@runweave/server-linux-x64 0.1.8 → 0.1.9

package/payload/plugins/toolkit/skills/youtube-reading/scripts/artifacts.py

@@ -0,0 +1,206 @@
+"""Versioned artifacts, atomic writes and mechanical (not semantic) review gates."""
+import contextlib
+import fcntl
+import hashlib
+import json
+import os
+from pathlib import Path
+import re
+import tempfile
+from datetime import datetime, timezone
+
+SCHEMA_VERSION = 1
+PROMPT_VERSION = '1'
+CONTRACT_VERSION = '1'
+
+
+class JobError(Exception):
+    def __init__(self, code, message):
+        super().__init__(message)
+        self.code = code
+
+
+def now():
+    return datetime.now(timezone.utc).isoformat()
+
+
+def digest(path):
+    h = hashlib.sha256()
+    with Path(path).open('rb') as f:
+        for block in iter(lambda: f.read(1024 * 1024), b''):
+            h.update(block)
+    return h.hexdigest()
+
+
+def identity(value):
+    return hashlib.sha256(json.dumps(value, sort_keys=True, ensure_ascii=False).encode()).hexdigest()
+
+
+def read_json(path):
+    return json.loads(Path(path).read_text(encoding='utf-8'))
+
+
+def atomic_write(path, content):
+    path = Path(path)
+    path.parent.mkdir(parents=True, exist_ok=True)
+    fd, tmp = tempfile.mkstemp(prefix='.' + path.name, dir=path.parent)
+    try:
+        with os.fdopen(fd, 'w', encoding='utf-8') as f:
+            f.write(content)
+            f.flush()
+            os.fsync(f.fileno())
+        os.replace(tmp, path)
+    finally:
+        if os.path.exists(tmp):
+            os.unlink(tmp)
+
+
+def write_json(path, value):
+    atomic_write(path, json.dumps(value, ensure_ascii=False, indent=2) + '\n')
+
+
+@contextlib.contextmanager
+def job_lock(directory):
+    directory = Path(directory)
+    directory.mkdir(parents=True, exist_ok=True)
+    # Keep the inode: flock releases on death, including SIGKILL, without stale
+    # PID reuse or lock-unlink races between competing writers.
+    with (directory / '.lock').open('a+') as f:
+        try:
+            fcntl.flock(f, fcntl.LOCK_EX | fcntl.LOCK_NB)
+        except BlockingIOError:
+            raise JobError('JOB_BUSY', '任务正在运行，请稍后 resume。')
+        try:
+            f.seek(0)
+            f.truncate()
+            f.write(json.dumps({'pid': os.getpid(), 'acquired_at': now()}))
+            f.flush()
+            yield
+        finally:
+            fcntl.flock(f, fcntl.LOCK_UN)
+
+
+def validate_source(source):
+    if not isinstance(source, dict) or source.get('schema_version') != SCHEMA_VERSION:
+        raise JobError('SOURCE_SCHEMA', '不支持的 source schema。')
+    previous = -1
+    ids = set()
+    if not isinstance(source.get('segments'), list) or not all(isinstance(s, dict) for s in source['segments']):
+        raise JobError('SOURCE_SCHEMA', 'segments 必须为段落对象数组。')
+    for s in source['segments']:
+        if (not isinstance(s['id'], str) or s['id'] in ids or
+                not isinstance(s['text'], str) or not s['text'].strip() or
+                not isinstance(s['start_ms'], int) or not isinstance(s['end_ms'], int) or
+                s['start_ms'] < 0 or s['start_ms'] < previous or s['end_ms'] <= s['start_ms']):
+            raise JobError('SOURCE_INVALID', '来源 ID、时间顺序或文本无效。')
+        ids.add(s['id'])
+        previous = s['start_ms']
+
+
+def chunk_source(source_path, work_dir):
+    source = read_json(source_path)
+    validate_source(source)
+    segments = source['segments']
+    groups, current, size = [], [], 0
+    for s in segments:
+        if len(s['text']) > 6000:
+            raise JobError('SEGMENT_TOO_LARGE', '单个来源段超过 6000 字符；先按来源时间拆段。')
+        if current and size + len(s['text']) > 6000:
+            groups.append(current)
+            current, size = [], 0
+        current.append(s)
+        size += len(s['text'])
+    if current:
+        groups.append(current)
+    out, offset = [], 0
+    root = Path(work_dir) / 'chunks'
+    root.mkdir(parents=True, exist_ok=True)
+    for i, group in enumerate(groups):
+        path = root / f'{i:04d}.json'
+        write_json(path, {'schema_version': 1, 'source_sha256': digest(source_path),
+                         'core': group, 'context_read_only': {
+                             'before': segments[max(0, offset-2):offset],
+                             'after': segments[offset+len(group):offset+len(group)+2]}})
+        out.append({'path': str(path), 'sha256': digest(path), 'segment_ids': [s['id'] for s in group]})
+        offset += len(group)
+    for stale in root.glob('*.json'):
+        if stale.name not in {Path(x['path']).name for x in out}:
+            stale.unlink()
+    return out
+
+
+def timestamp(ms):
+    seconds = ms // 1000
+    return f'{seconds//3600:02d}:{seconds//60%60:02d}:{seconds%60:02d}'
+
+
+def plain(text):
+    # External text is data, including Markdown headings and HTML from videos.
+    return re.sub(r'([\\*_{}\[\]<>#])', r'\\\1', ' '.join(text.split())).replace(chr(96), '\\' + chr(96))
+
+
+def validate_review(source, review, manifest, work_dir):
+    validate_source(source)
+    if not isinstance(review, dict):
+        raise JobError('REVIEW_SCHEMA', 'review 必须为 JSON 对象。')
+    if (review.get('schema_version') != 1 or
+            review.get('source_sha256') != manifest['source_sha256'] or
+            review.get('prompt_version') != PROMPT_VERSION or
+            review.get('contract_version') != CONTRACT_VERSION):
+        raise JobError('REVIEW_VERSION', '校对来源摘要或处理版本不匹配。')
+    rows = review.get('segments', [])
+    if not isinstance(rows, list) or not all(isinstance(r, dict) for r in rows):
+        raise JobError('REVIEW_SCHEMA', 'segments 必须为校对对象数组。')
+    if [r.get('segment_id') for r in rows] != [s['id'] for s in source['segments']]:
+        raise JobError('REVIEW_COVERAGE', '每个来源段必须按顺序恰好映射一次。')
+    kept = {r['segment_id'] for r in rows if r.get('action') != 'remove_duplicate'}
+    by_id = {s['id']: s for s in source['segments']}
+    changes = []
+    for s, r in zip(source['segments'], rows):
+        action, text = r.get('action'), r.get('text_zh')
+        if action not in {'keep', 'correct', 'translate', 'uncertain', 'remove_duplicate'} or not isinstance(text, str):
+            raise JobError('REVIEW_INVALID', '无效的 action 或 text_zh。')
+        if not isinstance(r.get('reason'), str) or (action != 'keep' and not r['reason'].strip()):
+            raise JobError('REVIEW_REASON', '修改必须解释依据。')
+        if action == 'remove_duplicate':
+            record_id = r.get('retranscription_id', '')
+            if not re.fullmatch(r'retry-\d+-\d+', record_id):
+                raise JobError('DUPLICATE_EVIDENCE', '删除重复必须引用局部重识别记录。')
+            p = Path(work_dir) / 'retranscriptions' / (record_id + '.json')
+            evidence = read_json(p) if p.exists() else {}
+            retained = by_id.get(r.get('retained_segment_id'))
+            if (r.get('retained_segment_id') not in kept or not retained or
+                    evidence.get('source_sha256') != manifest['source_sha256'] or
+                    not evidence.get('segments') or
+                    not all(evidence.get('start_ms', 1) <= x['start_ms'] < x['end_ms'] <= evidence.get('end_ms', 0)
+                            for x in (s, retained))):
+                raise JobError('DUPLICATE_EVIDENCE', '证据必须覆盖删除段和保留段且对应当前来源。')
+        elif not text.strip():
+            raise JobError('REVIEW_EMPTY', '正文不能为空。')
+        if action == 'keep' and text != s['text']:
+            raise JobError('REVIEW_ACTION', 'keep 不得修改原文。')
+        if action == 'uncertain' and not re.search(r'\[听辨不清，\d{2,}:[0-5]\d:[0-5]\d\]', text):
+            raise JobError('UNCERTAINTY_MARKER', '不确定段必须保留带时间的听辨不清标记。')
+        before = re.findall(r'\d+(?:[.,]\d+)*|不|没|无|未|别|not\b|never\b|no\b', s['text'], re.I)
+        after = re.findall(r'\d+(?:[.,]\d+)*|不|没|无|未|别|not\b|never\b|no\b', text, re.I)
+        if before != after and action != 'remove_duplicate':
+            if not str(r.get('reason', '')).strip():
+                raise JobError('INVARIANT_EXPLANATION', '数字或否定变化需要解释。')
+            changes.append({'segment_id': s['id'], 'before': before, 'after': after, 'reason': r['reason']})
+    return changes
+
+
+def render(manifest, source, review, guide):
+    if not guide.strip() or re.search(r'^\s*#{1,6}\s|<\s*/?\s*h[1-6]\b', guide, re.M | re.I):
+        raise JobError('GUIDE_STRUCTURE', '导读必须为不带标题的正文。')
+    lines = ['# ' + plain(manifest['title']), '',
+             f"来源：{manifest['canonical_url']}；频道：{plain(manifest.get('channel') or '未知')}；时长：{timestamp(manifest['duration_ms'])}", '']
+    if manifest.get('quality_hints') or any(r['action'] == 'uncertain' for r in review['segments']):
+        lines += ['说明：存在待核对片段；材料完整性与听辨标记请结合原视频核对。', '']
+    lines += ['## 导读', '', plain(guide), '', '## 正文', '']
+    for s, r in zip(source['segments'], review['segments']):
+        if r['action'] != 'remove_duplicate':
+            lines += [f"[{timestamp(s['start_ms'])}] {plain(r['text_zh'])}", '']
+    if not source['segments']:
+        lines += ['未识别到可确认的口述内容。', '']
+    return '\n'.join(lines)

package/payload/plugins/toolkit/skills/youtube-reading/scripts/runtime.py

@@ -0,0 +1,424 @@
+"""Private macOS runtime and bounded network/ASR operations. No shell commands."""
+import importlib.metadata
+import json
+import os
+from pathlib import Path
+import platform
+import re
+import shutil
+import subprocess
+import sys
+import tarfile
+import time
+import uuid
+import wave
+
+from artifacts import JobError, digest, identity, job_lock, now, read_json, write_json
+
+SKILL_ROOT = Path(__file__).resolve().parent.parent
+DEFAULT_ROOT = Path.home() / 'Library/Application Support/youtube-reading'
+LOCK = read_json(SKILL_ROOT / 'assets/runtime-lock.json')
+
+
+def run(args, log, timeout=120):
+    log = Path(log)
+    log.parent.mkdir(parents=True, exist_ok=True)
+    with log.open('a', encoding='utf-8') as f:
+        try:
+            p = subprocess.run([str(a) for a in args], stdout=f, stderr=f, timeout=timeout)
+        except subprocess.TimeoutExpired:
+            raise JobError('PROCESS_TIMEOUT', f'进程超时；日志：{log}')
+    if p.returncode:
+        raise JobError('PROCESS_FAILED', f'进程退出 {p.returncode}；日志：{log}')
+
+
+def active(root):
+    path = Path(root) / 'active.json'
+    if not path.exists():
+        raise JobError('RUNTIME_MISSING', '尚未安装专用运行时，请执行 setup。')
+    release = read_json(path)['release']
+    if not re.fullmatch(r'[a-zA-Z0-9-]+', release):
+        raise JobError('RUNTIME_INVALID', '运行时版本指针无效。')
+    return Path(root) / 'releases' / release
+
+
+def tool_version(tool):
+    path = shutil.which(tool)
+    if not path:
+        return {'path': None, 'version': None}
+    flag = '-version' if tool in {'ffmpeg', 'ffprobe'} else '--version'
+    try:
+        p = subprocess.run([path, flag], capture_output=True, text=True, timeout=15)
+        return {'path': path, 'version': (p.stdout or p.stderr).splitlines()[0]}
+    except (OSError, subprocess.SubprocessError, IndexError):
+        return {'path': path, 'version': None}
+
+
+def doctor(root=DEFAULT_ROOT, release=None):
+    issues, checks = [], {}
+    if platform.system() != 'Darwin' or platform.machine() != 'arm64':
+        issues.append('requires macOS Apple Silicon')
+    if sys.version_info < (3, 10):
+        issues.append('requires Python >=3.10')
+    for tool in ('ffmpeg', 'ffprobe', 'cmake', 'node', 'clang', 'curl'):
+        checks[tool] = tool_version(tool)
+        if not checks[tool]['version']:
+            issues.append('missing tool: ' + tool)
+    for tool, minimum in [('ffmpeg', (7, 0)), ('ffprobe', (7, 0)), ('node', (22, 0)), ('cmake', (3, 21))]:
+        version = checks[tool]['version'] or ''
+        match = re.search(r'(\d+)\.(\d+)', version)
+        if match and tuple(map(int, match.groups())) < minimum:
+            issues.append(f'{tool} requires >= {minimum[0]}.{minimum[1]}')
+    try:
+        release = Path(release) if release else active(root)
+        receipt = read_json(release / 'receipt.json')
+        if receipt['lock_sha256'] != identity(LOCK):
+            issues.append('runtime lock changed; run setup')
+        for rel, expected in receipt['files'].items():
+            file = release / rel
+            if not file.is_file() or digest(file) != expected:
+                issues.append('checksum mismatch: ' + rel)
+        for model in LOCK['models']:
+            file = release / 'models' / model['name']
+            checks[model['name']] = {'sha256': digest(file) if file.exists() else None,
+                                      'expected': model['sha256']}
+            if checks[model['name']]['sha256'] != model['sha256']:
+                issues.append('model checksum mismatch: ' + model['name'])
+        code = 'import importlib.metadata,json; print(json.dumps({n:importlib.metadata.version(n) for n in ' + repr([x['name'] for x in LOCK['python_packages']]) + '}))'
+        p = subprocess.run([str(release / 'venv/bin/python'), '-c', code], capture_output=True, text=True, timeout=20)
+        versions = json.loads(p.stdout) if p.returncode == 0 else {}
+        checks['python_packages'] = versions
+        for x in LOCK['python_packages']:
+            if versions.get(x['name']) != x['version']:
+                issues.append('package version mismatch: ' + x['name'])
+        binary = release / 'whisper/build/bin/whisper-cli'
+        version = subprocess.run([str(binary), '--version'], capture_output=True, text=True, timeout=15)
+        checks['whisper_binary_version'] = (version.stdout + version.stderr).strip()
+        if version.returncode or LOCK['whisper']['version'] not in checks['whisper_binary_version']:
+            issues.append('whisper binary version mismatch')
+        checks['release'] = str(release)
+        checks['whisper'] = LOCK['whisper']
+    except (JobError, OSError, ValueError, KeyError, subprocess.SubprocessError) as e:
+        issues.append(str(e))
+    return {'ok': not issues, 'runtime_root': str(Path(root).resolve()), 'checks': checks, 'issues': issues}
+
+
+def download(url, path, sha256, log):
+    path = Path(path)
+    if path.exists() and digest(path) == sha256:
+        return
+    temp = path.with_suffix(path.suffix + '.part')
+    run(['curl', '-4', '--http1.1', '--fail', '--location', '--silent', '--show-error',
+         '--connect-timeout', '15', '--max-time', '900', '--retry', '2',
+         '--output', temp, url], log, timeout=2750)
+    if digest(temp) != sha256:
+        raise JobError('CHECKSUM_MISMATCH', f'下载摘要不匹配：{path.name}')
+    os.replace(temp, path)
+
+
+def setup(root=DEFAULT_ROOT):
+    root = Path(root).expanduser().resolve()
+    with job_lock(root):
+        existing = doctor(root)
+        if existing['ok']:
+            return {**existing, 'cache_hit': True}
+        missing = [x for x in existing['issues'] if x.startswith(('requires ', 'missing tool:')) or ' requires >= ' in x]
+        if missing:
+            raise JobError('SETUP_PREREQUISITE', '; '.join(missing))
+        release = root / 'releases' / (identity(LOCK)[:12] + '-' + uuid.uuid4().hex[:8])
+        release.mkdir(parents=True)
+        cache = root / 'downloads'
+        cache.mkdir(exist_ok=True)
+        log = release / 'setup.log'
+        run([sys.executable, '-m', 'venv', release / 'venv'], log)
+        for x in LOCK['python_packages']:
+            download(x['url'], cache / x['filename'], x['sha256'], log)
+        run([release / 'venv/bin/python', '-m', 'pip', 'install', '--no-index', '--no-deps',
+             '--require-hashes', '--find-links', cache, '-r', SKILL_ROOT / 'requirements.lock'], log)
+        archive = cache / (LOCK['whisper']['commit'] + '.tar.gz')
+        download(LOCK['whisper']['url'], archive, LOCK['whisper']['sha256'], log)
+        source = release / 'whisper'
+        source.mkdir()
+        with tarfile.open(archive) as tar:
+            # Extract files only, dropping archive prefix and forbidding links/traversal.
+            for member in tar.getmembers():
+                if not member.isfile():
+                    continue
+                rel = Path(*Path(member.name).parts[1:])
+                if rel.is_absolute() or '..' in rel.parts:
+                    raise JobError('ARCHIVE_INVALID', '源码归档含不安全路径。')
+                dest = source / rel
+                dest.parent.mkdir(parents=True, exist_ok=True)
+                with tar.extractfile(member) as src, dest.open('wb') as out:
+                    shutil.copyfileobj(src, out)
+        run(['cmake', '-S', source, '-B', source / 'build', '-DCMAKE_BUILD_TYPE=Release',
+             '-DGGML_METAL=ON', '-DWHISPER_BUILD_TESTS=OFF', '-DWHISPER_BUILD_SERVER=OFF',
+             '-DWHISPER_BUILD_COMMIT=' + LOCK['whisper']['commit']], log, 180)
+        run(['cmake', '--build', source / 'build', '--config', 'Release', '-j', '4'], log, 600)
+        (release / 'models').mkdir()
+        for model in LOCK['models']:
+            download(model['url'], cache / model['name'], model['sha256'], log)
+            shutil.copy2(cache / model['name'], release / 'models' / model['name'])
+        files = {}
+        for file in (source / 'build').rglob('*'):
+            if file.is_file() and (file.suffix == '.dylib' or file.name == 'whisper-cli'):
+                files[str(file.relative_to(release))] = digest(file)
+        write_json(release / 'receipt.json', {'schema_version': 1, 'lock_sha256': identity(LOCK),
+                                             'installed_at': now(), 'files': files})
+        result = doctor(root, release)
+        if not result['ok']:
+            raise JobError('SETUP_VERIFICATION', '; '.join(result['issues']))
+        write_json(root / 'active.json', {'release': release.name})
+        return result
+
+
+def activate_python(root):
+    release = active(root)
+    site = list((release / 'venv/lib').glob('python*/site-packages'))
+    if not site:
+        raise JobError('RUNTIME_MISSING', '缺少 Python 专用环境。')
+    sys.path.insert(0, str(site[0]))
+    return release
+
+
+class NetworkAdapter:
+    """Inject this boundary for isolated failures; production uses real requests."""
+    def __init__(self, root, work):
+        activate_python(root)
+        import requests
+        import urllib3.util.connection
+        urllib3.util.connection.HAS_IPV6 = False
+        class Session(requests.Session):
+            def request(self, *args, **kwargs):
+                kwargs['timeout'] = 15
+                for attempt in range(3):
+                    try:
+                        response = super().request(*args, **kwargs)
+                        response.raise_for_status()
+                        return response
+                    except (requests.Timeout, requests.ConnectionError):
+                        if attempt == 2:
+                            raise
+        self.session = Session()
+        self.work = Path(work)
+
+    def metadata(self, url):
+        import yt_dlp
+        with yt_dlp.YoutubeDL(self.options()) as ydl:
+            return ydl.extract_info(url, download=False)
+
+    def options(self):
+        adapter = self
+        class Logger:
+            def debug(self, msg):
+                pass
+            def warning(self, msg):
+                self.error(msg)
+            def error(self, msg):
+                with (adapter.work / 'network.log').open('a') as f:
+                    f.write(msg + '\n')
+        return {'quiet': True, 'no_warnings': True, 'logger': Logger(), 'noplaylist': True,
+                'source_address': '0.0.0.0', 'socket_timeout': 15, 'retries': 2,
+                'fragment_retries': 2, 'extractor_retries': 2, 'file_access_retries': 2,
+                'js_runtimes': {'node': {'path': shutil.which('node')}},
+                'cachedir': str(self.work / 'yt-cache')}
+
+    def transcripts(self, video_id, language):
+        from youtube_transcript_api import YouTubeTranscriptApi
+        from youtube_transcript_api._errors import TranscriptsDisabled, NoTranscriptFound
+        try:
+            tracks = list(YouTubeTranscriptApi(http_client=self.session).list(video_id))
+        except (TranscriptsDisabled, NoTranscriptFound):
+            return {'tracks': [], 'segments': [], 'language': language, 'source_kind': None}
+        inventory = [{'language': t.language_code, 'generated': t.is_generated} for t in tracks]
+        # A known original language is mandatory; do not choose translated manual tracks.
+        prefix = (language or '').split('-')[0]
+        matching = [t for t in tracks if prefix and t.language_code.split('-')[0] == prefix]
+        matching.sort(key=lambda t: t.is_generated)
+        if not matching:
+            return {'tracks': inventory, 'segments': [], 'language': language, 'source_kind': None}
+        t = matching[0]
+        rows = t.fetch().to_raw_data()
+        return {'tracks': inventory, 'segments': rows, 'language': t.language_code,
+                'source_kind': 'automatic_caption' if t.is_generated else 'manual_caption'}
+
+    def audio(self, url):
+        import yt_dlp
+        options = {**self.options(), 'format': 'bestaudio/best',
+                   'outtmpl': str(self.work / 'audio.%(ext)s')}
+        with yt_dlp.YoutubeDL(options) as ydl:
+            info = ydl.extract_info(url, download=True)
+            return Path(ydl.prepare_filename(info))
+
+
+def acquire_metadata(url, adapter):
+    # Adapters own request retries (three attempts, 15 s each). Do not multiply retries.
+    try:
+        return adapter.metadata(url)
+    except Exception as e:
+        name = type(e).__name__
+        code = 'NETWORK_TIMEOUT' if 'timeout' in name.lower() else 'ACQUISITION_DENIED'
+        raise JobError(code, '获取元数据失败：' + name) from e
+
+
+def convert_audio(audio, work):
+    dest = Path(work) / 'audio.wav'
+    stamp = Path(work) / 'audio-conversion.json'
+    input_hash = digest(audio)
+    if stamp.exists() and dest.exists():
+        saved = read_json(stamp)
+        if saved.get('input_sha256') == input_hash and saved.get('sha256') == digest(dest):
+            return dest
+    temp = dest.with_name('audio.pending.wav')
+    run(['ffmpeg', '-v', 'error', '-y', '-i', audio, '-ar', '16000', '-ac', '1', '-c:a', 'pcm_s16le', temp],
+        Path(work) / 'ffmpeg.log', 600)
+    os.replace(temp, dest)
+    write_json(stamp, {'input_sha256': input_hash, 'sha256': digest(dest)})
+    return dest
+
+
+def audio_duration(path):
+    with wave.open(str(path), 'rb') as wav:
+        return round(wav.getnframes() / wav.getframerate() * 1000)
+
+
+def transcribe_audio(audio_path, work_dir, options):
+    """Same resumable core for CLI and local audio acceptance. options: runtime_root,
+    language (auto default), start_ms/end_ms, vad, timeout_seconds, on_chunk callback.
+    Each core <=600 s; reconcile unowned boundary candidates after all cores finish.
+    """
+    work = Path(work_dir)
+    work.mkdir(parents=True, exist_ok=True)
+    release = active(options.get('runtime_root', DEFAULT_ROOT))
+    duration = audio_duration(audio_path)
+    start = options.get('start_ms', 0)
+    end = min(options.get('end_ms', duration), duration)
+    if not 0 <= start < end:
+        raise JobError('AUDIO_RANGE', '无效音频时间区间。')
+    settings = {k: v for k, v in options.items() if k not in {'on_chunk', 'runtime_root'}}
+    key = identity({'audio_sha256': digest(audio_path), 'lock': identity(LOCK),
+                    'settings': settings, 'boundary_version': 1})
+    result, boundary_candidates = [], []
+    for core_start in range(start, end, 600000):
+        core_end = min(core_start + 600000, end)
+        context = options.get('context_ms', 2000)
+        clip_start = max(0, core_start - context)
+        clip_end = min(duration, core_end + context)
+        name = f'{core_start}-{core_end}'
+        directory = work / key[:16]
+        directory.mkdir(exist_ok=True)
+        receipt = directory / (name + '.done.json')
+        if receipt.exists():
+            saved = read_json(receipt)
+            output = directory / (name + '.json')
+            if saved.get('key') == key and output.exists() and digest(output) == saved['sha256']:
+                result.extend(saved['segments'])
+                boundary_candidates.extend(saved['boundary_candidates'])
+                if options.get('on_chunk'):
+                    options['on_chunk'](saved, True)
+                continue
+        clip = directory / (name + '.wav')
+        with wave.open(str(audio_path), 'rb') as src, wave.open(str(clip), 'wb') as out:
+            out.setparams(src.getparams())
+            src.setpos(round(clip_start * src.getframerate() / 1000))
+            out.writeframes(src.readframes(round((clip_end-clip_start) * src.getframerate() / 1000)))
+        output_prefix = directory / name
+        timeout = options.get('timeout_seconds') or max(120, (clip_end-clip_start) / 2000)
+        args = [release / 'whisper/build/bin/whisper-cli', '-m', release / 'models' / LOCK['models'][0]['name'],
+                '-f', clip, '-l', options.get('language', 'auto'), '-t', '4', '-oj', '-of', output_prefix]
+        if options.get('vad', True):
+            args += ['--vad', '-vm', release / 'models' / LOCK['models'][1]['name']]
+        started = time.monotonic()
+        try:
+            run(args, directory / (name + '.log'), timeout)
+        except JobError as e:
+            if e.code == 'PROCESS_TIMEOUT':
+                raise JobError('ASR_TIMEOUT', str(e)) from e
+            raise
+        raw = read_json(directory / (name + '.json'))
+        rows, candidates = [], []
+        for item in raw.get('transcription', []):
+            a = clip_start + item['offsets']['from']
+            b = min(clip_end, clip_start + item['offsets']['to'])
+            text = item['text'].strip()
+            if text and b > a and a < core_end and b > core_start:
+                target = rows if core_start <= (a+b)/2 < core_end else candidates
+                target.append({'start_ms': a, 'end_ms': b, 'text': text,
+                               'language': raw.get('result', {}).get('language', options.get('language', 'auto')),
+                               'source_kind': 'ASR', 'core_start_ms': core_start,
+                               'raw_path': str(directory / (name + '.json'))})
+        saved = {'key': key, 'sha256': digest(directory / (name + '.json')), 'segments': rows,
+                 'boundary_candidates': candidates,
+                 'core_start_ms': core_start, 'core_end_ms': core_end, 'timeout_seconds': timeout,
+                 'elapsed_seconds': round(time.monotonic()-started, 3), 'completed_at': now()}
+        write_json(receipt, saved)
+        result.extend(rows)
+        boundary_candidates.extend(candidates)
+        if options.get('on_chunk'):
+            options['on_chunk'](saved, False)
+    # Independent timestamps can put the same sentence outside BOTH owning
+    # cores. Retain those candidates until all cores are available; restoring
+    # a candidate also covers its counterpart, so it is not restored twice.
+    for candidate in sorted(boundary_candidates, key=lambda x: x['end_ms']-x['start_ms'], reverse=True):
+        midpoint = (candidate['start_ms'] + candidate['end_ms']) / 2
+        if start <= midpoint < end and not any(s['start_ms'] <= midpoint < s['end_ms'] for s in result):
+            result.append({**candidate, 'boundary_recovery': True})
+    result.sort(key=lambda x: x['start_ms'])
+    # Only trim a literal repeated boundary prefix when different core windows
+    # claim the same audio time. Keep raw text/times as provenance. Ordinary
+    # repetitions within one window or at distinct times remain untouched.
+    for previous, current in zip(result, result[1:]):
+        overlap = previous['end_ms'] - current['start_ms']
+        if previous['core_start_ms'] == current['core_start_ms'] or overlap < 500:
+            continue
+        left = re.sub(r'[\s，,。.!！?？]', '', previous['text'])
+        right = current['text']
+        for n in range(min(len(left), len(right)), 3, -1):
+            if left.endswith(right[:n]) and n < len(right):
+                current['boundary_overlap'] = {
+                    'original_text': right, 'removed_prefix': right[:n],
+                    'overlap_ms': overlap, 'retained_raw_path': previous['raw_path']}
+                current['text'] = right[n:].lstrip('，,。 ')
+                current['start_ms'] = previous['end_ms']
+                break
+    if options.get('vad', True) and options.get('recover_gaps', True):
+        # Language can change mid-clip. Re-detect language in uncovered intervals
+        # using VAD, once only; silence remains empty rather than hallucinated.
+        covered = [{'start_ms': max(start, x['start_ms'])-start,
+                    'end_ms': min(end, x['end_ms'])-start} for x in result]
+        for gap_start, gap_end in caption_gaps(covered, end-start):
+            recovered = transcribe_audio(audio_path, work / 'coverage', {
+                **options, 'start_ms': start+gap_start, 'end_ms': start+gap_end,
+                'language': 'auto', 'context_ms': 0, 'recover_gaps': False})
+            for segment in recovered:
+                segment['coverage_recovery'] = True
+            result.extend(recovered)
+        result.sort(key=lambda x: x['start_ms'])
+    return result
+
+
+def caption_gaps(segments, duration_ms):
+    cursor, gaps = 0, []
+    for s in sorted(segments, key=lambda x: x['start_ms']):
+        if s['start_ms']-cursor > 3000:
+            gaps.append((cursor, s['start_ms']))
+        cursor = max(cursor, s['end_ms'])
+    if duration_ms-cursor > 3000:
+        gaps.append((cursor, duration_ms))
+    return gaps
+
+
+def fill_gaps(segments, audio, work, options):
+    gaps = caption_gaps(segments, audio_duration(audio))
+    additions = []
+    for a, b in gaps:
+        rows = transcribe_audio(audio, Path(work) / 'gaps', {
+            **options, 'start_ms': a, 'end_ms': b, 'vad': True, 'context_ms': 0})
+        # Feed only missing audio: clipping timestamps cannot remove words
+        # spoken in neighbouring captions. Keep ASR timestamps as provenance.
+        for r in rows:
+            r = {**r, 'source_kind': 'gap_fill'}
+            if r['end_ms'] > r['start_ms']:
+                additions.append(r)
+    write_json(Path(work) / 'gap-check.json', {'schema_version': 1, 'candidates': gaps, 'additions': additions})
+    return sorted(segments + additions, key=lambda x: x['start_ms'])