Peach: Local-first video metadata scraper and media manager.
fix(media): 修复尾部索引 MP4 的播放头与旧失败缓存
simon15614e950fb2f23d
diff --git a/CHANGELOG.md b/CHANGELOG.mdindex 07889f5b1..1c42dbaee 100644--- a/CHANGELOG.md+++ b/CHANGELOG.md@@ -23,6 +23,7 @@ Removed / Fixed / Security。### 修复+- **播放**:缺少显示时间戳表的 MP4 可在索引位于文件末尾时重建播放头;旧版失败缓存会重新检查这类布局,无法判断编辑起点的文件继续转码。- **配置**:同源检查统一识别 HTTP、HTTPS 的默认端口,避免地址显式包含 `:80` 或 `:443` 时被误拒绝。- **馆藏**:处理失败后的线程清理完成前保留停止中状态,避免立即重试撞上仍持有的处理锁;起步失败时显示本次错误,不再退回上一轮结果。- **采集**:云下载能力未取得时,在本机诊断记录保留令牌、挂载点或目录探测阶段及失败分类,便于区分配置与网络问题。diff --git a/docs/REUSE.md b/docs/REUSE.mdindex 80ed67580..0109dcca5 100644--- a/docs/REUSE.md+++ b/docs/REUSE.md@@ -815,8 +815,9 @@ A 盘垃圾隔离使用短名称 `peach-purge-<32 位 UUID>.peach-quarantine`,- 远程媒体持久缓存复用 HTTPX 0.28.1 的流式响应、Starlette 1.7.0 的文件 Range、filelock 4.0.12 和 `fsutil.atomic_path`;缓存身份、失败保留与按需取得由 Peach 管理。HLS 清单解析用 [m3u8 6.0.0](https://pypi.org/project/m3u8/)(MIT,Python ≥3.7,纯 Python、24.1 KB wheel、无传递依赖),已读取固定版 `model.py` 并验证嵌套清单、密钥、初始化段和字节范围。Hishel 1.4.0(BSD-3-Clause)已读取 `_sync_cache.py` 与 Range 判据,通用 HTTP 缓存按来源 URL、缓存头和重新验证工作;此处需按稳定条目身份避开签名 URL 更新、永久保留成功媒体,因此复用现有持久文件路径。规则及验收见 [媒体缓存](MEDIA_CACHE.md)。- 账本路径在 Windows 只做 `abspath`,不 `resolve()`:PikPak 的 A: 是 WinFsp 网络驱动器,`resolve()` 得到 UNC 形态,之后每个经 `file_for()` 的播放请求都慢几个数量级;115 的 B: 不受影响。- 不兼容片源(HEVC、mp3 以外的音轨、非 MP4 容器)按 6 秒片重编码给 HLS,账本没记时长或为负时用 ffprobe 的时长切,探测也拿不到才回 Range。分片与整片共享硬件后端探测及回退链,见[硬件转码后端](#硬件转码后端),同一分片只起一个 FFmpeg。Range 响应按 1 MiB 读文件,`BufferedFileResponse` 自己盯 `http.disconnect`(uvicorn 断开后 `send()` 静默返回,Starlette `FileResponse` 不监听),否则拖一次进度条就留下一个把整部片拉完的幽灵读者。-- 有 B 帧却没有 `ctts` 的 MP4 是时间戳错乱,浏览器丢掉倒着走的帧。这类片源不重编码:`mp4repair.py` 重建一份 `moov`(游程编码的 `ctts`、编辑列表补整体平移、`stco`/`co64` 按头长差平移)存成 `transcode_root` 里的 `.mp4hdr` 边车,`/stream` 用「边车的头 + 原文件那段 mdat」按 Range 发。显示顺序取自 `ffprobe -ignore_editlist 1 -show_entries frame=pts`,`pkt_dts` 不能用。边车算出来前走 HLS 转码,同一部片后台只算一次,算不出就跳过。- 编辑列表的半帧舍入判据采用 peach-patcher 的 [PR #175](https://github.com/peach-mitao/peach/pull/175)(`4106d30cc740863cd1d44a207f80c5f75291141f`),按最短正帧间距判断已有补偿,保留真实裁切、空编辑和段速率。新版边车标识为 `PEACHMP4` 第 2 版;第 1 版在读取时按偏移游程的加权均值还原整体平移,仅修正重复补偿,不读原片、不展开整片时间轴、不重写边车或媒体。已有无效边车可在所选媒体库的修复任务中重建,其旧失败结论不再遮蔽这项修复;正常边车、正常扫描结论和无产物的既有解码失败继续复用。临时旧边车、批处理和 Range 拼接均有回归;未新增依赖。+- 有 B 帧却没有 `ctts` 的 MP4 是时间戳错乱,浏览器丢掉倒着走的帧。这类片源不重编码:`mp4repair.py` 重建一份 `moov`(游程编码的 `ctts`、编辑列表补整体平移、`stco`/`co64` 按头长差平移)存成 `transcode_root` 里的 `.mp4hdr` 边车,`/stream` 用「边车的头 + 原文件那段 mdat」按 Range 发。`moov` 位于媒体前后都支持;位于末尾时只读索引尾段,把它放到媒体前,媒体后的其他盒子原样保留。显示顺序取自 `ffprobe -ignore_editlist 1 -show_entries frame=pts`,`pkt_dts` 不能用。边车算出来前走 HLS 转码,同一部片后台只算一次,算不出就跳过。+- 尾部索引布局采用 simon1561 的 [PR #157](https://github.com/peach-mitao/peach/pull/157)(`e67ead5a35a63fc36aa8153e5d62998a001779db`),补充扩展盒头、单个媒体盒和索引读取边界。非零编辑起点超过重建补偿范围时无法区分预卷与裁剪,保留转码;临时 B 帧视频已验证适用样本 50 帧逐帧一致、歧义样本拒修。扫描缓存第 1 版的失败结论按尾部布局重新诊断,第 2 版保留当前失败结论,不清空正常缓存或自动运行媒体修复。- 「只采集在线资料」对齐全的行(番号已落库、字段有着落或已在候选表、封面在位)不碰磁盘;文件在不在看目录列表、同目录只列一次,扫描用目录列表自带的大小与时间、不逐个 stat;候选 CSV 每 5 秒写盘一次,被打断也在收尾写全。- 一个控件在两页出现时,选中态也归它:首页、资料页与关注页的「全部/没看过/稍后看/已标记」共用 `use-view-glide.ts` 那块滑动玻璃与 `[data-view-glide]`,填充只由玻璃给,各排自己不铺底。- 本地详情与关注详情的「接着看」共用 `RelatedSection`、横滚接线与卡片 `next` 版式,数量读同一设置。空结果复用 `EmptyState`,读取失败复用 `RetryNote`,可原位重试;关闭推荐时不取数。关注推荐复用 `/api/follow` 分页及 TanStack Query 缓存,最多读取三页;优先同创作者和共同标签,排除当前合集、已看、忽略与不同媒体类型。创作者头像链接复用卡片的 hover,下载链接复用舞台操作键几何;未新增依赖,推荐读取不写 ledger。diff --git a/src/peach/mp4repair.py b/src/peach/mp4repair.pyindex 0f21f3f74..cc94d6ff4 100644--- a/src/peach/mp4repair.py+++ b/src/peach/mp4repair.py@@ -297,11 +297,22 @@ def _shift_chunk_offsets(box: _Box, delta: int) -> None:_Span = tuple[bytes, int, int, int]-def _layout(source: Path) -> tuple[bytes, _Span, _Span]:- """读出 `mdat` 之前那一整段,并确认这个布局拼得回去。+@dataclass(frozen=True)+class _Layout:+ before: bytes+ mdat_header: bytes+ suffix: bytes+ mdat: _Span+ moov_box: bytes+ moov_body: int+ moov_before_start: int | None+++def _layout(source: Path) -> _Layout:+ """读出 `mdat` 和 `moov`,返回拼回虚拟文件所需的几段。- 拼接只换头、原样接上 `mdat`,所以 `mdat` 必须是最后一个盒子,前面的东西也得- 一次读得完。不满足就抛 `RepairUnavailable`,让调用方回到转码那条路。+ faststart 文件把 `moov` 放在媒体前面,尾部只允许没有盒子;下载器常见的布局则把+ `moov` 放在末尾。后一种只多读媒体之后的索引段,媒体仍按原字节流发送,不复制原片。"""from .smb_media import SmbResourcewith (source.open() if isinstance(source, SmbResource) else open(source, 'rb')) as handle:@@ -309,22 +320,44 @@ def _layout(source: Path) -> tuple[bytes, _Span, _Span]:kinds = {kind for kind, *_ in boxes}if not {b"moov", b"mdat"} <= kinds:raise RepairUnavailable("顶层缺 moov 或 mdat")+ if any(sum(box[0] == kind for box in boxes) != 1 for kind in (b"moov", b"mdat")):+ raise RepairUnavailable("顶层必须各有一个 moov 和 mdat")mdat = next(box for box in boxes if box[0] == b"mdat")moov = next(box for box in boxes if box[0] == b"moov")- if moov[1] > mdat[1]:- raise RepairUnavailable("moov 在 mdat 之后,这个布局还没支持")- if any(box[1] > mdat[1] for box in boxes):- raise RepairUnavailable("mdat 后面还有盒子,这个布局还没支持")if mdat[3] - mdat[2] <= 0 or moov[3] - moov[1] > HEADER_LIMIT:raise RepairUnavailable("moov 太大或 mdat 是空的")+ if mdat[2] > HEADER_LIMIT:+ raise RepairUnavailable("媒体前的索引段太大")+ if moov[1] < mdat[1]:+ if any(box[1] > mdat[1] for box in boxes):+ raise RepairUnavailable("mdat 后面还有盒子,这个布局还没支持")+ handle.seek(moov[1])+ moov_box = handle.read(moov[3] - moov[1])+ handle.seek(0)+ before = handle.read(mdat[1])+ handle.seek(mdat[1])+ mdat_header = handle.read(mdat[2] - mdat[1])+ return _Layout(before, mdat_header, b"", mdat, moov_box, moov[2] - moov[1], moov[1])+ after_size = handle.seek(0, 2) - mdat[3]+ if after_size > HEADER_LIMIT - mdat[2]:+ raise RepairUnavailable("媒体后的索引段太大")handle.seek(0)- return handle.read(mdat[2]), moov, mdat+ before = handle.read(mdat[1])+ handle.seek(mdat[1])+ mdat_header = handle.read(mdat[2] - mdat[1])+ handle.seek(mdat[3])+ after = handle.read(after_size)+ start = moov[1] - mdat[3]+ stop = moov[3] - mdat[3]+ return _Layout(before, mdat_header, after[:start] + after[stop:], mdat,+ after[start:stop], moov[2] - moov[1], None)def repaired_header(source: Path, ffprobe: Path) -> RepairedHeader:"""算出修好的头。修不了就抛 `RepairUnavailable`。"""- head, moov, mdat = _layout(source)- tree = _parse(head, moov[2], moov[3])+ layout = _layout(source)+ mdat = layout.mdat+ tree = _parse(layout.moov_box, layout.moov_body, len(layout.moov_box))root = _Box(b"moov", None, tree)traks = [child for child in root.children if child.kind == b"trak"]video = next((trak for trak in traks if _is_video(trak)), None)@@ -346,6 +379,13 @@ def repaired_header(source: Path, ffprobe: Path) -> RepairedHeader:stbl.children.insert(stbl.children.index(stts) + 1, _Box(b"ctts", _ctts(offsets), []))edts = video.find(b"edts")elst = edts.find(b"elst") if edts else None+ if layout.moov_before_start is None and elst and elst.payload:+ def checked_start(value: int) -> int:+ if value > shift + _half_frame(decode_times):+ raise RepairUnavailable("尾部索引的编辑起点无法区分预卷与裁剪")+ return value++ _edit_times(elst.payload, checked_start)payload = _retimed_edit_list(elst.payload if elst and elst.payload else None, shift, _track_duration(root, video),_half_frame(decode_times))@@ -356,14 +396,31 @@ def repaired_header(source: Path, ffprobe: Path) -> RepairedHeader:else:video.children.insert(1, _Box(b"edts", None, [_Box(b"elst", payload, [])]))+ def assemble(rebuilt: bytes) -> bytes:+ if layout.moov_before_start is not None:+ stop = layout.moov_before_start + len(layout.moov_box)+ return (layout.before[:layout.moov_before_start] + rebuilt+ + layout.before[stop:] + layout.mdat_header)+ return layout.before + rebuilt + layout.mdat_header+rebuilt = root.serialize()- delta = len(rebuilt) - (moov[3] - moov[1])+ delta = len(assemble(rebuilt)) - mdat[2]_shift_chunk_offsets(root, delta)rebuilt = root.serialize()- prefix = head[:moov[1]] + rebuilt + head[moov[3]:]+ prefix = assemble(rebuilt)if len(prefix) - mdat[2] != delta:raise RepairUnavailable("重建后的头长度对不上")- return RepairedHeader(prefix=prefix, payload_start=mdat[2], payload_end=mdat[3])+ return RepairedHeader(prefix=prefix, payload_start=mdat[2], payload_end=mdat[3], suffix=layout.suffix)+++def has_tail_moov(source: Path) -> bool:+ """只读顶层盒子位置,识别旧布局判据会跳过的尾部索引。"""+ from .smb_media import SmbResource+ with (source.open() if isinstance(source, SmbResource) else open(source, 'rb')) as handle:+ boxes = _top_level(handle)+ moovs = [box for box in boxes if box[0] == b'moov']+ mdats = [box for box in boxes if box[0] == b'mdat']+ return len(moovs) == len(mdats) == 1 and moovs[0][1] > mdats[0][1]def write_sidecar(path: Path, header: RepairedHeader) -> None:diff --git a/src/peach/web_media_repair.py b/src/peach/web_media_repair.pyindex e275e5507..ed218b890 100644--- a/src/peach/web_media_repair.py+++ b/src/peach/web_media_repair.py@@ -24,7 +24,7 @@ import subprocessimport timefrom pathlib import Path, PureWindowsPath-from . import media_libraries, media_probe, mp4recover, settings_file+from . import media_libraries, media_probe, mp4recover, mp4repair, settings_filefrom .mp4repair import RepairUnavailablefrom .platform import translate_ledger_path@@ -34,6 +34,7 @@ SCAN_CACHE_NAME = "mp4-repair-scan.json"#: 文件身份对应的判定缓存;修复产物由边车或新文件核验。CLEAN = "clean"UNAVAILABLE = "unavailable"+LEGACY_UNAVAILABLE = "unavailable-layout-v1"#: 缺时间戳表,修的是边车。HEADER = "header"@@ -53,7 +54,10 @@ def _load_cache(contract) -> dict[str, str]:except (OSError, ValueError):return {}entries = payload.get("entries") if isinstance(payload, dict) else None- return {str(key): str(value) for key, value in entries.items()} if isinstance(entries, dict) else {}+ if not isinstance(entries, dict):+ return {}+ return {str(key): LEGACY_UNAVAILABLE if value == UNAVAILABLE and payload.get('version', 1) == 1+ else str(value) for key, value in entries.items()}def _save_cache(contract, entries: dict[str, str]) -> None:@@ -61,7 +65,7 @@ def _save_cache(contract, entries: dict[str, str]) -> None:try:path.parent.mkdir(parents=True, exist_ok=True)temporary = path.with_name(path.name + ".tmp")- temporary.write_text(json.dumps({"version": 1, "entries": entries}), encoding="utf-8")+ temporary.write_text(json.dumps({"version": 2, "entries": entries}), encoding="utf-8")temporary.replace(path)except OSError:# 结论缓存丢了只是下次重扫一遍,不该让整轮任务失败。@@ -117,6 +121,15 @@ def _survey(contract, job, job_id: str, rows, cache: dict[str, str]) -> list[tup_advance(job, job_id, index, skipped=True)continuekind = cache.get(key)+ if kind == LEGACY_UNAVAILABLE:+ try:+ kind = None if mp4repair.has_tail_moov(source) else UNAVAILABLE+ except OSError:+ _advance(job, job_id, index, skipped=True)+ continue+ if kind == UNAVAILABLE:+ cache[key] = kind+ _save_cache(contract, cache)if kind == UNAVAILABLE and store.has_invalid_sidecar(asset_id, source):kind = HEADERif kind in (CLEAN, UNAVAILABLE):diff --git a/tests/test_mp4repair.py b/tests/test_mp4repair.pyindex b6ddcb963..bd57265ec 100644--- a/tests/test_mp4repair.py+++ b/tests/test_mp4repair.py@@ -81,6 +81,20 @@ def legacy_sidecar(path: Path, rows: list[tuple[int, int, int]], *, version=0) -return header+def tail_mp4(**options) -> bytes:+ """把正常布局里的 moov 搬到媒体后,并把块偏移改到新的媒体起点。"""+ blob = minimal_mp4(timescale=1000, sample_delta=400, samples=len(DECODE_TO_DISPLAY),+ keyframe_every=6, **options)+ moov_span = next((body - 8, stop) for kind, body, stop in _iter_boxes(blob, 0, len(blob))+ if kind == b"moov")+ start, stop = moov_span+ moov = blob[start:stop]+ tree = mp4repair._parse(moov, 8, len(moov))+ root = mp4repair._Box(b"moov", None, tree)+ mp4repair._shift_chunk_offsets(root, -len(moov))+ return blob[:start] + blob[stop:] + root.serialize()++class Mp4HeaderRepairTests(unittest.TestCase):def setUp(self) -> None:self._temporary = TemporaryDirectory()@@ -132,6 +146,87 @@ class Mp4HeaderRepairTests(unittest.TestCase):self.assertEqual(after[-1][0],header.size - SAMPLE_SIZE, "最后一个块没落在文件末尾那一段上")+ def test_a_tail_moov_is_moved_ahead_of_the_media_without_rewriting_samples(self):+ source = self.root / "tail.mp4"+ source.write_bytes(tail_mp4())+ header = self.repair(source)+ blob = source.read_bytes()++ self.assertEqual(blob[header.payload_start:header.payload_end],+ sample_bytes(len(DECODE_TO_DISPLAY)))+ self.assertEqual(header.suffix, b"")+ moov = moov_of(header.prefix)+ self.assertEqual(next(_iter_boxes(moov, 0, len(moov)))[0], b"mvhd")++ def test_tail_layout_chunk_offsets_point_into_the_spliced_media(self):+ source = self.root / "tail.mp4"+ source.write_bytes(tail_mp4())+ header = self.repair(source)+ after = entries(table(moov_of(header.prefix), b"stco"), ">I")++ self.assertEqual(after[0][0], len(header.prefix))+ self.assertEqual(after[-1][0], header.size - SAMPLE_SIZE)++ def test_other_tail_boxes_stay_after_the_media_when_moov_moves(self):+ source = self.root / "tail.mp4"+ media = tail_mp4()+ moov_start = media.rfind(b"moov") - 4+ source.write_bytes(media[:moov_start] + b"\x00\x00\x00\x10freeABCDEFGH" + media[moov_start:])+ header = self.repair(source)++ self.assertEqual(header.suffix, b"\x00\x00\x00\x10freeABCDEFGH")++ def test_tail_layout_moves_64_bit_chunk_offsets_with_the_media(self):+ blob = tail_mp4()+ start = blob.rfind(b'moov') - 4+ root = mp4repair._Box(b'moov', None, mp4repair._parse(blob[start:], 8, len(blob) - start))+ video = next(child for child in root.children if child.kind == b'trak')+ stco = mp4repair._sample_table(video).find(b'stco')+ offsets = entries(stco.payload, '>I')+ stco.kind = b'co64'+ stco.payload = stco.payload[:8] + b''.join(struct.pack('>Q', value) for value, in offsets)+ source = self.root / 'co64.mp4'+ source.write_bytes(blob[:start] + root.serialize())++ header = self.repair(source)++ self.assertEqual(entries(table(moov_of(header.prefix), b'co64'), '>Q'),+ [(len(header.prefix) + index * SAMPLE_SIZE,)+ for index in range(len(DECODE_TO_DISPLAY))])++ def test_extended_size_tail_moov_uses_its_actual_payload_start(self):+ blob = tail_mp4()+ start = blob.rfind(b'moov') - 4+ moov = blob[start:]+ source = self.root / 'extended.mp4'+ source.write_bytes(blob[:start] + struct.pack('>I4sQ', 1, b'moov', len(moov) + 8) + moov[8:])+ header = self.repair(source)+ self.assertEqual(entries(table(moov_of(header.prefix), b'stco'), '>I')[0][0], len(header.prefix))+ normal = self.broken_mp4(self.root / 'normal.mp4')+ self.assertEqual(composition_times(moov_of(header.prefix), 400),+ composition_times(moov_of(self.repair(normal).prefix), 400))++ def test_tail_layout_requires_one_media_box_and_bounded_metadata(self):+ source = self.root / 'tail.mp4'+ source.write_bytes(tail_mp4() + b'\x00\x00\x00\x0cmdatDATA')+ with self.assertRaisesRegex(RepairUnavailable, '各有一个'):+ self.repair(source)+ source.write_bytes(tail_mp4())+ with patch.object(mp4repair, 'HEADER_LIMIT', 24):+ with self.assertRaises(RepairUnavailable):+ self.repair(source)++ def test_ambiguous_tail_edit_start_keeps_the_transcode_fallback(self):+ source = self.root / 'tail.mp4'+ source.write_bytes(tail_mp4(edit_list=(5200, 800)))+ with self.assertRaisesRegex(RepairUnavailable, '预卷与裁剪'):+ self.repair(source)+ for start in (0, 400):+ with self.subTest(start=start):+ source.write_bytes(tail_mp4(edit_list=(5200, start)))+ self.assertEqual(entries(edit_list_of(self.repair(source)), '>IiI'),+ [(5200, 400, 1 << 16)])+def test_an_edit_list_that_already_compensates_the_shift_is_left_alone(self):"""封装器为 ctts 补过的那一刀要原样留着,再补一次会真的切掉开头那两帧。"""source = self.broken_mp4(self.root / "broken.mp4", edit_list=(5200, 400))diff --git a/tests/test_streaming.py b/tests/test_streaming.pyindex 15bb7f115..a42520d73 100644--- a/tests/test_streaming.py+++ b/tests/test_streaming.py@@ -15,7 +15,7 @@ from peach.streaming import CancellableFileResponse, StreamSessionRegistryfrom peach.mp4index import segment_planfrom peach.routes_media import hls_segmentfrom peach.segments import HlsSegmentService, build_hls_playlist-from test_mp4repair import edit_list_of, entries, legacy_sidecar+from test_mp4repair import DECODE_TO_DISPLAY, edit_list_of, entries, legacy_sidecar, tail_mp4async def _collect(response, requested: bytes | None, method: str = "GET"):@@ -329,6 +329,28 @@ class StreamSessionRegistryTests(unittest.IsolatedAsyncioTestCase):self.assertEqual(status_code, 416)self.assertEqual(headers["content-range"], f"bytes */{len(whole)}")+ async def test_tail_moov_repair_serves_both_seams_and_the_saved_suffix(self):+ from peach import mp4repair+ from peach.streaming import SplicedMp4Response++ with tempfile.TemporaryDirectory() as tmp:+ path = Path(tmp) / 'tail.mp4'+ path.write_bytes(tail_mp4() + b'\x00\x00\x00\x10freeABCDEFGH')+ original = path.read_bytes()+ with patch.object(mp4repair, '_display_order', return_value=list(DECODE_TO_DISPLAY)):+ header = mp4repair.repaired_header(path, Path('ffprobe'))+ sidecar = Path(tmp) / 'repair.mp4hdr'+ mp4repair.write_sidecar(sidecar, header)+ header = mp4repair.read_sidecar(sidecar)+ whole = header.prefix + original[header.payload_start:header.payload_end] + header.suffix+ for seam in (len(header.prefix), len(whole) - len(header.suffix)):+ status, headers, body = await _collect(+ SplicedMp4Response(path, header), f'bytes={seam - 4}-{seam + 4}'.encode())+ self.assertEqual(status, 206)+ self.assertEqual(body, whole[seam - 4:seam + 5])+ self.assertEqual(headers['content-length'], '9')+ self.assertEqual(path.read_bytes(), original)+async def test_a_spliced_head_request_reports_the_size_without_a_body(self):import tempfilefrom pathlib import Pathdiff --git a/tests/test_web_media_repair.py b/tests/test_web_media_repair.pyindex e2ea8be59..7d4e35e50 100644--- a/tests/test_web_media_repair.py+++ b/tests/test_web_media_repair.py@@ -4,6 +4,7 @@修好之后文件和账本落成什么样。重建头的算术在 `test_mp4repair` 里,重建索引的判据在`test_mp4recover` 里。"""+import jsonimport sqlite3import tempfileimport unittest@@ -18,6 +19,7 @@ from peach.mp4repair import RepairUnavailablefrom peach import mp4repairfrom support.ledger import fresh_ledger+from test_mp4repair import tail_mp4SEP = chr(92)#: 顶层只有 `ftyp` 和写到文件末尾的 `mdat`:下载没写完、缺整个索引的那一类。@@ -264,6 +266,40 @@ class MediaRepairRunTests(unittest.TestCase):saved = rm_repair._load_cache(self.contract)self.assertEqual(saved[rm_repair._key(3, self.media / 'fine.mp4')], rm_repair.CLEAN)+ def test_old_failed_tail_layout_is_rechecked_without_rescanning_other_files(self):+ (self.media / 'favourite.mp4').write_bytes(tail_mp4())+ entries = {rm_repair._key(asset_id, self.media / name): verdict+ for asset_id, name, verdict in (+ (1, 'seldom.mp4', rm_repair.UNAVAILABLE),+ (2, 'favourite.mp4', rm_repair.UNAVAILABLE),+ (3, 'fine.mp4', rm_repair.CLEAN),+ (6, 'cut.mp4', rm_repair.UNAVAILABLE))}+ path = rm_repair._cache_path(self.contract)+ path.parent.mkdir(parents=True, exist_ok=True)+ path.write_text(json.dumps({'version': 1, 'entries': entries}), encoding='utf-8')++ state = self.run_once()++ self.assertEqual(self.transcodes.asked, ['favourite.mp4'])+ self.assertEqual(self.store.repaired, [2])+ self.assertEqual((state['repaired'], state['failed']), (1, 0))+ saved = json.loads(path.read_text(encoding='utf-8'))+ self.assertEqual(saved['version'], 2)+ self.assertEqual(saved['entries'][rm_repair._key(1, self.media / 'seldom.mp4')],+ rm_repair.UNAVAILABLE)++ def test_a_current_failed_tail_layout_stays_cached(self):+ (self.media / 'favourite.mp4').write_bytes(tail_mp4())+ cache = {rm_repair._key(asset_id, self.media / name): rm_repair.UNAVAILABLE+ for asset_id, name in ((1, 'seldom.mp4'), (2, 'favourite.mp4'),+ (3, 'fine.mp4'), (6, 'cut.mp4'))}+ rm_repair._save_cache(self.contract, cache)++ self.run_once()++ self.assertEqual(self.transcodes.asked, [])+ self.assertEqual(self.store.repaired, [])+def test_a_changed_file_is_judged_again(self):"""结论跟着文件的大小和改动时间走:换了内容,上一轮的结论就不作数。"""self.run_once()