fix(ceus): 降低长视频内存峰值，快照开跑即写，build 核对 CEUS 运行与交付包

- 内存：prepare_segmentation 改为流式解码，只留分屏和配准的抽样帧，绿框逐帧在两半上找；
  分割那一半由 SegmentationInput.target_frames() 到用时再解码，segment_video 边读边写 jpg；
  MedSAM2 init_state 的读帧换成直接在存放设备上建张量（不改 MedSAM2 源码，结果逐位相同）；
  传播结果写进预分配数组。帧数最多的一段视频常驻内存峰值 11.2 GiB → 2.3 GiB，分辨率最大的
  一段原代码被 OOM 杀掉、现为 4.2 GiB；两段的掩码和全部产物与改前逐字节相同，33 段视频的
  第 1–2 步结果与改前一致，29 例配准与 review.csv 一致。
- 快照：ceus_quant 开跑时写 snapshot/（git、ceus.yaml、review.csv、uv.lock、segment.json），
  配置和核对表从快照副本读；收尾由新函数 runs.finish_snapshot 只新建 finished.json。
  build 与 ceus_align_review 同样改用 finish_snapshot 记结束时间。
- build：summary.json 加 video_sha256；装 CEUS 结果时核对快照里的 delivery_dir、逐例核对
  视频 SHA-256 和推注时间，不同就整次停下；新字段做类型检查，不进 tic.json。

来源：主会话任务（问题 1–3 及补充的 delivery_dir 比对）。

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

scripts/ceus_align_review.py

@@ -23,7 +23,7 @@ import numpy.typing as npt
 from cadx import ceus
 from cadx.ceus import CeusConfig, HalfRegistration, ViewRect, crop_frame
 from cadx.delivery import read_delivery
-from cadx.runs import make_build_id, read_git_state, write_snapshot
+from cadx.runs import finish_snapshot, make_build_id, read_git_state, write_snapshot
 from ceus_quant import REPO_ROOT, REVIEW_HEADER, file_sha256, load_config, resolve_paths
 
 KEY_HEADER = ("id", "video_id", "is_control", "dx", "dy", "video_sha256")
@@ -35,13 +35,6 @@ type _Gray = npt.NDArray[np.float32]
 type _Mask = npt.NDArray[np.bool_]
 
 
-def _indices(count: int, wanted: int) -> set[int]:
-    """与 `sample_frames_for_split` 相同的均匀抽帧下标。"""
-    if count <= wanted:
-        return set(range(count))
-    return {int(i) for i in np.linspace(0, count - 1, num=wanted, dtype=np.int32).tolist()}
-
-
 @dataclass(frozen=True)
 class Material:
     """一例出图要的东西：两个视图；整帧的时间平均灰度、抽样中间那帧的灰度、绿框紫圈像素；两半扇形。"""
@@ -59,8 +52,8 @@ def step1(video_path: Path, cfg: CeusConfig) -> tuple[Material, HalfRegistration
     """第 1 步：逐帧解码只留分屏和配准要的抽样帧，找分屏线，配准两半。"""
     video = ceus.VideoFile(video_path)
     _, frame_count, _, _ = video.probe()
-    split_at = _indices(frame_count, cfg.views.split_sample_frames)
-    register_at = _indices(frame_count, cfg.registration.sample_frames)
+    split_at = set(ceus.sample_indices(frame_count, cfg.views.split_sample_frames))
+    register_at = set(ceus.sample_indices(frame_count, cfg.registration.sample_frames))
     split_frames, frames = list[_U8](), list[_U8]()
     decoded = 0
     for index, frame in video.iter_frames():
@@ -169,9 +162,9 @@ def render(config_path: Path) -> None:
         csv.writer(handle).writerows([KEY_HEADER, *sorted(rows)])
     with (partial / "failed.csv").open("w", encoding="utf-8", newline="") as handle:
         csv.writer(handle).writerows([("video_id", "step", "reason"), *failures])
-    finished_at, lock = datetime.now().astimezone(), REPO_ROOT / "uv.lock"
     configs = {"ceus.yaml": config_path}  # align_check 段的随机数种子随 ceus.yaml 进快照
-    write_snapshot(partial / "snapshot", run_id, git, started_at, finished_at, lock, configs)
+    write_snapshot(partial / "snapshot", run_id, git, started_at, REPO_ROOT / "uv.lock", configs)
+    finish_snapshot(partial / "snapshot", datetime.now().astimezone(), {})
     if run_dir.exists():
         raise FileExistsError(f"运行目录已存在，不覆盖：{run_dir}")
     partial.rename(run_dir)

scripts/ceus_quant.py

@@ -7,6 +7,8 @@
 第 3 步（MedSAM2 分割）的结果按内容键缓存在 `<work_root>/segment_cache/<键>/masks.npz`，键相同就
 直接读缓存、不调显卡。某一例在第 1–6 步失败（`CeusCaseError`）记进 `failed.csv`，接着跑下一例；
 其余异常整次运行停下，留下 `.partial` 目录。终端只打印计数，不打印视频号。
+快照在开跑时写进 `.partial/snapshot/`（代码、配置和核对表原文、权重与 MedSAM2 版本），配置和核对表
+随后从快照里的副本读；收尾只补 `finished.json`（结束时间、显存峰值、分割缓存命中数）。
 """
 
 import argparse
@@ -44,7 +46,7 @@ from cadx.ceus import (
     tic_plot_markers,
 )
 from cadx.delivery import CaseRecord, read_delivery
-from cadx.runs import make_build_id, read_git_state, write_snapshot
+from cadx.runs import finish_snapshot, make_build_id, read_git_state, write_snapshot
 from cadx.segment import SegmentConfig, cuda_peak_memory_bytes, medsam2_commit, segment_video
 
 REPO_ROOT = Path(__file__).resolve().parents[1]
@@ -241,7 +243,7 @@ def _masks(
     if final.exists():
         with np.load(final / "masks.npz", allow_pickle=False) as data:
             return data["masks"], True
-    masks = segment_video(prepared.target_frames, prepared.prompt_masks, segment_cfg)
+    masks = segment_video(prepared.target_frames(), prepared.prompt_masks, segment_cfg)
     partial = cache_root / f"{key}.partial"
     # 上次中途停下留下的 .partial 会让这里报错：work/ 可以删掉重算，删掉它再跑。
     partial.mkdir(parents=True, exist_ok=False)
@@ -283,6 +285,7 @@ def _run_case(
     registration = prepared.registration
     summary = {
         "video_id": case.video_id,
+        "video_sha256": video_sha256,
         "segment_cache_key": key,
         "registration": {
             "contrast_side": prepared.contrast.side,
@@ -323,10 +326,31 @@ def main(argv: Sequence[str]) -> None:
             raise FileExistsError(f"运行目录已存在，不覆盖：{path}")
     partial.mkdir(parents=True, exist_ok=False)
 
-    review = read_review(cfg.paths.align_review)
-    delivery = read_delivery(cfg.paths.delivery_dir, None)
     checkpoint_sha256 = file_sha256(cfg.segment.checkpoint)
     commit = medsam2_commit(cfg.segment.medsam2_dir)
+    # 开跑时就写快照：运行期间改了 ceus.yaml 或 review.csv，快照里仍是这次用的那份。
+    snapshot_dir = partial / "snapshot"
+    write_snapshot(
+        snapshot_dir=snapshot_dir,
+        build_id=run_id,
+        git=git,
+        started_at=started_at,
+        uv_lock=REPO_ROOT / "uv.lock",
+        configs={"ceus.yaml": config_path, "align_review.csv": cfg.paths.align_review},
+    )
+    segment_record = {
+        "checkpoint_sha256": checkpoint_sha256,
+        "medsam2_commit": commit,
+        "device": cfg.segment.device,
+    }
+    (snapshot_dir / "segment.json").write_text(
+        json.dumps(segment_record, ensure_ascii=False, indent=2) + "\n", encoding="utf-8"
+    )
+    # 配置和核对表读快照里的副本，与快照记的是同一份字节。
+    written = load_config(snapshot_dir / "ceus.yaml")
+    cfg = resolve_paths(written)
+    review = read_review(snapshot_dir / "align_review.csv")
+    delivery = read_delivery(cfg.paths.delivery_dir, None)
 
     # 缓存键用配置文件里的写法（路径不随仓库位置变）。
     def key_of(video_sha256: str) -> str:
@@ -352,24 +376,10 @@ def main(argv: Sequence[str]) -> None:
         writer.writerow(_FAILED_HEADER)
         writer.writerows(failures)
     peak = cuda_peak_memory_bytes(cfg.segment.device)
-    write_snapshot(
-        snapshot_dir=partial / "snapshot",
-        build_id=run_id,
-        git=git,
-        started_at=started_at,
-        finished_at=datetime.now().astimezone(),
-        uv_lock=REPO_ROOT / "uv.lock",
-        configs={"ceus.yaml": config_path, "align_review.csv": cfg.paths.align_review},
-    )
-    segment_record = {
-        "checkpoint_sha256": checkpoint_sha256,
-        "medsam2_commit": commit,
-        "device": cfg.segment.device,
-        "cuda_peak_memory_bytes": peak,
-        "segment_cache_hits": cache_hits,
-    }
-    (partial / "snapshot" / "segment.json").write_text(
-        json.dumps(segment_record, ensure_ascii=False, indent=2) + "\n", encoding="utf-8"
+    finish_snapshot(
+        snapshot_dir,
+        datetime.now().astimezone(),
+        {"cuda_peak_memory_bytes": peak, "segment_cache_hits": cache_hits},
     )
     # POSIX 的 rename 会悄悄覆盖空目录，所以改名前再查一次。
     if run_dir.exists():