@oh-story/dsh 0.1.8 → 0.1.9

package/lib/drama/skills/short-drama-edit/scripts/selftest.py

@@ -0,0 +1,456 @@
+#!/usr/bin/env python3
+"""Offline self-test for cut-list parsing and mechanical checks.
+
+Runs without ffmpeg: every assertion here is about the document, not the media.
+The media-dependent paths are exercised by the suite's tests instead.
+"""
+
+from __future__ import annotations
+
+import re
+import sys
+import tempfile
+from pathlib import Path
+
+sys.path.insert(0, str(Path(__file__).resolve().parent))
+
+from edit_tool import (  # noqa: E402
+    DEFAULT_REMOTION_CONCURRENCY,
+    REMOTION_SOURCE,
+    REMOTION_SOURCE_FILES,
+    EditError,
+    _ass_text,
+    _build_ass,
+    _shot_match_filter,
+    _subtitle_cues,
+    check_cuts,
+    parse_cut_list,
+)
+
+MINIMUM_PYTHON = (3, 9)
+if sys.version_info < MINIMUM_PYTHON:
+    raise SystemExit("selftest.py requires Python 3.9 or newer")
+
+MOTION = """# EP001 视频提示词
+
+## MOTION-EP001-010 · 手指停在发布键上
+
+## MOTION-EP001-011 · 认证弹窗展开
+"""
+
+SCREENPLAY = """# EP001
+
+## EP001-SC003 内 · 房间 · 夜
+
+角色甲：这条路我自己走。
+"""
+
+CUT_LIST = """# EP001 剪辑单
+
+- 成片目标时长：5.00 秒
+- 交付响度：-16 LUFS
+- 未采用镜头：MOTION-EP001-009（理由：文件缺失——尚未生产）
+
+## CUT-EP001-001 · 手指停在发布键上
+
+- 来源：MOTION-EP001-010 · media/010.mp4
+- 入点：0.60
+- 出点：3.60
+- 时长：3.00
+- 取舍：入点=手指已经在下压；出点=界面开始响应
+- 声音：保留原声
+- 字幕：无
+
+## CUT-EP001-002 · 那句话
+
+- 来源：MOTION-EP001-011 · media/011.mp4
+- 入点：0.00
+- 出点：2.00
+- 时长：2.00
+- 取舍：入点=文件起点；出点=句尾收音后
+- 声音：保留原声
+- 字幕：这条路我自己走
+"""
+
+
+def require(condition: bool, message: str) -> None:
+    if not condition:
+        raise AssertionError(message)
+
+
+def build(root: Path, cut_list: str) -> Path:
+    episode = root / "剧集" / "EP001"
+    (episode / "media").mkdir(parents=True)
+    (episode / "视频提示词.md").write_text(MOTION, encoding="utf-8")
+    (episode / "剧本.md").write_text(SCREENPLAY, encoding="utf-8")
+    (episode / "剪辑单.md").write_text(cut_list, encoding="utf-8")
+    for name in ("010.mp4", "011.mp4"):
+        (episode / "media" / name).write_bytes(b"")
+    return episode
+
+
+def check_subtitle_geometry() -> None:
+    """FontSize is already in PlayRes units; scaling it by the frame twice ruins it.
+
+    Letting ffmpeg convert an SRT hands libass a 384-high canvas and every size
+    is then multiplied by height/384 on the way to the frame. Computing the size
+    from the real height on top of that produced type a third of the frame wide,
+    sitting in the middle of the picture with both ends of the line cut off — and
+    it shipped, because the render after the style change was checked with
+    numbers instead of a frame.
+    """
+
+    width, height = 768, 1344
+    ass = _build_ass([(1.0, 2.0, "这条路我自己走")], width, height)
+    require(f"PlayResX: {width}" in ass, "PlayResX 必须等于画面宽")
+    require(f"PlayResY: {height}" in ass, "PlayResY 必须等于画面高")
+    style = next(line for line in ass.splitlines() if line.startswith("Style:"))
+    fields = style.split(":", 1)[1].split(",")
+    font_size = float(fields[2])
+    require(
+        height * 0.02 <= font_size <= height * 0.06,
+        f"字号 {font_size} 不在画面高度的 2%–6% 之间",
+    )
+    margin_v = float(fields[21])
+    require(margin_v < height * 0.2, f"底边距 {margin_v} 会把字幕推离安全区")
+    margin_h = float(fields[19])
+    require(margin_h > 0, "左右边距为 0 时长句会顶到画面边缘")
+    require("这条路我自己走" in ass, "台词原文必须原样进 ASS")
+
+
+def check_subtitle_timing() -> None:
+    """Cues follow the segments that were written, not the numbers in the document.
+
+    Each segment lands on a frame boundary, so it runs a few milliseconds past
+    its declared length. Accumulating the declared numbers drifts, and the
+    subtitle leaves before the actor stops speaking.
+    """
+
+    with tempfile.TemporaryDirectory() as scratch:
+        root = Path(scratch)
+        episode = build(root, CUT_LIST)
+        _, cuts, _ = parse_cut_list(episode / "剪辑单.md")
+        declared = [cut.end - cut.start for cut in cuts]
+        measured = [span + 0.04 for span in declared]
+
+        drifting = _subtitle_cues(cuts, declared)
+        真实 = _subtitle_cues(cuts, measured)
+        require(len(真实) == 1, f"应有一条字幕，实际 {len(真实)}")
+        require(
+            真实[0][0] > drifting[0][0],
+            "分段比声明更长时，后面的字幕必须相应后移",
+        )
+        require(
+            abs(真实[0][0] - (drifting[0][0] + 0.04)) < 1e-6,
+            "位移必须等于前面各段的实测差之和",
+        )
+
+
+def check_shot_match() -> None:
+    """A declared correction is applied; an undeclared one never is.
+
+    Generated shots drift a stop apart, so the join reads as a mistake. The
+    correction has to be visible in the document — a tool that measured clips
+    and adjusted them on its own would be changing pictures nobody could review.
+    """
+
+    with tempfile.TemporaryDirectory() as scratch:
+        root = Path(scratch)
+        episode = build(root, CUT_LIST)
+        _, cuts, _ = parse_cut_list(episode / "剪辑单.md")
+        require(_shot_match_filter(cuts[0]) == "", "没写画面的段不得被改动")
+
+        listed = CUT_LIST.replace(
+            "- 声音：保留原声\n- 字幕：无",
+            "- 声音：保留原声\n- 画面：亮度 +0.06；色温 -6\n- 字幕：无", 1)
+        (episode / "剪辑单.md").write_text(listed, encoding="utf-8")
+        _, cuts, _ = parse_cut_list(episode / "剪辑单.md")
+        chain = _shot_match_filter(cuts[0])
+        require("brightness=0.06" in chain, f"亮度没进滤镜链: {chain}")
+        require("colorbalance" in chain and "rm=-0.06" in chain, f"色温没进滤镜链: {chain}")
+
+        # ffmpeg 的 eq 只接受 0..3 的饱和度；负数在文档里通过、渲染时才炸。
+        negative = CUT_LIST.replace(
+            "- 声音：保留原声\n- 字幕：无",
+            "- 声音：保留原声\n- 画面：饱和 -1.5\n- 字幕：无", 1)
+        (episode / "剪辑单.md").write_text(negative, encoding="utf-8")
+        try:
+            parse_cut_list(episode / "剪辑单.md")
+        except EditError as error:
+            require("超出允许范围" in str(error), f"负饱和度报错不对: {error}")
+        else:
+            raise AssertionError("饱和 -1.5 应当被拒绝——ffmpeg 不接受负饱和度")
+
+        # 超出范围是重新调色，不是接镜，必须挡住。
+        wild = CUT_LIST.replace(
+            "- 声音：保留原声\n- 字幕：无",
+            "- 声音：保留原声\n- 画面：亮度 +0.9\n- 字幕：无", 1)
+        (episode / "剪辑单.md").write_text(wild, encoding="utf-8")
+        try:
+            parse_cut_list(episode / "剪辑单.md")
+        except EditError as error:
+            require("超出允许范围" in str(error), f"越界报错不对: {error}")
+        else:
+            raise AssertionError("亮度 +0.9 应当被拒绝")
+
+        # 写了内容却没有可执行项，是写错了，不能静默当成不校正。
+        vague = CUT_LIST.replace(
+            "- 声音：保留原声\n- 字幕：无",
+            "- 声音：保留原声\n- 画面：调暖一点\n- 字幕：无", 1)
+        (episode / "剪辑单.md").write_text(vague, encoding="utf-8")
+        try:
+            parse_cut_list(episode / "剪辑单.md")
+        except EditError as error:
+            require("没有可执行的项" in str(error), f"含混报错不对: {error}")
+        else:
+            raise AssertionError("「调暖一点」应当被拒绝")
+
+
+def check_multi_subtitle() -> None:
+    """A shot can carry several lines, and all of them must reach the screen.
+
+    The exchange "就是什么 / 就是少了点东西 / 少了什么" is one over-shoulder shot.
+    A cut list that holds only one subtitle per cut silently drops the other
+    two: the film plays lines that never appear, and nothing reports it.
+    """
+
+    with tempfile.TemporaryDirectory() as scratch:
+        root = Path(scratch)
+        episode = build(root, CUT_LIST)
+        (episode / "剧本.md").write_text(
+            SCREENPLAY + "\n江晨：就是什么。\n江晨：就是少了点东西。\n江晨：少了什么。\n",
+            encoding="utf-8",
+        )
+        multi = CUT_LIST.replace(
+            "- 字幕：这条路我自己走",
+            "- 字幕 1：0.10-0.60 就是什么\n"
+            "- 字幕 2：0.80-1.40 就是少了点东西\n"
+            "- 字幕 3：1.50-1.90 少了什么",
+        )
+        (episode / "剪辑单.md").write_text(multi, encoding="utf-8")
+        _, cuts, _ = parse_cut_list(episode / "剪辑单.md")
+        require(len(cuts[1].subtitles) == 3, f"三句应全部解析，实际 {len(cuts[1].subtitles)}")
+        require(cuts[1].subtitles[2][2] == "少了什么", "字幕顺序错了")
+        require(not check_cuts(episode, cuts, root, probe=False), "三句合法字幕不该报错")
+        cues = _subtitle_cues(cuts, [cut.end - cut.start for cut in cuts])
+        require(len(cues) == 3, f"三句应各自成为一条 cue，实际 {len(cues)}")
+
+        # 编号必须连续：跳号意味着有一句被漏掉了。
+        gap = CUT_LIST.replace(
+            "- 字幕：这条路我自己走",
+            "- 字幕 1：0.10-0.60 就是什么\n- 字幕 3：1.50-1.90 少了什么",
+        )
+        (episode / "剪辑单.md").write_text(gap, encoding="utf-8")
+        try:
+            parse_cut_list(episode / "剪辑单.md")
+        except EditError as error:
+            require("连续" in str(error), f"跳号报错不对: {error}")
+        else:
+            raise AssertionError("字幕编号跳号应当被拒绝")
+
+        # 两句不能同时在屏上。
+        overlap = CUT_LIST.replace(
+            "- 字幕：这条路我自己走",
+            "- 字幕 1：0.10-1.00 就是什么\n- 字幕 2：0.80-1.40 就是少了点东西",
+        )
+        (episode / "剪辑单.md").write_text(overlap, encoding="utf-8")
+        try:
+            parse_cut_list(episode / "剪辑单.md")
+        except EditError as error:
+            require("重叠" in str(error), f"重叠报错不对: {error}")
+        else:
+            raise AssertionError("字幕时间重叠应当被拒绝")
+
+        # 多句时每句必须自带时间，否则无从摆放。
+        untimed = CUT_LIST.replace(
+            "- 字幕：这条路我自己走",
+            "- 字幕 1：就是什么\n- 字幕 2：0.80-1.40 就是少了点东西",
+        )
+        (episode / "剪辑单.md").write_text(untimed, encoding="utf-8")
+        try:
+            parse_cut_list(episode / "剪辑单.md")
+        except EditError as error:
+            require("各自带时间" in str(error), f"缺时间报错不对: {error}")
+        else:
+            raise AssertionError("多句字幕缺时间应当被拒绝")
+
+
+def check_stale_window() -> None:
+    """Regenerating a shot invalidates every subtitle window bound to it.
+
+    The windows were reverse-engineered from the old take's audio. Nothing else
+    notices the change: the numbers stay in range, the render still succeeds,
+    and the subtitle appears at a moment when nobody is speaking.
+    """
+
+    import os
+    import time
+
+    with tempfile.TemporaryDirectory() as scratch:
+        root = Path(scratch)
+        # Anchored on the field, not on the line's wording: the fixture's dialogue
+        # is illustrative and gets rewritten, and a test that breaks when an
+        # example changes is a test nobody trusts.
+        timed = re.sub(
+            r"^(- 字幕：(?!无).+)$",
+            r"\1\n- 字幕时间：0.30-1.60",
+            CUT_LIST,
+            count=1,
+            flags=re.MULTILINE,
+        )
+        require(timed != CUT_LIST, "夹具里应当有一条带台词的字幕")
+        episode = build(root, timed)
+        _, cuts, _ = parse_cut_list(episode / "剪辑单.md")
+        require(
+            not check_cuts(episode, cuts, root, probe=False),
+            "素材不比剪辑单新时不该报陈旧",
+        )
+
+        media = episode / "media" / "011.mp4"
+        listing = episode / "剪辑单.md"
+        newer = listing.stat().st_mtime + 60
+        os.utime(media, (newer, newer))
+        findings = check_cuts(episode, cuts, root, probe=False)
+        require(
+            any("素材比剪辑单新" in item for item in findings),
+            f"重出后的陈旧字幕时间没抓到: {findings}",
+        )
+
+        # 改过剪辑单之后就不再报——那正是「重测并改」的动作。
+        time.sleep(0.01)
+        listing.write_text(timed, encoding="utf-8")
+        later = newer + 60
+        os.utime(listing, (later, later))
+        _, cuts, _ = parse_cut_list(listing)
+        require(
+            not any("素材比剪辑单新" in item for item in check_cuts(episode, cuts, root, probe=False)),
+            "改过剪辑单之后不该继续报陈旧",
+        )
+
+
+def check_ass_escaping() -> None:
+    """A quoted line keeps every character it had in 剧本.md.
+
+    `{`…`}` is an override block in ASS: unescaped, libass drops the braces and
+    everything between them, on screen, with no error anywhere. EDT-05 exists to
+    keep the burned text equal to the screenplay's character for character.
+    """
+
+    line = "他念出来：{姓名}，请签字。"
+    escaped = _ass_text(line)
+    require("\\{" in escaped and "\\}" in escaped, f"花括号没有转义: {escaped}")
+    require("姓名" in escaped, "转义把字弄丢了")
+    ass = _build_ass([(1.0, 2.0, line)], 1080, 1920)
+    dialogue = [row for row in ass.splitlines() if row.startswith("Dialogue")][0]
+    require(dialogue.endswith(escaped), f"Dialogue 行没有用转义后的正文: {dialogue}")
+    require(
+        _ass_text("反斜杠 \\N 不是换行").count("\\\\") == 1,
+        "反斜杠没有转义，\\N 会被当成换行",
+    )
+
+
+def check_remotion_sources_all_shipped() -> None:
+    """The sync list is a whitelist; a file left out of it is missing at render.
+
+    Nothing else notices until someone has already paid for that render.
+    """
+
+    # `as_posix`, because the whitelist is written with forward slashes and
+    # `relative_to` yields backslashes on Windows.
+    on_disk = {
+        path.relative_to(REMOTION_SOURCE).as_posix()
+        for path in REMOTION_SOURCE.rglob("*")
+        if path.is_file()
+        and path.suffix in {".ts", ".tsx", ".json"}
+        and "node_modules" not in path.parts
+    }
+    listed = set(REMOTION_SOURCE_FILES)
+    require(not (on_disk - listed), f"Remotion 源文件没进同步清单: {sorted(on_disk - listed)}")
+    require(not (listed - on_disk), f"同步清单里有不存在的文件: {sorted(listed - on_disk)}")
+
+
+def check_remotion_concurrency_is_capped() -> None:
+    """Remotion defaults to one browser per core, each holding a full frame."""
+
+    require(DEFAULT_REMOTION_CONCURRENCY >= 1, "并发上限必须是正数")
+    source = Path(__file__).resolve().with_name("edit_tool.py").read_text(encoding="utf-8")
+    require("--concurrency=" in source, "render 必须显式传 --concurrency，不能用 Remotion 的默认值")
+
+
+def main() -> int:
+    with tempfile.TemporaryDirectory() as scratch:
+        root = Path(scratch)
+        episode = build(root, CUT_LIST)
+        delivery, cuts, unused = parse_cut_list(episode / "剪辑单.md")
+        require(delivery.target_seconds == 5.0, "目标时长没有解析出来")
+        require(delivery.loudness_lufs == -16.0, "交付响度没有解析出来")
+        require(len(cuts) == 2, f"应解析出 2 段，实际 {len(cuts)}")
+        require(len(unused) == 1, "未采用镜头没有解析出来")
+        require(cuts[1].subtitles[0][2] == "这条路我自己走", "字幕文字没有解析出来")
+
+        clean = check_cuts(episode, cuts, root, probe=False)
+        require(not clean, f"完好的剪辑单不该有 findings: {clean}")
+
+        # 出点 - 入点 与声明时长不符必须抓到。
+        drifted = CUT_LIST.replace("- 时长：3.00", "- 时长：2.50")
+        (episode / "剪辑单.md").write_text(drifted, encoding="utf-8")
+        _, cuts, _ = parse_cut_list(episode / "剪辑单.md")
+        findings = check_cuts(episode, cuts, root, probe=False)
+        require(any("与「时长" in item for item in findings), f"时长漂移没抓到: {findings}")
+
+        # 字幕必须能在剧本里找到原文；转写猜出来的字不行。
+        invented = CUT_LIST.replace("字幕：这条路我自己走", "字幕：这条陆我自己走")
+        (episode / "剪辑单.md").write_text(invented, encoding="utf-8")
+        _, cuts, _ = parse_cut_list(episode / "剪辑单.md")
+        findings = check_cuts(episode, cuts, root, probe=False)
+        require(any("找不到原文" in item for item in findings), f"编造的字幕没抓到: {findings}")
+
+        # 未知 MOTION 必须抓到。
+        unknown = CUT_LIST.replace("MOTION-EP001-011", "MOTION-EP001-099")
+        (episode / "剪辑单.md").write_text(unknown, encoding="utf-8")
+        _, cuts, _ = parse_cut_list(episode / "剪辑单.md")
+        findings = check_cuts(episode, cuts, root, probe=False)
+        require(any("不在《视频提示词.md》中" in item for item in findings), f"未知 MOTION 没抓到: {findings}")
+
+        # 素材缺失必须抓到。
+        (episode / "media" / "011.mp4").unlink()
+        (episode / "剪辑单.md").write_text(CUT_LIST, encoding="utf-8")
+        _, cuts, _ = parse_cut_list(episode / "剪辑单.md")
+        findings = check_cuts(episode, cuts, root, probe=False)
+        require(any("素材不存在" in item for item in findings), f"缺素材没抓到: {findings}")
+
+        # 同一素材上的两段不得重叠。
+        overlapping = CUT_LIST.replace(
+            "- 来源：MOTION-EP001-011 · media/011.mp4", "- 来源：MOTION-EP001-011 · media/010.mp4"
+        ).replace("- 入点：0.00", "- 入点：1.00").replace("- 出点：2.00", "- 出点：3.00")
+        (episode / "media" / "011.mp4").write_bytes(b"")
+        (episode / "剪辑单.md").write_text(overlapping, encoding="utf-8")
+        _, cuts, _ = parse_cut_list(episode / "剪辑单.md")
+        findings = check_cuts(episode, cuts, root, probe=False)
+        require(any("区间重叠" in item for item in findings), f"同源重叠没抓到: {findings}")
+
+        # 缺字段是文档缺陷，报出来而不是崩掉。
+        (episode / "剪辑单.md").write_text(
+            CUT_LIST.replace("- 出点：3.60\n", ""), encoding="utf-8"
+        )
+        try:
+            parse_cut_list(episode / "剪辑单.md")
+        except EditError as error:
+            require("出点" in str(error), f"缺字段的报错没点名字段: {error}")
+        else:
+            raise AssertionError("缺「出点」应当报错")
+
+    check_subtitle_geometry()
+    check_subtitle_timing()
+    check_shot_match()
+    check_ass_escaping()
+    check_multi_subtitle()
+    check_stale_window()
+    check_remotion_sources_all_shipped()
+    check_remotion_concurrency_is_capped()
+    print("short-drama-edit self-tests passed")
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

package/lib/drama/skills/short-drama-image-prompts/assets/image-prompts.md

@@ -2,6 +2,11 @@
 
 此文件由已接受规格生成。元信息用于核对来源；只复制引用块内的通用提示词。不要手写供应商参数。
 
+**这份模板属于结构化管线**（`image-prompt-specs.jsonl` 一路）。creator-first 项目的《图片提示词.md》骨架以 `SKILL.md` 为准：`## IMG-... · 中文名`，字段写成不加粗的
+`- 参考：`（取「无外部参考」或完整 `REF-...` 语法），正文标题是 `### 可复制提示词`。
+不要把这里的加粗字段名、`参考图用途`、`本次呈现` 或 `### 可复制通用提示词` 搬进 creator-first 文档——
+`creator_markdown_check.py` 按 `SKILL.md` 的骨架校验，照本模板写会被判「缺少参考字段」。
+
 ```markdown
 # EP<编号> · 资产图片提示词
 

package/lib/drama/skills/short-drama-image-prompts/assets/lookdev-prompts.md

@@ -4,6 +4,9 @@
 只复制引用块内的自然语言提示词；本文件不触发图片或媒体服务，实际生产交
 `$short-drama-produce` 并单独确认。
 
+**这是项目级文件，不是某一集的创作文档。** 它不参与 `creator_markdown_check.py` 的
+五文档校验；某一集要用的资产提示词写进该集的《图片提示词.md》，骨架以 `SKILL.md` 为准。
+
 ```markdown
 # 项目 Look Development 提示词
 

package/lib/drama/skills/short-drama-novel-analyze/references/aggregation-and-entities.md

@@ -69,13 +69,13 @@
 | 指标 | 参考值 | 算法 | 不达标怎么办 |
 |---|---|---|---|
 | 归属置信 | ≥ 0.85 | 有明确归属的情节点 / 单元内情节点总数 | 低于线的单元标「待复核」，不删 |
+| 覆盖率 | 85%–95% | 已归类情节点 / 总情节点数 | < 85% 走散落兜底；> 95% 复核是否过度归并 |
+| 重叠率 | ≤ 35% | 跨单元共享的情节点 / 总情节点数 | > 35% 说明两个单元其实是一个，考虑合并 |
 
 「有明确归属」指这条情节点服务于本单元的戏剧目标或对抗力量，而不只是恰好落在本单元的
 章节范围内。按连续章节切分时，若把「落在范围内」当作归属，这个比值恒等于 1，指标失去
 意义——所以逐条判的是**它推动的是不是这个单元的那件事**，跨单元共享的情节点计入重叠率
 而不计入本单元的分子。
-| 覆盖率 | 85%–95% | 已归类情节点 / 总情节点数 | < 85% 走散落兜底；> 95% 复核是否过度归并 |
-| 重叠率 | ≤ 35% | 跨单元共享的情节点 / 总情节点数 | > 35% 说明两个单元其实是一个，考虑合并 |
 
 **小体量素材的例外**：不足 30 章、单主线连续叙事的材料按时序切分时，覆盖率天然接近
 100%、散落接近 0，这是线性结构的正常结果，不是过度归并。此时复核的重点是

package/lib/drama/skills/short-drama-novel-analyze/references/chapter-extraction.md

@@ -103,9 +103,12 @@ P=$(grep -cE '^P[0-9]+ ' "$F")
 grep -cE '^P[0-9]+ [^|]+\|[^|]*[^|[:space:]][^|]*\|[^|]*涉及' "$F"   # 必须 == $P
 grep -cE '^P[0-9]+ .*\|.*\|.*\|[^|]*功能：' "$F"                     # 必须 == $P
 grep -hoE '^P[0-9]+ [^|]+' "$F" | awk '{print $2}' | sort -u        # 必须 ⊆ 类型表
-grep -cE '^\|[^|]*\|[^|]*\|' "$F"                                  # 必须 >= 2
+awk '/^## 载体记录/{f=1;next} /^## /{f=0} f&&/^\|/{n++} END{print n+0}' "$F"  # 必须 >= 3
 ```
 
+- 载体表计数少于 3：该章只写了表头没写内容。阈值是 3 不是 2，因为表头行与 `|---|` 分隔行
+  本身就占 2 行——按 2 计的话，一张空表照样过闸。**这一条也必须只数「载体记录」小节内的
+  表格行**：数全文的 `|` 行会把「出场人物」表算进来，同样让空载体表蒙混过关。
 - 白描段计数少于 `$P`：有情节点缺白描，或字段顺序、分隔符不对。
 - 功能段计数少于 `$P`：有情节点没回答改变了什么。注意全角冒号——半角会漏。
   这一条只数情节点行第四段里的 `功能：`：早先的写法数的是全文任何位置的这四个字符，

package/lib/drama/skills/short-drama-produce/SKILL.md

@@ -75,11 +75,13 @@ python3 <本技能目录>/scripts/production_tool.py prepare <project> --job <jo
 python3 <本技能目录>/scripts/production_tool.py confirm <project> --job-id <id> --confirmation "CONFIRM <id> <code>"
 python3 <本技能目录>/scripts/production_tool.py run <project> --job-id <id> --adapter-config <outside-project-config.json>
 python3 <本技能目录>/scripts/production_tool.py status <project> --job-id <id>
+python3 <本技能目录>/scripts/production_tool.py collect <project> --job-id <id> --adapter-config <outside-project-config.json>
 python3 <本技能目录>/scripts/production_tool.py audit <project>
 ```
 
 `prepare` 只验证并预览，不生产。`confirm` 只保存与当前 job 指纹绑定的一次性确认。
-`run` 才启动 adapter。`audit` 只对账本地任务历史、失败后恢复、重复内容尝试和当前输出字节，
+`run` 才启动 adapter。`collect` 只取回**已经提交过、已经计费**的那次任务的结果，
+不重新提交、也不需要新的确认——见下面「中断不等于要重跑」。`audit` 只对账本地任务历史、失败后恢复、重复内容尝试和当前输出字节，
 不会调用供应商，也不把技术成功、文件存在或哈希一致写成媒体质量结论。同一 job 存在未决
 `running` attempt 时禁止重新 prepare、confirm 或 run；先等待完成或排查遗留 attempt。
 
@@ -97,8 +99,8 @@ python3 <本技能目录>/scripts/production_tool.py audit <project>
 - **tts**：从 `剧本.md` 读取原句与表演要求，声音参考由用户或现有媒体明确提供。不得在生产 job
   中改词，也不为 TTS 新建第六份创作文档。
 - **music**：读取 `视频提示词.md` 中创作者已确认的时间线音乐章节；主题曲使用已确认歌词，纯配乐
-  不携带歌词。供应商不能精确承诺时长时，生成源音轨后仍由剪辑按文档里的混音意图完成落点、循环、
-  淡入淡出和对白 ducking。
+  不携带歌词。供应商不能精确承诺时长时，生成源音轨后仍由 `$short-drama-edit` 按文档里的混音意图完成落点、
+  循环、淡入淡出和对白 ducking。
 
 一个 job 不混合 modality。大批量工作拆成创作者能看清数量和成本边界的小 job；不为方便把整季
 隐式塞进一次确认。
@@ -133,6 +135,23 @@ adapter 配置必须在项目外，只包含 argv 命令和超时；凭据由 ad
 
 仓库自带 `fixture_adapter.py` 只用于离线测试，不代表真实生成质量或默认生产 adapter。
 
+## 中断不等于要重跑
+
+视频任务在**提交那一刻**就已经计费，不是在拿到结果时。提交之后的一切——轮询几分钟、下载——
+都可能被进程被杀、断网、机器休眠打断。内置 adapter 因此在拿到供应商任务 ID 的第一时间就把它
+写进 `handle_path`（早于第一次轮询），这个路径不随本次尝试一起删除。
+
+于是中断之后有三条确定的动作，不必再花一次钱：
+
+1. `audit` 会把带着任务 ID 的未完成尝试报成 `orphaned_provider_job`，
+   `action` 是 `collect_before_retry`；
+2. `collect` 用那个 ID 取回结果并把这次尝试标成成功；
+3. 只有在 `collect` 也确认那边确实失败之后，才走重新确认与重投的老路。
+
+**不要在 `audit` 报出 `orphaned_provider_job` 时直接重投**——那是在为同一个镜头付第二次钱。
+`collect` 不走确认闸门是有意的：闸门防的是意外花钱，而 collect 不花钱；
+如果它也要求重新确认，那么中断之后最省事的路径就变成再付一次，正好是闸门要防的事。
+
 ## 结果与复核
 
 成功后回报实际输出路径、媒体类型和运行状态；不要把“adapter 返回成功”写成质量结论。
@@ -145,6 +164,8 @@ adapter 配置必须在项目外，只包含 argv 命令和超时；凭据由 ad
 构图与角色一致、画面本身合规的图；音频被拒重录那句台词。改动写进新的 job 重新 prepare，让创
 作者在预览里看到改的是哪一项再确认；原样重投的那次确认不产生修复，只产生一笔费用。重复内容
 缺陷回到对应 prompt/spec owner。
+生产结束后不自动进入剪辑。素材要装配成成片时由用户点名 `$short-drama-edit`；它只取舍已有帧，
+不回头改本阶段的 job，也不生成新素材。
 如需质量复核，报告可把已有结果另行交给 `$short-drama-review`；不要在生产调用中自动启动复核。
 Dashboard 只负责展示这些文件和运行摘要，不提供 adapter 设置或生产按钮。
 

package/lib/drama/skills/short-drama-produce/references/adapter-contract.md

@@ -113,6 +113,34 @@ on the machine locale. The document contains the confirmed job plus:
   production tool. Every adapter output source must be a direct regular-file
   child of this directory. The whole directory is deleted after success or
   failure.
+- `handle_path`: absolute path where the adapter records the provider's task id
+  **as soon as it has one, before it starts polling** — write
+  `{"provider_job_id": "<id>"}`, atomically. Unlike `output_root` this path
+  survives the attempt.
+- `collect_provider_job_id`: present only on a collect call. When it is set the
+  adapter must **not submit anything**; it polls and downloads that existing
+  task and returns its outputs as usual.
+
+## Why the handle exists
+
+A video task is billed the moment it is submitted, not when its result is
+collected. Everything after submission — polling for minutes, downloading — can
+be interrupted by a killed process, a dropped connection, or a sleeping laptop.
+Without a durable id the attempt record says `failed` while the provider's task
+is alive and already paid for, and the only way forward is to submit again and
+pay a second time for the same shot.
+
+So: write the handle before the first poll, and treat failing to write it as
+non-fatal — a submitted task must never be failed because its id could not be
+recorded locally. The tool copies the handle onto the attempt record on both the
+success and the failure path, `audit` reports any unfinished attempt that
+carries one as `orphaned_provider_job` with `action: collect_before_retry`, and
+`production_tool.py collect` fetches it.
+
+`collect` deliberately sits outside the confirmation gate. That gate exists to
+prevent an unintended charge; collecting spends nothing because the charge
+already happened. Requiring a fresh confirmation would make paying again the
+cheapest way out of an interruption — the exact outcome the gate is for.
 
 It may translate provider-neutral parameters into its chosen SDK/API. Optional
 provider adapters under `scripts/` document and implement known translations;

package/lib/drama/skills/short-drama-produce/references/providers/minimax-h3-video.md

@@ -31,6 +31,12 @@ required and are only accepted when the runtime profile above permits their valu
 optional for reference-conditioned jobs and **required** for text-to-video, where `adaptive` is
 refused because there is no reference frame to adapt to.
 
+The adapter validates and forwards `duration`; it does not estimate speech length or retime dialogue.
+Before preparing a speaking shot, compare the accepted shot duration, the prompt's spoken timeline,
+and the job's `duration`. Resolve insufficient speaking time in the storyboard and regenerate the
+dependent prompt before preparing the job. Prompt timestamps alone do not change the request duration.
+The official video API has no separate speech-rate parameter.
+
 The prompt is compiled into one `text` item of the multimodal `content` array and is refused above
 7000 characters. Each declared reference becomes one further `content` item carrying an explicit
 `role`: `first_frame`, `last_frame`, `reference_image`, `reference_video`, or `reference_audio`.
@@ -73,11 +79,24 @@ what changes upstream is only which axes of the target-model profile a project d
 video-prompt skill's target-model profile. Two consequences are worth stating here because they
 show up as production defects rather than as API errors:
 
-- an unstated audio channel is not a silent one. A shot that says nothing about music or speech can
-  come back with invented score and invented lines, so the shot's own sound intent, including what
-  must **not** be produced, has to be in the copyable body;
-- readable on-screen text is likewise produced unless it is excluded, so a shot with no readable-text
-  obligation states that exclusion instead of relying on the model to omit it.
+- an unstated **music** channel is not a silent one. Probing this suite's own prompts against H3, a
+  body carrying `non_diegetic_music: N/A` came back without a score in 14 of 14 runs, while a body
+  that said nothing about music came back with an invented score in 2 of 3. State the music intent.
+- the **speech** channel did not behave the same way in those probes: no run invented dialogue,
+  including 8 runs whose body carried a contradictory `No dialogue` exclusion. Treat "an unstated
+  vocal channel gets filled" as unverified rather than as a known failure, and write the shot's
+  sound intent because the body should say what this shot sounds like — not on the strength of a
+  mechanism this suite has not observed.
+- readable on-screen text was **not** produced unmasked in those probes either: across 44 sampled
+  frames from 11 clips — including dialogue clips carrying no exclusion sentence, on H3 and on
+  Seedance 2.0/2.5 — no burned-in subtitle or caption appeared. The exclusion sentence this suite
+  mandates is therefore precautionary **for video** rather than evidence-driven. Keep writing it:
+  a one-line exclusion costs nothing, these probes are small enough that they only show the failure
+  was not produced here, and generator behaviour changes between versions. What should change is the
+  justification — do not defend it as a fix for an observed defect. Note the asymmetry before carrying this reasoning anywhere else:
+  the same probe run against `gpt-image-2` found the opposite — an unconstrained still came back
+  covered in invented readable text, and the exclusion sentence cleaned it up. The image stages
+  keep their text policy on evidence; only the video stage's copy of it is precautionary.
 
 Neither is written here as a fixed phrase, and this suite never injects one. The wording belongs to
 the shot and to the project's declared prompt language.