok.any() else None


def ref_progress(xy, tracked, ref, closed):
    R = np.vstack([ref[:, :2], ref[:, :2]]) if closed else ref[:, :2]
    S = np.concatenate([[0], np.cumsum(np.linalg.norm(np.diff(R, axis=0), axis=1))])
    out = np.zeros(len(xy))
    for k, (p, i) in enumerate(zip(xy, tracked)):
        best = (np.inf, S[i])
        for j in (i - 1, i):
            if j < 0 or j + 1 >= len(R):
                continue
            a, b = R[j], R[j + 1]
            L = np.linalg.norm(b - a)
            t = np.clip((p - a) @ (b - a) / L ** 2, 0, 1)
            dd = np.linalg.norm(a + t * (b - a) - p)
            if dd < best[0]:
                best = (dd, S[j] + t * L)
        out[k] = best[1]
    return np.maximum.accumulate(out)


# ----------------------------------------------------------------------------------------------- references
def find_waypoint_file(name, dirs):
    for d in dirs:
        p = os.path.join(d, name + '.csv')
        if os.path.exists(p):
            return p
    raise FileNotFoundError(f'waypoint file {name}.csv not found in {dirs}')


def build_reference(spec, wp_dirs):
    """Copy of runner art_runner/references.build_reference (array part); returns (array, sha256 of array bytes)."""
    kind = spec['type']
    if kind in ('released_csv', 'csv_transformed'):
        with open(find_waypoint_file(spec['name'], wp_dirs)) as f:
            arr = np.loadtxt(f, delimiter=',')
        if kind == 'csv_transformed':
            rot = float(spec.get('rotate', 0.0))
            c, s = np.cos(rot), np.sin(rot)
            out = arr.copy()
            out[:, 0] = c * arr[:, 0] - s * arr[:, 1] + float(spec.get('dx', 0.0))
            out[:, 1] = s * arr[:, 0] + c * arr[:, 1] + float(spec.get('dy', 0.0))
            if rot != 0.0:
                out[:, 2] = wrap(arr[:, 2] + rot)
            arr = out
    elif kind == 'straight':
        x0, y0 = float(spec.get('x0', 0.0)), float(spec.get('y0', 0.0))
        h = float(spec.get('heading', 0.0))
        spacing = float(spec.get('spacing', 0.1))
        s0, s1 = float(spec.get('s_start', -10.0)), float(spec.get('s_end', 60.0))
        n = int(round((s1 - s0) / spacing)) + 1
        s = s0 + spacing * np.arange(n)
        arr = np.column_stack((x0 + s * np.cos(h), y0 + s * np.sin(h),
                               np.full(n, float(wrap(h))), np.full(n, float(spec.get('v_ref', 1.0)))))
    elif kind == 'array':
        arr = np.asarray(spec['points'], dtype=float)
    else:
        raise ValueError(f'unknown reference type {kind}')
    arr = np.asarray(arr, dtype=float)
    return arr, sha256_bytes(np.ascontiguousarray(arr).tobytes())


# ----------------------------------------------------------------------------------------------- metric kernels
def first_index(mask):
    k = np.flatnonzero(mask)
    return int(k[0]) if len(k) else None


def first_sustained_index(inside, n_req):
    """First k with inside[k:k+n_req] all True (n_req consecutive samples), else None."""
    if n_req <= 1:
        return first_index(inside)
    if len(inside) < n_req:
        return None
    c = np.convolve(inside.astype(np.int64), np.ones(n_req, dtype=np.int64), 'valid')
    return first_index(c >= n_req)


def residence_samples(D, dt_log=0.01):
    """Closed-interval residence [t_k, t_k + D] at the 100 Hz log: D / dt + 1 samples."""
    return int(round(D / dt_log)) + 1


def bounded_score(entry, H, fail_t, log_end_t, terminal, dur=0.0):
    """Bounded settling-time score. Returns (value or None if censored, reason).
    A controller failure at or before entry (and before H) scores H: after a failure the runner holds the last
    command (released ChROSDriverInputsHandler behaviour), so a later entry is not credited to the controller.
    Divergence alone does not score H ('diverged, then recovered' scores the first entry before H)."""
    if fail_t is not None and fail_t <= H + EPS_T and (entry is None or fail_t <= entry + EPS_T):
        return float(H), 'controller_failure_before_entry'
    if entry is not None and entry + dur <= H + EPS_T:
        return float(entry), 'success'
    if log_end_t + EPS_T >= H:
        return float(H), 'no_entry'
    if terminal:
        return float(H), 'terminal_stop'
    return None, 'censored'


def spin_events(yaw_rate, speed, yaw_rate_min=1.5, factor=2.0, r_min=1.02, n_consecutive=3):
    """Copy of runner metrics.spin_events."""
    yr = np.abs(np.asarray(yaw_rate, dtype=float))
    sp = np.asarray(speed, dtype=float)
    if len(yr) == 0:
        return None, 0, None
    flag = (yr > yaw_rate_min) & (yr > factor * sp / r_min)
    k0 = None
    if len(flag) >= n_consecutive:
        run = np.convolve(flag.astype(int), np.ones(n_consecutive, dtype=int), 'valid') >= n_consecutive
        k = np.flatnonzero(run)
        k0 = int(k[0]) if len(k) else None
    return k0, int(flag.sum()), float(yr.max())


def divergence(steps, dist_min, dist_m, hold_s):
    """Runner rule (episode.py): dist_min > dist_m continuously for >= hold_s (100 Hz log). Returns onset step of
    the first qualifying excursion (None), total time beyond dist_m (s), max dist_min."""
    hold = int(round(hold_s / DT))
    out = dist_min > dist_m
    onset, since = None, None
    for s, o in zip(steps, out):
        if o:
            since = s if since is None else since
            if s - since >= hold and onset is None:
                onset = since
        else:
            since = None
    return onset, float(out.sum() * 0.01), float(dist_min.max()) if len(dist_min) else None


def sign_changes(x, band=0.0):
    """Number of sign changes of x, ignoring samples with |x| <= band (deadband) and exact zeros."""
    x = np.asarray(x, dtype=float)
    s = np.sign(x[np.abs(x) > band])
    if len(s) < 2:
        return 0
    return int(np.sum(s[1:] != s[:-1]))


# ----------------------------------------------------------------------------------------------- per-episode
def label_from_controller(controller, params):
    if controller in ('mpc', 'nnmpc', 'pid', 'nnhd'):
        return controller
    if controller == 'plugin:wrapped_released':
        b = params.get('base')
        k = params.get('steering_scale', 1.0)
        lab = b if k == 1.0 else f'{b}x{int(round(float(k) * 100)):03d}'
        if params.get('shared_speed'):
            lab += '_shspd'
        return lab
    return controller.replace('plugin:', '')


def episode_identity(cfg, summ, batch, run_dir):
    tags = cfg.get('tags') or {}
    params = cfg.get('controller_params') or {}
    lat = cfg['latency']
    pert = cfg.get('perturbation') or {}
    noise = pert.get('obs_noise') or {}
    term = cfg['termination']
    ref = cfg['reference']
    pl = tags.get('placement') or {}
    controller = cfg['controller']
    base = params.get('base') if controller == 'plugin:wrapped_released' else controller
    scale = float(params.get('steering_scale', 1.0)) if controller == 'plugin:wrapped_released' else 1.0
    nnhd_gain = scale if base == 'nnhd' else None
    prov = summ.get('provenance') or {}
    fail = summ.get('failure') or {}
    c0 = float(cfg.get('control_start_s', 0.0))
    c0_steps = int(round(c0 * 1000))
    # geometric (e, phi) of the chassis start in the reference frame (straight references)
    e_geom = phi_geom = None
    if ref['type'] == 'straight':
        h = float(ref.get('heading', 0.0))
        dx = cfg['initial_pose']['x'] - float(ref.get('x0', 0.0))
        dy = cfg['initial_pose']['y'] - float(ref.get('y0', 0.0))
        e_geom = -math.sin(h) * dx + math.cos(h) * dy
        phi_geom = float(wrap(np.array(cfg['initial_pose']['yaw'] - h)))
    fail_step = fail.get('step') if fail.get('category') in CTRL_FAIL else None
    row = {
        'batch': batch, 'run_dir': run_dir, 'episode_id': cfg['episode_id'], 'scenario_id': cfg.get('scenario_id'),
        'kind': 'recovery' if term['type'] == 'fixed_horizon' else 'fullpath',
        'controller': controller, 'ctrl_label': tags.get('ctrl_label') or label_from_controller(controller, params),
        'base_controller': base, 'steering_scale': scale, 'nnhd_gain': nnhd_gain,
        'controller_params_json': json.dumps(params, sort_keys=True, separators=(',', ':')),
        'ref_type': ref['type'], 'ref_name': ref.get('name', ref['type']),
        'path_key': PATH_KEY.get(ref.get('name'), ref.get('name', ref['type'])),
        'latency_model': lat['model'], 'controller_period_s': lat.get('controller_period_s'),
        'controller_phase_s': lat.get('controller_phase_s'), 'compute_delay_s': lat.get('compute_delay_s'),
        'state_delay_s': lat.get('state_delay_s'), 'control_start_s': c0, 'horizon_s': cfg['horizon_s'],
        'init_x': cfg['initial_pose']['x'], 'init_y': cfg['initial_pose']['y'], 'init_yaw': cfg['initial_pose']['yaw'],
        'init_frame': cfg['initial_pose'].get('frame'),
        'placement_heading': pl.get('heading'), 'placement_x0': pl.get('x0'), 'placement_y0': pl.get('y0'),
        'e_m': tags.get('e_m'), 'phi_rad': tags.get('phi_rad'), 'e_geom_m': e_geom, 'phi_geom_rad': phi_geom,
        'steering_bias': pert.get('steering_bias'), 'actuation_delay_s': pert.get('actuation_delay_s'),
        'friction': pert.get('friction'), 'noise_std_x': noise.get('std_x'), 'noise_std_y': noise.get('std_y'),
        'noise_std_yaw': noise.get('std_yaw'), 'noise_seed': noise.get('seed'),
        'tube_lat_m': term.get('tube_lat_m'), 'tube_head_rad': term.get('tube_head_rad'),
        'divergence_dist_m': term.get('divergence_dist_m'), 'divergence_hold_s': term.get('divergence_hold_s'),
        'stop_on_divergence': term.get('stop_on_divergence'), 'on_controller_failure': cfg.get('on_controller_failure'),
        'osqp_rho_interval': cfg.get('mpc_osqp_adaptive_rho_interval'),
        'tags_json': json.dumps(tags, sort_keys=True, separators=(',', ':')),
        'default_config_sha256': tags.get('default_config_sha256'),
        'config_hash': summ.get('config_hash'),
        'code_digest': prov.get('code_digest'), 'host_fingerprint': prov.get('host_fingerprint'),
        'run_id': prov.get('run_id', summ.get('run_id')),
        'attempt': prov.get('attempt'), 'reference_sha256': summ.get('reference_sha256'),
        'status': summ.get('status'), 'outcome': summ.get('outcome'),
        'failure_category': fail.get('category'), 'failure_step': fail_step,
        'failure_t_ctrl': None if fail_step is None else fail_step * DT - c0_steps * DT,
        'final_step': summ.get('final_step'), 'sim_s': summ.get('sim_s'),
        'completion_step': summ.get('completion_step'), 'summary_divergence_step': summ.get('divergence_step'),
        'n_ticks': summ.get('n_ticks'), 'n_published': summ.get('n_published'),
        'n_rate_limited': summ.get('n_rate_limited'), 'n_state_rows': summ.get('n_state_rows'),
        'server_spawn_to_ready_s': summ.get('server_spawn_to_ready_s'),
        'physics_pure_step_wall_s': summ.get('physics_pure_step_wall_s'),
        'chassis_touchdown_step': summ.get('chassis_touchdown_step'),
        'first_command_apply_step': summ.get('first_command_apply_step'),
        'cog_z_at_first_command_apply': summ.get('cog_z_at_first_command_apply'),
    }
    for k in TAG_KEYS:
        v = tags.get(k)
        row['tag_' + k] = json.dumps(v, sort_keys=True) if isinstance(v, (dict, list)) else v
    w = summ.get('wall') or {}
    for k in ('total_s', 'init_s', 'stepping_s', 'controller_s', 'other_s'):
        row['wall_' + k] = w.get(k)
    if row['chassis_touchdown_step'] is not None and row['first_command_apply_step'] is not None:
        row['first_cmd_airborne'] = int(row['first_command_apply_step'] < row['chassis_touchdown_step'])
    # canonical config hash (runner config.config_hash)
    blob = json.dumps(cfg, sort_keys=True, separators=(',', ':'), allow_nan=False)
    row['config_hash_ok'] = int(hashlib.sha256(blob.encode()).hexdigest() == summ.get('config_hash'))
    return row, c0_steps


def tick_status_counts(tk):
    st = tk['status'] if 'status' in tk else np.array([], dtype=str)
    return {'n_tick_ok': int(np.sum(st == 'ok')), 'n_tick_no_state': int(np.sum(st == 'no_state')),
            'n_tick_controller_dead': int(np.sum(st == 'controller_dead')),
            'n_tick_solver_failure': int(np.sum(st == 'controller_solver_failure')),
            'n_tick_exception': int(np.sum(st == 'controller_exception')),
            'n_tick_nonfinite': int(np.sum(st == 'controller_nonfinite_output'))}


def recovery_metrics(cfg, summ, row, c0_steps, st, tk, ref):
    out = {}
    steps = icol(st, 'step')
    t_all = steps * DT - c0_steps * DT           # identical arithmetic to episode.py (t = step*DT - c0*DT)
    after = t_all >= 0
    t = t_all[after]
    x, y, yaw, speed = (fcol(st, k) for k in ('x', 'y', 'yaw', 'speed'))
    yaw_rate = fcol(st, 'yaw_rate')
    lat_all, head_all = fcol(st, 'e_lat'), fcol(st, 'e_head')
    lat, head = lat_all[after], head_all[after]
    log_end_t = float(t[-1]) if len(t) else 0.0
    out['log_end_t'] = log_end_t
    N_h = int(round(float(cfg['horizon_s']) * 1000))
    final_step = summ.get('final_step') or (int(steps[-1]) if len(steps) else 0)
    stopped_early = final_step < N_h
    terminal = bool(stopped_early and (summ.get('outcome') == 'diverged' or
                                       (summ.get('outcome') in CTRL_FAIL and cfg.get('on_controller_failure') == 'terminate')))
    out['terminal_stop'] = int(terminal)
    fail_t = row['failure_t_ctrl']
    # recompute the zero-lookahead error from the logged true state (check of state100 e_lat / e_head)
    E0, _ = error_state_vec(x, y, yaw, speed, ref, 0.0)
    out['chk_err_recompute_maxdiff'] = float(max(np.max(np.abs(E0[:, 1] - lat_all)), np.max(np.abs(E0[:, 2] - head_all)))) if len(x) else 0.0
    EL, _ = error_state_vec(x[after], y[after], yaw[after], speed[after], ref, 1.0)
    la_lat, la_head = EL[:, 1], EL[:, 2]
    # geometric variant on the straight line
    rs = cfg['reference']
    h = float(rs.get('heading', 0.0)) if rs['type'] == 'straight' else None
    if h is not None:
        ux, uy = math.cos(h), math.sin(h)
        px, py = x[after] - float(rs.get('x0', 0.0)), y[after] - float(rs.get('y0', 0.0))
        g_lat = -uy * px + ux * py
        g_head = wrap(yaw[after] - h)
        s_along = ux * px + uy * py
    else:
        g_lat = g_head = None
        s_along = fcol(st, 'arc_s')[after]
    xa, ya = x[after], y[after]
    odo = np.concatenate([[0.0], np.cumsum(np.hypot(np.diff(xa), np.diff(ya)))]) if len(xa) else np.zeros(0)
    # ticks (published)
    pubm = tk['published'] == '1' if len(tk.get('published', [])) else np.zeros(0, bool)
    tt = fcol(tk, 't_ctrl')[pubm] if pubm.any() else np.zeros(0)
    et1 = fcol(tk, 'et1')[pubm] if pubm.any() else np.zeros(0)
    et2 = fcol(tk, 'et2')[pubm] if pubm.any() else np.zeros(0)
    e1 = fcol(tk, 'e1')[pubm] if pubm.any() else np.zeros(0)
    e2 = fcol(tk, 'e2')[pubm] if pubm.any() else np.zeros(0)

    def entry_t(tv, a, b, L, Hd):
        k = first_index((np.abs(a) < L) & (np.abs(b) < Hd))
        return (float(tv[k]), k) if k is not None else (None, None)

    within = {Hh: t <= Hh + EPS_T for Hh in HORIZONS}
    prim = {}
    for (L, Hd) in CELLS:
        c = cell_name(L, Hd)
        inside = (np.abs(lat) < L) & (np.abs(head) < Hd)
        k = first_index(inside)
        te = float(t[k]) if k is not None else None
        out[f'trk100_{c}_first'] = te
        out[f'trk100_{c}_inside0'] = int(bool(inside[0])) if len(inside) else None
        for D in RESIDENCE_S:
            ks = first_sustained_index(inside, residence_samples(D))
            out[f'trk100_{c}_{res_name(D)}'] = float(t[ks]) if ks is not None else None
        if k is not None:
            w = within[25.0]
            seg = inside[k:][w[k:]]
            out[f'trk100_{c}_exits_H25'] = int(np.sum(np.diff(seg.astype(int)) == -1)) if len(seg) else None
            out[f'trk100_{c}_fracin_H25'] = float(seg.mean()) if len(seg) else None
            out[f'trk100_{c}_dprog'] = float(s_along[k] - s_along[0])
            out[f'trk100_{c}_dodo'] = float(odo[k])
        else:
            for kk in ('exits_H25', 'fracin_H25', 'dprog', 'dodo'):
                out[f'trk100_{c}_{kk}'] = None
        out[f'trk10_{c}_first'] = entry_t(tt, et1, et2, L, Hd)[0]
        out[f'la10_{c}_first'] = entry_t(tt, e1, e2, L, Hd)[0]
        out[f'la100_{c}_first'] = entry_t(t, la_lat, la_head, L, Hd)[0]
        for Hh in HORIZONS:
            out[f'T{int(Hh)}_trk100_{c}_first'] = bounded_score(te, Hh, fail_t, log_end_t, terminal)[0]
        if (L, Hd) == PRIMARY_CELL:
            prim = {'inside': inside, 'k': k, 'te': te}
    pc = PCELL
    Lp, Hp = PRIMARY_CELL
    if g_lat is not None:
        gin = (np.abs(g_lat) < Lp) & (np.abs(g_head) < Hp)
        kg = first_index(gin)
        out[f'geo100_{pc}_first'] = float(t[kg]) if kg is not None else None
        out[f'geo100_{pc}_inside0'] = int(bool(gin[0])) if len(gin) else None
    else:
        out[f'geo100_{pc}_first'] = out[f'geo100_{pc}_inside0'] = None
    for Hh in HORIZONS:
        hh = int(Hh)
        for D in RESIDENCE_S:
            out[f'T{hh}_trk100_{pc}_{res_name(D)}'] = bounded_score(out[f'trk100_{pc}_{res_name(D)}'], Hh, fail_t,
                                                                    log_end_t, terminal, dur=D)[0]
        for dfn in ('trk10', 'la10', 'la100', 'geo100'):
            out[f'T{hh}_{dfn}_{pc}_first'] = bounded_score(out[f'{dfn}_{pc}_first'], Hh, fail_t, log_end_t, terminal)[0]
        out[f'reason{hh}_trk100_{pc}_first'] = bounded_score(prim['te'], Hh, fail_t, log_end_t, terminal)[1]
        w = within[Hh]
        k = prim['k']
        if k is not None:
            seg = prim['inside'][k:][w[k:]]
            out[f'exits_{pc}_H{hh}'] = int(np.sum(np.diff(seg.astype(int)) == -1)) if len(seg) else None
            out[f'fracin_{pc}_H{hh}'] = float(seg.mean()) if len(seg) else None
        else:
            out[f'exits_{pc}_H{hh}'] = out[f'fracin_{pc}_H{hh}'] = None
        out[f'hmean_abs_lat_H{hh}'] = float(np.abs(lat[w]).mean()) if w.any() else None
        out[f'hmean_abs_head_H{hh}'] = float(np.abs(head[w]).mean()) if w.any() else None
        out[f'hmean_complete_H{hh}'] = int(log_end_t + EPS_T >= Hh)
    # final stay (summary stay_in_tube_time_100hz_s, full log)
    ins = prim['inside']
    if prim['k'] is not None and ins[-1]:
        outside = np.flatnonzero(~ins)
        out[f'trk100_{pc}_final_stay_t'] = float(t[outside[-1] + 1]) if len(outside) else float(t[0])
    else:
        out[f'trk100_{pc}_final_stay_t'] = None
    out['max_abs_lat_all'] = float(np.abs(lat_all).max()) if len(lat_all) else None
    out['final_e_lat'] = float(lat_all[-1]) if len(lat_all) else None
    out['final_e_head'] = float(head_all[-1]) if len(head_all) else None

    # ------------------------------------------------------------- behaviour diagnostics, within DIAG_H
    wH = within[DIAG_H]
    out['diag_H'] = DIAG_H
    latH, headH, tH = lat[wH], head[wH], t[wH]
    out['e_lat0'] = float(lat[0]) if len(lat) else None
    out['e_head0'] = float(head[0]) if len(head) else None
    s0 = np.sign(lat[0]) if len(lat) else 0.0
    kz = first_index(latH * s0 <= 0) if s0 != 0 else None
    if kz is not None:
        seg = -s0 * latH[kz:]
        kk = int(np.argmax(seg))
        out['first_zero_cross_t'], out['overshoot_m'], out['overshoot_t'] = float(tH[kz]), float(max(seg[kk], 0.0)), float(tH[kz + kk])
    else:
        out['first_zero_cross_t'], out['overshoot_m'], out['overshoot_t'] = None, 0.0, None
    out['n_lat_zero_cross'] = sign_changes(latH)
    kp = prim['k']
    if kp is not None and kp < len(latH):
        out['n_lat_sign_changes_after_entry'] = sign_changes(latH[kp:])
        out['max_abs_lat_after_entry'] = float(np.abs(latH[kp:]).max())
        out['max_abs_head_after_entry'] = float(np.abs(headH[kp:]).max())
    else:
        out['n_lat_sign_changes_after_entry'] = out['max_abs_lat_after_entry'] = out['max_abs_head_after_entry'] = None
    last5 = tH >= DIAG_H - 5.0 - EPS_T
    out['p2p_lat_last5s'] = float(latH[last5].max() - latH[last5].min()) if last5.any() and log_end_t + EPS_T >= DIAG_H else None
    out['lat_at_H'] = float(latH[-1]) if len(latH) else None
    out['head_at_H'] = float(headH[-1]) if len(headH) else None
    out['max_abs_lat_H'] = float(np.abs(latH).max()) if len(latH) else None
    out['max_abs_head_H'] = float(np.abs(headH).max()) if len(headH) else None
    # commands on published ticks within DIAG_H
    if pubm.any():
        wt = tt <= DIAG_H + EPS_T
        ps = fcol(tk, 'pub_steering')[pubm][wt]
        pt = fcol(tk, 'pub_throttle')[pubm][wt]
        rt = fcol(tk, 'raw_throttle')[pubm][wt]
        rl = fcol(tk, 'rate_limited')[pubm][wt]
        out['frac_steer_sat'] = float(np.mean(np.abs(ps) >= 1 - SAT_EPS)) if len(ps) else None
        out['frac_rate_limited'] = float(np.mean(rl == 1)) if len(rl) else None
        out['n_steer_reversals'] = sign_changes(ps, STEER_REVERSAL_BAND)
        out['n_steer_sign_changes_raw'] = sign_changes(ps)
        out['thr_mean'] = float(pt.mean()) if len(pt) else None
        out['thr_sd'] = float(pt.std()) if len(pt) else None
        out['frac_thr_one'] = float(np.mean(pt >= 1 - SAT_EPS)) if len(pt) else None
        out['frac_thr_zero'] = float(np.mean(pt <= SAT_EPS)) if len(pt) else None
        out['raw_thr_mean'] = float(np.nanmean(rt)) if len(rt) else None
        out['raw_thr_max'] = float(np.nanmax(rt)) if len(rt) else None
        te = prim['te']
        pre = tt[wt] <= (te if te is not None else DIAG_H) + EPS_T
        out['thr_mean_to_entry'] = float(pt[pre].mean()) if pre.any() else None
    spH = speed[after][wH]
    out['speed_mean_H'] = float(spH.mean()) if len(spH) else None
    out['speed_max_H'] = float(spH.max()) if len(spH) else None
    k9 = first_index(speed[after] >= 0.9)
    out['t_speed_09'] = float(t[k9]) if k9 is not None else None
    if prim['k'] is not None:
        out['speed_mean_to_entry'] = float(speed[after][:prim['k'] + 1].mean())
    out['max_abs_yaw_rate_H'] = float(np.abs(yaw_rate[after][wH]).max()) if wH.any() else None
    k0, nflag, yrmax = spin_events(yaw_rate, speed, **SPIN)
    out['spin_flag'] = int(k0 is not None)
    out['spin_first_step'] = int(steps[k0]) if k0 is not None else None
    out['spin_n_flagged'] = nflag
    out['max_abs_yaw_rate_all'] = yrmax
    term = cfg['termination']
    dmin = fcol(st, 'dist_min')
    onset, beyond, dmax = divergence(steps, dmin, term['divergence_dist_m'], term['divergence_hold_s'])
    out['div_flag'] = int(onset is not None)
    out['div_onset_step'] = onset
    out['div_onset_t'] = None if onset is None else onset * DT - c0_steps * DT
    out['div_recovered_tube'] = (int(prim['te'] is not None and prim['te'] > out['div_onset_t'])
                                 if onset is not None else None)
    out['time_beyond_div_dist_s'] = beyond
    out['max_dist_min'] = dmax

    # ------------------------------------------------------------- checks against the runner summary
    chk = {}
    chk['chk_tube_first'] = same_opt(out[f'trk100_{cell_name(term["tube_lat_m"], term["tube_head_rad"])}_first'],
                                     summ.get('tube_entry_time_100hz_s'), 0.0) if (term['tube_lat_m'], term['tube_head_rad']) in CELLS else True
    chk['chk_tube_tick'] = same_opt(out[f'trk10_{pc}_first'], summ.get('tube_entry_time_tick_s'), 0.0)
    if 'tube_entry_time_tick_lookahead_s' in summ:
        chk['chk_tube_tick_la'] = same_opt(out[f'la10_{pc}_first'], summ.get('tube_entry_time_tick_lookahead_s'), 0.0)
        chk['chk_tube_100hz_la'] = same_opt(out[f'la100_{pc}_first'], summ.get('tube_entry_time_100hz_lookahead_s'), 0.0)
    if prim['k'] is not None and 'fraction_inside_tube_after_entry' in summ:
        ins = prim['inside']
        chk['chk_fracin_all'] = abs(float(ins[prim['k']:].mean()) - summ['fraction_inside_tube_after_entry']) <= 1e-12
        chk['chk_exits_all'] = int(np.sum(np.diff(ins[prim['k']:].astype(int)) == -1)) == summ['tube_exits_after_entry']
        chk['chk_final_stay'] = same_opt(out[f'trk100_{pc}_final_stay_t'], summ.get('stay_in_tube_time_100hz_s'), 0.0)
    chk['chk_max_abs_lat'] = same_opt(out['max_abs_lat_all'], summ.get('max_abs_lat_100hz_m'), 0.0)
    chk['chk_final_e'] = same_opt(out['final_e_lat'], summ.get('final_e_lat_m'), 0.0) and \
        same_opt(out['final_e_head'], summ.get('final_e_head_rad'), 0.0)
    if 'spin_flag' in summ:
        chk['chk_spin'] = (bool(out['spin_flag']) == bool(summ['spin_flag']) and
                           out['spin_first_step'] == summ.get('spin_first_step') and
                           same_opt(out['max_abs_yaw_rate_all'], summ.get('max_abs_yaw_rate_100hz'), 0.0))
    if summ.get('divergence_step') is not None:
        chk['chk_divergence'] = out['div_onset_step'] == summ['divergence_step']
    chk['chk_err_recompute'] = out['chk_err_recompute_maxdiff'] <= 1e-9
    # guard: a nearest-waypoint index at either end of the straight reference is a design error
    nref = len(ref)
    ends = [icol(st, 'nearest_idx')] if 'nearest_idx' in st else []
    if pubm.any():
        ends += [icol(tk, k)[pubm] for k in ('idx_lookahead', 'idx_track') if k in tk]
    out['ref_index_at_end'] = int(any(np.any((v == 0) | (v == nref - 1)) for v in ends if len(v)))
    # the same guard evaluated on [0, c0 + H] of the episode (t_ctrl <= H) for every analysed horizon H, so a
    # truncation after c0 + H cannot remove a scenario from the H analysis. The whole-log flag above is kept.
    t_ends = ([t_all] if 'nearest_idx' in st else []) + \
        ([tt for k in ('idx_lookahead', 'idx_track') if k in tk] if pubm.any() else [])
    for Hh in HORIZONS:
        out[f'ref_index_at_end_H{int(Hh)}'] = int(any(np.any(((v == 0) | (v == nref - 1)) & (tv <= Hh + EPS_T))
                                                      for v, tv in zip(ends, t_ends) if len(v)))
    for k, v in chk.items():
        if k != 'chk_err_recompute':
            out[k] = int(bool(v))
    out['chk_all'] = int(all(bool(v) for v in chk.values()))
    return out


def fullpath_metrics(cfg, summ, row, c0_steps, st, tk, ref):
    out = {}
    name = cfg['reference'].get('name')
    closed = CLOSED_PATHS.get(name, False)
    n = len(ref)
    out['n_waypoints'], out['closed_path'] = n, int(closed)
    out['ref_length_m'] = float(np.linalg.norm(np.diff(ref[:, :2], axis=0), axis=1).sum())
    out['online_outcome'] = summ.get('outcome')
    pubm = tk['published'] == '1'
    x, y, yaw = fcol(tk, 'obs_x')[pubm], fcol(tk, 'obs_y')[pubm], fcol(tk, 'obs_theta')[pubm]
    xt, yt, yawt = fcol(tk, 'x')[pubm], fcol(tk, 'y')[pubm], fcol(tk, 'yaw')[pubm]
    tstep = icol(tk, 'step')[pubm]
    t = tstep * 1e-3
    steps100 = icol(st, 'step')
    out['log_end_t'] = float(steps100[-1] * DT - c0_steps * DT) if len(steps100) else 0.0
    out['terminal_stop'] = 0
    if len(x) == 0:
        out['eval_outcome'] = 'failed'
        out['chk_all'] = 0
        return out
    xy = np.column_stack([x, y])
    tr = track(xy, ref, closed)
    ci = completion_index(xy, tr, ref)
    lat, head = audit_errors(x, y, yaw, ref)
    latT, headT = audit_errors(xt, yt, yawt, ref)
    out['max_tracked_index'] = int(tr.max())
    fallback = False
    if ci is not None:
        end, outcome = ci + 1, 'completed'
        out['completion_time_s'] = float(t[ci] - t[0])
        out['sim_s_to_offline_completion'] = float(t[ci])
    elif summ.get('outcome') == 'completed' and summ.get('completion_step') is not None:
        # compare_calibration.py: the runner stops at its 100 Hz completion step (margin 0), so the last tick can lie
        # just before the final waypoint; the window is then all published ticks.
        end, outcome, fallback = len(x), 'completed', True
        out['completion_time_s'] = float(summ['completion_step'] * 1e-3 - t[0])
        out['sim_s_to_offline_completion'] = float(summ['completion_step'] * 1e-3)
    elif tr.max() >= n - 10:
        end, outcome = int(np.flatnonzero(tr == tr.max())[0]) + 1, 'end_departure'
    else:
        end, outcome = len(x), 'failed'
    out['eval_outcome'] = outcome
    out['completion_from_100hz_fallback'] = int(fallback)
    out['completion_index'] = ci
    out['post_completion_departure'] = int(outcome == 'completed' and not fallback and summ.get('outcome') != 'completed')
    out['eval_end_index'], out['eval_samples'] = end, end
    out['mae_lat_m'] = float(np.abs(lat[:end]).mean())
    out['mae_head_rad'] = float(np.abs(head[:end]).mean())
    out['mae_lat_true_m'] = float(np.abs(latT[:end]).mean())
    out['mae_head_true_rad'] = float(np.abs(headT[:end]).mean())
    out['max_abs_lat_window_m'] = float(np.abs(lat[:end]).max())
    end_step = int(tstep[end - 1])
    e100 = fcol(st, 'e_lat')
    out['max_abs_lat_100hz_to_end_m'] = float(np.abs(e100[steps100 <= end_step]).max()) if len(e100) else None
    # common window [0, L - 2 m] of reference arc length (sensitivity)
    s = ref_progress(xy[:end], tr[:end], ref, closed)
    kc = first_index(s >= out['ref_length_m'] - 2.0)
    endc = kc if kc is not None else end
    endc = max(endc, 1)
    out['common_window_end_index'] = endc
    out['mae_lat_common_m'] = float(np.abs(lat[:endc]).mean())
    out['mae_head_common_rad'] = float(np.abs(head[:endc]).mean())
    # online window (runner summary.run_mae_*): published ticks with step <= completion_step (or the last tick)
    tk_steps = icol(tk, 'step')
    end_online = summ.get('completion_step') or (int(tk_steps[-1]) if len(tk_steps) else 0)
    on = tstep <= end_online
    if on.any():
        out['online_mae_lat_m'] = float(np.abs(lat[on]).mean())
        out['online_mae_head_rad'] = float(np.abs(head[on]).mean())
        out['online_mae_samples'] = int(on.sum())
    # behaviour diagnostics on the evaluation window
    ps = fcol(tk, 'pub_steering')[pubm][:end]
    pt = fcol(tk, 'pub_throttle')[pubm][:end]
    rt = fcol(tk, 'raw_throttle')[pubm][:end]
    rl = fcol(tk, 'rate_limited')[pubm][:end]
    sp = fcol(tk, 'speed')[pubm][:end]
    out['fp_frac_steer_sat'] = float(np.mean(np.abs(ps) >= 1 - SAT_EPS))
    out['fp_frac_rate_limited'] = float(np.mean(rl == 1))
    out['fp_n_steer_reversals'] = sign_changes(ps, STEER_REVERSAL_BAND)
    out['fp_thr_mean'] = float(pt.mean())
    out['fp_frac_thr_one'] = float(np.mean(pt >= 1 - SAT_EPS))
    out['fp_frac_thr_zero'] = float(np.mean(pt <= SAT_EPS))
    out['fp_raw_thr_mean'] = float(np.nanmean(rt))
    out['fp_speed_mean'] = float(sp.mean())
    out['fp_speed_max'] = float(sp.max())
    w100 = steps100 <= end_step
    yr = fcol(st, 'yaw_rate')
    sp100 = fcol(st, 'speed')
    out['fp_max_abs_yaw_rate'] = float(np.abs(yr[w100]).max()) if w100.any() else None
    k0, _, yrmax = spin_events(yr, sp100, **SPIN)
    out['fp_spin_flag'] = int(k0 is not None)
    dmin = fcol(st, 'dist_min')
    out['fp_max_dist_min'] = float(dmin.max()) if len(dmin) else None
    chk = {'chk_online_mae': same_opt(out.get('online_mae_lat_m'), summ.get('run_mae_lateral_m'), 1e-15) and
           same_opt(out.get('online_mae_head_rad'), summ.get('run_mae_heading_rad'), 1e-15)}
    if 'spin_flag' in summ:
        chk['chk_spin'] = bool(out['fp_spin_flag']) == bool(summ['spin_flag']) and \
            same_opt(yrmax, summ.get('max_abs_yaw_rate_100hz'), 0.0)
    for k, v in chk.items():
        out[k] = int(bool(v))
    out['chk_all'] = int(all(chk.values()))
    return out


def extract_episode(task):
    """task = (episode_dir, batch, run_dir_label, wp_dirs). Returns (kind, row) or ('error', {...})."""
    ep_dir, batch, run_label, wp_dirs = task
    try:
        with open(os.path.join(ep_dir, 'summary.json'), 'rb') as f:
            sraw = f.read()
        summ = json.loads(sraw)
        with open(os.path.join(ep_dir, 'episode_config.json')) as f:
            cfg = json.load(f)
        if summ.get('status') != 'ok':
            return 'error', {'episode_dir': ep_dir, 'error': f"status {summ.get('status')} in episodes/"}
        row, c0_steps = episode_identity(cfg, summ, batch, run_label)
        row['summary_sha256'] = sha256_bytes(sraw)
        with open(os.path.join(ep_dir, 'ticks.csv.gz'), 'rb') as f:
            traw = f.read()
        with open(os.path.join(ep_dir, 'state100.csv.gz'), 'rb') as f:
            sraw100 = f.read()
        row['ticks_sha256'], row['state_sha256'] = sha256_bytes(traw), sha256_bytes(sraw100)
        files = summ.get('files') or {}
        row['files_hash_ok'] = int(files.get('ticks.csv.gz') == row['ticks_sha256'] and
                                   files.get('state100.csv.gz') == row['state_sha256'])
        tk, _ = read_csv_gz_bytes(traw)
        st, _ = read_csv_gz_bytes(sraw100)
        ref, ref_sha = build_reference(cfg['reference'], wp_dirs)
        row['ref_sha_ok'] = int(ref_sha == summ.get('reference_sha256'))
        row.update(tick_status_counts(tk))
        if row['kind'] == 'recovery':
            row.update(recovery_metrics(cfg, summ, row, c0_steps, st, tk, ref))
        else:
            row.update(fullpath_metrics(cfg, summ, row, c0_steps, st, tk, ref))
        row['chk_all'] = int(bool(row.get('chk_all', 0)) and row['files_hash_ok'] == 1 and row['ref_sha_ok'] == 1
                             and row['config_hash_ok'] == 1)
        return row['kind'], row
    except Exception as e:  # noqa: BLE001 - recorded, never silently dropped
        import traceback
        return 'error', {'episode_dir': ep_dir, 'error': f'{type(e).__name__}: {e}', 'traceback': traceback.format_exc()[-3000:]}


# ----------------------------------------------------------------------------------------------- writing
def fmt(v):
    if v is None:
        return ''
    if isinstance(v, (bool, np.bool_)):
        return '1' if v else '0'
    if isinstance(v, (int, np.integer)):
        return str(int(v))
    if isinstance(v, (float, np.floating)):
        v = float(v)
        if math.isnan(v):
            return ''
        return repr(v)
    s = str(v)
    if any(ch in s for ch in ',"\n'):
        s = '"' + s.replace('"', '""') + '"'
    return s


def write_shard(path, columns, rows):
    unknown = sorted({k for r in rows for k in r} - set(columns))
    if unknown:
        raise RuntimeError(f'extractor bug: keys not in the column list: {unknown[:10]}')
    buf = io.StringIO()
    buf.write(','.join(columns) + '\n')
    for r in rows:
        buf.write(','.join(fmt(r.get(c)) for c in columns) + '\n')
    tmp = path + f'.tmp{os.getpid()}'
    data = buf.getvalue().encode()
    with open(tmp, 'wb') as raw:
        with gzip.GzipFile(fileobj=raw, mode='wb', mtime=0, compresslevel=6) as gz:
            gz.write(data)
    os.replace(tmp, path)
    return sha256_file(path), sha256_bytes(data)


def write_errors(path, rows):
    tmp = path + f'.tmp{os.getpid()}'
    with open(tmp, 'w') as f:
        for r in rows:
            f.write(json.dumps(r, sort_keys=True) + '\n')
    os.replace(tmp, path)
    return sha256_file(path)


def split_csv_line(line):
    """Minimal CSV split supporting the quoted fields written by fmt()."""
    out, cur, q, i = [], [], False, 0
    while i < len(line):
        ch = line[i]
        if q:
            if ch == '"':
                if i + 1 < len(line) and line[i + 1] == '"':
                    cur.append('"')
                    i += 1
                else:
                    q = False
            else:
                cur.append(ch)
        elif ch == '"':
            q = True
        elif ch == ',':
            out.append(''.join(cur))
            cur = []
        else:
            cur.append(ch)
        i += 1
    out.append(''.join(cur))
    return out


def existing_ids(batch_dir):
    """episode key -> summary_sha256 from complete shards (resume)."""
    done = {}
    for p in sorted(glob.glob(os.path.join(batch_dir, '*_part*.csv.gz'))):
        with gzip.open(p, 'rt') as f:
            head = split_csv_line(f.readline().rstrip('\n'))
            i_run, i_ep, i_sha = head.index('run_dir'), head.index('episode_id'), head.index('summary_sha256')
            for line in f:
                v = split_csv_line(line.rstrip('\n'))
                done[(v[i_run], v[i_ep])] = v[i_sha]
    return done


def next_part(batch_dir):
    ks = [int(os.path.basename(p).split('_part')[1].split('.')[0])
          for p in glob.glob(os.path.join(batch_dir, '*_part*.*')) if '_part' in p and '.tmp' not in p]
    return max(ks) + 1 if ks else 0


def run_level_info(run_dir):
    info = {'run_dir': run_dir, 'manifests': [], 'host_timing': None,
            'n_infrastructure_error_dirs': 0, 'infrastructure_error_dirs': []}
    for m in sorted(glob.glob(os.path.join(run_dir, 'manifest_*.json'))):
        try:
            d = json.load(open(m))
        except Exception as e:  # noqa: BLE001
            info['manifests'].append({'file': m, 'error': str(e)})
            continue
        info['manifests'].append({'file': os.path.basename(m), 'sha256': sha256_file(m), 'run_id': d.get('run_id'),
                                  'status': d.get('status'), 'exit_code': d.get('exit_code'),
                                  'workers': d.get('workers'), 'wall_s': d.get('wall_s'),
                                  'n_scheduled': d.get('n_scheduled'), 'n_skipped_existing': d.get('n_skipped_existing'),
                                  'counts': d.get('counts'), 'started_utc': d.get('started_utc'),
                                  'finished_utc': d.get('finished_utc'),
                                  'infrastructure_exhausted': d.get('infrastructure_exhausted'),
                                  'n_infrastructure_events': len(d.get('infrastructure_events') or []),
                                  'code_digest': (d.get('provenance') or {}).get('code_digest'),
                                  'host_fingerprint': (d.get('provenance') or {}).get('host_fingerprint'),
                                  'per_episode_wall_total_s_sum': float(sum((e.get('wall') or {}).get('total_s', 0.0) or 0.0
                                                                            for e in d.get('per_episode_wall') or []))})
    ht = os.path.join(run_dir, 'host_timing.json')
    if os.path.exists(ht):
        info['host_timing'] = json.load(open(ht))
        info['host_timing_sha256'] = sha256_file(ht)
    errs = sorted(glob.glob(os.path.join(run_dir, 'infrastructure_errors', '*')))
    info['n_infrastructure_error_dirs'] = len(errs)
    info['infrastructure_error_dirs'] = [os.path.basename(e) for e in errs]
    return info


def main(argv=None):
    ap = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
    ap.add_argument('--out', help='output root')
    ap.add_argument('--batch', help='batch name (sub-directory of --out)')
    ap.add_argument('--workers', type=int, default=max(1, min(8, (os.cpu_count() or 2) - 2)))
    ap.add_argument('--shard-size', type=int, default=2000)
    ap.add_argument('--waypoints-dir', action='append', default=[])
    ap.add_argument('--run-label-root', default=None,
                    help='run_dir values are written relative to this directory (default: absolute paths)')
    ap.add_argument('--limit', type=int, default=None, help='extract at most K new episodes (testing)')
    ap.add_argument('--selftest', action='store_true')
    ap.add_argument('run_dirs', nargs='*')
    a = ap.parse_args(argv)
    if a.selftest:
        return selftest()
    if not (a.out and a.batch and a.run_dirs):
        ap.error('--out, --batch and at least one RUN_DIR are required')
    wp_dirs = a.waypoints_dir + DEFAULT_WAYPOINT_DIRS
    bdir = os.path.join(a.out, a.batch)
    os.makedirs(bdir, exist_ok=True)
    done = existing_ids(bdir)
    tasks, refused = [], []
    for rd in a.run_dirs:
        rd = os.path.abspath(rd)
        label = os.path.relpath(rd, a.run_label_root) if a.run_label_root else rd
        for sp in sorted(glob.glob(os.path.join(rd, 'episodes', '*', 'summary.json'))):
            ep = os.path.dirname(sp)
            eid = os.path.basename(ep)
            if eid.startswith('.tmp_'):
                continue
            key = (label, eid)
            if key in done:
                if sha256_file(sp) != done[key]:
                    refused.append(ep)
                continue
            tasks.append((ep, a.batch, label, wp_dirs))
    if refused:
        print(f'REFUSED: {len(refused)} already extracted episodes have a changed summary.json (e.g. {refused[0]}); '
              'use a new batch directory', file=sys.stderr)
        return 2
    if a.limit is not None:
        tasks = tasks[:a.limit]
    t0 = time.time()
    part = next_part(bdir)
    buffers = {'recovery': [], 'fullpath': [], 'error': []}
    cols = {'recovery': ID_COLUMNS + recovery_columns(), 'fullpath': ID_COLUMNS + fullpath_columns()}
    written = []

    def flush(kind, force=False):
        nonlocal part
        rows = buffers[kind]
        if not rows or (len(rows) < a.shard_size and not force):
            return
        if kind == 'error':
            p = os.path.join(bdir, f'errors_part{part:04d}.jsonl')
            h, hc = write_errors(p, rows), None
        else:
            p = os.path.join(bdir, f'{kind}_part{part:04d}.csv.gz')
            h, hc = write_shard(p, cols[kind], sorted(rows, key=lambda r: (r['run_dir'], r['episode_id'])))
        written.append({'file': os.path.basename(p), 'sha256': h, 'content_sha256': hc, 'rows': len(rows)})
        part += 1
        buffers[kind] = []

    n_done = 0
    if tasks:
        ctx = mp.get_context('fork') if hasattr(os, 'fork') else mp.get_context()
        with ctx.Pool(a.workers) as pool:
            for kind, row in pool.imap_unordered(extract_episode, tasks, chunksize=4):
                buffers[kind].append(row)
                n_done += 1
                flush(kind)
                if n_done % 500 == 0:
                    print(f'{n_done}/{len(tasks)} episodes, {time.time() - t0:.0f} s', flush=True)
    for k in ('recovery', 'fullpath', 'error'):
        flush(k, force=True)
    # manifest (rewritten; lists every shard present)
    shards = []
    for p in sorted(glob.glob(os.path.join(bdir, '*_part*.*'))):
        if '.tmp' in p:
            continue
        rec_ = {'file': os.path.basename(p), 'sha256': sha256_file(p)}
        if p.endswith('.csv.gz'):
            with gzip.open(p, 'rb') as fh:
                rec_['content_sha256'] = sha256_bytes(fh.read())   # gzip bytes depend on the zlib build; content does not
        shards.append(rec_)
    man_path = os.path.join(bdir, 'MANIFEST.json')
    old = json.load(open(man_path)) if os.path.exists(man_path) else {}
    invocations = old.get('invocations', [])
    invocations.append({'utc': time.strftime('%Y-%m-%dT%H:%M:%SZ', time.gmtime()), 'argv': sys.argv,
                        'n_new_episodes': n_done, 'written': written, 'wall_s': time.time() - t0,
                        'workers': a.workers})
    n_err = sum(1 for s in shards if s['file'].startswith('errors_'))
    manifest = {
        'extractor': EXTRACTOR_VERSION, 'extractor_sha256': sha256_file(os.path.abspath(__file__)),
        'python': sys.version, 'numpy': np.__version__, 'platform': platform.platform(),
        'batch': a.batch, 'run_dirs': [os.path.abspath(r) for r in a.run_dirs],
        'run_level': [run_level_info(os.path.abspath(r)) for r in a.run_dirs],
        'waypoint_dirs': wp_dirs, 'shards': shards, 'n_error_shards': n_err, 'invocations': invocations,
        'definitions': {'lat_grid_m': LAT_GRID, 'head_grid_rad': HEAD_GRID, 'primary_cell': PRIMARY_CELL,
                        'residence_s': RESIDENCE_S, 'residence_rule': 'closed interval [t_k, t_k + D], D/0.01 + 1 samples',
                        'horizons_s': HORIZONS, 'diag_H_s': DIAG_H, 'steer_reversal_band': STEER_REVERSAL_BAND,
                        'spin': SPIN},
    }
    with open(man_path + '.tmp', 'w') as f:
        json.dump(manifest, f, indent=1, sort_keys=True, default=str)
    os.replace(man_path + '.tmp', man_path)
    print(json.dumps({'batch': a.batch, 'new_episodes': n_done, 'written': written, 'wall_s': round(time.time() - t0, 1)}))
    return 0 if not any(w['file'].startswith('errors_') for w in written) else 1


# ----------------------------------------------------------------------------------------------- self test
def selftest():
    ok = True
    ins = np.array([0, 1, 1, 0, 1, 1, 1, 1, 0], bool)
    ok &= first_sustained_index(ins, 3) == 4 and first_sustained_index(ins, 5) is None and first_index(ins) == 1
    ok &= residence_samples(0.5) == 51 and residence_samples(1.0) == 101
    ok &= bounded_score(5.0, 25, None, 25.0, False) == (5.0, 'success')
    ok &= bounded_score(None, 25, None, 25.0, False) == (25.0, 'no_entry')
    ok &= bounded_score(None, 40, None, 25.0, False) == (None, 'censored')
    ok &= bounded_score(10.0, 40, None, 25.0, False) == (10.0, 'success')
    ok &= bounded_score(None, 25, None, 6.5, True) == (25.0, 'terminal_stop')
    ok &= bounded_score(5.0, 25, 3.0, 25.0, False) == (25.0, 'controller_failure_before_entry')
    ok &= bounded_score(5.0, 25, 7.0, 25.0, False) == (5.0, 'success')
    ok &= bounded_score(24.8, 25, None, 25.0, False, dur=0.5) == (25.0, 'no_entry')
    ok &= sign_changes([0.3, 0.01, -0.02, -0.3, 0.2], 0.05) == 2 and sign_changes([1, -1, 1]) == 2
    steps = np.arange(0, 6000, 10)
    d = np.zeros(len(steps)); d[100:400] = 6.0
    on, beyond, mx = divergence(steps, d, 5.0, 2.0)
    ok &= on == 1000 and abs(beyond - 3.0) < 1e-9
    d2 = np.zeros(len(steps)); d2[100:250] = 6.0
    ok &= divergence(steps, d2, 5.0, 2.0)[0] is None
    # error state on a straight +x line: vehicle at y = -1 heading 0 -> lateral +1 (reference to the left)
    ref = build_reference({'type': 'straight', 's_start': -30, 's_end': 60, 'spacing': 0.1}, [])[0]
    E, _ = error_state_vec([0.0], [-1.0], [0.0], [0.0], ref, 0.0)
    ok &= abs(E[0, 1] - 1.0) < 1e-12 and abs(E[0, 2]) < 1e-12
    print('selftest', 'PASS' if ok else 'FAIL')
    return 0 if ok else 1


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