@lemoncat7/dsh-partner 2.5.0 → 2.7.0

package/resources/skills/drama-production/upstream/skills/short-drama-video-prompts/references/stage-contract.md

@@ -0,0 +1,60 @@
+# 视频提示词阶段契约
+
+本阶段只拥有 `剧集/<EP>/视频提示词.md` 中的 `MOTION-...` 项与按需的时间线音乐章节。它继承镜头
+起点、终点、时长、对白和冻结帧，不回写分镜或视觉设定：镜头边界、职责、起终状态和关键帧正文都不在
+本阶段手里。
+
+唯一的例外是参考准备。SKILL 工作流第 2、3 步要求在同一请求内刷新受影响镜头的「输入参考图」
+（必要时一并回填「视觉依据」），那是**分镜 owner 在本请求内的动作**，按分镜阶段的字段规则执行，
+不是本阶段改写分镜的授权。它只更新这两条依据字段，不动镜头边界、职责、起终状态或关键帧正文。
+
+每项正文只描述从当前冻结起点到终点的变化；多镜打包、补拍或替代关系确有需要时，也写在同一文档
+的可读章节中，不建立 motion spec、container、readiness、QA 或接受记录。
+
+## 本阶段规则
+
+### `VID`
+
+| ID | Class | Knowledge |
+|---|---|---|
+| VID-01 | structural_invariant | Motion reads but cannot rewrite shot start/end/duration/dialogue and next-shot state. The duration a MOTION declares is compared against its SHOT: they must be the same number, because 时长 is what reaches the execution end and what VID-04/VID-13 arithmetic is built on. A downstream copy left behind by an upstream revision is a defect even though every string still parses. |
+| VID-02 | craft_default | Write start anchor, ordered subject motion, camera behavior, timing, and end report; add performance change and environment/audio only when this shot actually carries them. |
+| VID-03 | structural_invariant | Choose image-to-video only when the matching storyboard's reference set is complete, declares readable REF inputs, and is copied unchanged. No REF is not by itself consent to text-to-video: first discover and bind matching real project images; preserve any verified partial bindings, but if required images remain missing, report them and stop before the final motion document. Choose text-to-video only after the creator explicitly selects it, and carry a non-empty static visual anchor in the copyable text. Only image-to-video may omit appearance/composition already carried by its real reference frame. |
+| VID-04 | structural_invariant | Explicit segment timing sums exactly to its shot's accepted duration—neither exceeding it nor leaving an unallocated remainder. |
+| VID-05 | reviewed_invariant | Untimed action load must be feasible enough to preserve the intended performance and story change. |
+| VID-06 | structural_invariant | Locked and moving camera instructions cannot govern the same interval without an explicit transition. |
+| VID-07 | taste_option | Camera may be locked or moving; audio/lip-sync detail follows the chosen production profile. |
+| VID-08 | reviewed_invariant | Structured motion names this shot's exact subjects, actions, contacts, and results rather than reusable placeholders; when a performance path is present, it names only the actors and visible changes this shot actually carries. |
+| VID-09 | structural_invariant | The next start names an existing `SHOT-...` or is explicitly unresolved; never invent a hidden source key, record or hash. |
+| VID-10 | craft_default | Resolve one accepted production profile for the current delivery scope; local variants may coexist when their range and precedence are explicit, without overriding source coverage or exact-readable obligations. |
+| VID-11 | reviewed_invariant | A selective transform names its trigger, exact target scope, end geometry/state, and preserve set so non-target people, props, text surfaces, and spatial anchors do not change with it. |
+| VID-12 | reviewed_invariant | A pickup or alternate names stable `MOTION-...` IDs and maps each source requirement to this version, the master, another pickup, or a requested storyboard revision; it never silently replaces the master. |
+| VID-13 | structural_invariant | A multi-shot delivery group contains contiguous shots from one scene with one geography/asset chain. Its duration equals the sum of member shots, and grouping changes neither shot boundaries nor per-shot reviewability. |
+| VID-14 | craft_default | Music intent may be annotated per shot as a relative entry/exit/duck against neighbours, but its realization belongs to the timeline layer; no deliverable—single-shot or multi-shot container—carries a baked-in music bed unless the project accepted otherwise or the source is diegetic. Dialogue, off-screen sources, ambience, and event effects stay with the deliverable. |
+| VID-15 | structural_invariant | Within one episode a shot belongs to at most one delivery group; grouped and loose shots account for the visible shot set exactly once. |
+| VID-16 | reviewed_invariant | When performance changes, multi-character motion differentiates the actors who actually carry it and keeps each chosen signal readable in the accepted framing; it does not require an arc for non-performing shots, force every craft field, or duplicate one emotion across the cast. |
+| VID-17 | reviewed_invariant | Multiple references use stable visible IDs and an explicit order, so reordering cannot silently change which reference controls which property. |
+| VID-18 | reviewed_invariant | Readiness is reported from the current visible inputs and real blocking gaps, not persisted as a motion fact; text readiness never claims generated identity, performance, lip-sync, mix, edit or market quality. |
+| VID-19 | reviewed_invariant | When the creator profile declares required literal tokens for a delivery route, that route's delivery text preserves them byte-for-byte and outside the verbatim-dialogue fence. Paraphrase, translation, reordering, or omission is treated as a defect because a literal-matching surface has no reason to reject the rewritten text—the failure is silent rather than reported. The suite asserts no specific surface's behaviour, and whether a given result took the route is returned adherence, provable only by a bound production observation. Tokens declare a route and never substitute for the start state, action, or endpoint. Absent a declared token list the suite invents none; until the profile exposes a machine-readable list at a pinned field path, the reviewer cites the profile against the delivery text rather than claiming mechanical enforcement. |
+| VID-20 | reviewed_invariant | Packing routes change delivery granularity only, leaving shot boundaries, shot purpose, and per-shot reviewability intact. A single long-form generation carrying several accepted shots *is* a multi-shot container and is billed under VID-13 and VID-15; it introduces no separate accounting and no exemption from the contiguity, binding-chain, and scene-boundary constraints. A continuation route instead starts from a previously generated result, which is observation evidence and not an accepted artifact: the accepted shot start boundary stays the sole authority, and any claim about the observed state binds a production observation record or remains `unverified`. |
+| VID-21 | craft_default | When the project's accepted production profile has generated imagery carry the frame, actions are written as high-frequency, whole-body or single-limb movements common in everyday footage; precise interception, invisible internal states, negative actions, and three-or-more-step two-handed choreography are rewritten into equivalent common-action combinations, with the dramatic information carried by combination and timing. The rule is inactive for live action or an undeclared profile, and a rewrite may never change accepted shot boundaries, terminal states, or screenplay fact. |
+| VID-22 | reviewed_invariant | Copyable text describes the intended image and action. IDs, workflow status, hashes and file paths stay outside it. Prefer observable states where they clarify execution; necessary negative constraints are allowed. Unless explicitly authored, prohibit non-diegetic subtitles, captions and dialogue overlays while preserving approved in-scene text and its exact content, carrier and placement. |
+| VID-23 | reviewed_invariant | For models generating audio with video, specify the intended dialogue, VO, OS, vocalizations, effects, ambience and silence. Keep source, speaker, language and verbatim dialogue boundaries. Summaries must match the actual sound layers; follow model-specific music syntax without assuming missing layers will always be generated. |
+| VID-24 | structural_invariant | An image-to-video shot carries the storyboard's `REF-...` and `PLAN-...` slots unchanged, each with its 用途, so the motion document itself answers which start frame and which identity, geography, and prop pictures this shot sends — and, for `PLAN-...`, in which order the creator attaches them. The reference mode — first frame, first and last frame, or multi-slot reference — is read from that 用途 combination and translated by the matched model dialect; where a model makes frame and reference conditioning mutually exclusive, the start frame joins the reference group rather than splitting the request. |
+| VID-25 | structural_invariant | Quoted Chinese in a copyable body is delivered, not authored. Every run of four or more Han characters inside a `<d>` tag or quotation marks resolves, ignoring punctuation and spacing, to text that already exists in `剧本.md`, `视觉设定.md` or `分镜.md`. A line the shot wishes the character had said belongs upstream in the screenplay, not in the prompt. Runs of three characters or fewer are in-frame labels and interjections far more often than lines, so they stay with review rather than the mechanical check. |
+
+### `CON`
+
+| ID | Class | Knowledge |
+|---|---|---|
+| CON-01 | structural_invariant | Linked end and next start states match or have an explicit owner revision. |
+| CON-02 | reviewed_invariant | Knowledge, injury, ownership, weather, light, or physical state does not teleport/regress without story cause. |
+| CON-03 | craft_default | Track downstream-relevant deltas, not the whole 设定集 in every shot. |
+| CON-04 | structural_invariant | A continuity change states before, after, cause/source scene, effective range, and affected visible IDs. |
+| CON-05 | taste_option | Declared montage, ellipsis, dream, or subjective imagery may intentionally break ordinary continuity. |
+| CON-06 | structural_invariant | A continuity change names every existing downstream document it affects; future work is described, not pre-created. |
+| CON-07 | structural_invariant | A continuity lock declared in `视觉设定.md` fixes one verbatim surface plus its shot and image scope; every in-scope frozen keyframe, motion prompt, and named image prompt carries that surface unchanged. |
+
+规则分级由高到低：`structural_invariant`（结构缺陷，阻断）、
+`reviewed_invariant`（需证据判断）、`craft_default`（常用做法，可覆盖）、
+`taste_option`（创作者选择，不作缺陷）。创作者已接受的事实优先于本表。

package/resources/skills/drama-production/upstream/skills/short-drama-video-prompts/references/target-model-profile.md

@@ -0,0 +1,116 @@
+# 目标模型能力档案
+
+## 目录
+
+- [这份档案解决什么](#这份档案解决什么)
+- [九条会改变正文的能力轴](#九条会改变正文的能力轴)
+- [声音同轨生成](#声音同轨生成)
+- [参考条件方式](#参考条件方式)
+- [未声明时怎么写](#未声明时怎么写)
+- [自检](#自检)
+
+## 这份档案解决什么
+
+本套件的模型相关规则一律写成“按目标模型确认”，但“确认”在实践中经常没有落点：没人知道要确认
+哪几件事，于是提示词按最保守的写法交出去，执行端再靠自己补——补出来的往往就是漂移。
+
+本文只做一件事：把**需要确认什么**列成有限的几条。通用档案不猜供应商能力；已核对官方契约的
+模型可以由本技能的版本化方言文件提供推荐值。每一条取值都由创作者按自己接入的执行端声明一次，
+本集共用。生产端具体接哪个供应商、用什么字段，属于 `$short-drama-produce` 的 adapter 档案，不进入创作文档。
+
+**`craft_default`**： 能力轴是**声明**出来的，不是推断出来的。没有声明就按“未声明时怎么写”写；
+不要因为某个执行端“据说支持”某种写法就改写已接受的镜头边界、时长或起止状态。
+
+档案写在 `short-drama.json#/creator_authority/production_profile/choices`。最小结构如下；未知字段留空，
+不能从模型简称补全：
+
+```json
+{
+  "target_video_model": "versioned-model-profile",
+  "video_prompt_dialect": "dialect-id",
+  "video_prompt_language": "zh-CN",
+  "native_duration_seconds": {"min": 4, "max": 15},
+  "supported_generation_modes": ["text", "reference"],
+  "audio_generation": "same_pass"
+}
+```
+
+## 九条会改变正文的能力轴
+
+| 能力轴 | 声明什么 | 声明后正文允许怎么变 |
+|---|---|---|
+| 模型版本 | 项目使用的版本化模型档案键；账号侧 Endpoint / model ID 留在生产配置 | 只路由到该版本的方言；不因名称相近套用另一版本 |
+| 提示词方言 | 通用自然语言 / 目标原生结构及版本 | 命中已有方言时按其字段、素材标记和声音语法写；语言标签不能代替方言 |
+| 单次原生时长 | 一次生成能覆盖的最短与最长秒数 | 分镜规划时就让每个投产镜头落在区间内；短动作在合法镜长内到达终点后 hold，长动作按闭合状态拆镜 |
+| 参考条件方式 | 无 / 首帧 / 首尾帧 / 多槽参考（及槽位数上限） | 见下节 |
+| 声音是否同轨生成 | `same_pass` 或 `separate` | 见下节 |
+| 画幅与分辨率 | 执行端实际接受的画幅与分辨率集合 | 只写项目已接受的那一个；不在正文里罗列备选 |
+| 正文长度上限 | 一次能接受的字符数 | 超限时先删已被真实参考帧承载的重复描述；《视觉设定.md》连续性锁的锁面、逐字对白和已批准文字显示位置不得删 |
+| 运镜词汇 | 执行端是否接受具名运镜术语 | 不接受时把运镜写成可见的画面关系变化（主体占画比、边缘进出、前后景遮挡关系），不是换一个术语 |
+| 视频提示词正文语言 | 目标视频模型本次使用 `zh-CN`、`en` 或其他明确语言 | 视频可复制正文使用该语言；逐字对白、VO、OS 仍保留原口语语言，不翻译、不加外语开场 |
+
+## 声音同轨生成
+
+有些执行端只产出画面，声音在后期补；有些执行端在同一次生成里同时产出对白、环境声与音乐。
+两者对正文的要求不同，而且**这是本套件里最容易被忽略的一条**：
+
+- **不同轨**（默认假设）：声音段落写本镜的声音**职责**——谁说话、什么环境声承担信息、
+  转场声落在哪里。它服务于后期与审查，不需要写成可执行指令。
+- **同轨**：声音段落变成会被真的生成出来的东西。可复制正文写完本镜的完整声音时间线：需要的
+  对白、VO、OS、非语言发声、SFX 与相对时机，以及各事件之间承接的 ambience 或静默。逐字对白
+  仍按 [交付形态](delivery-profile.md) 的围栏处理。没有定义的时间与通道会被执行端自行填满，
+  和没写终点的动作会被自行补动作是同一件事。
+
+有对白、VO 或 OS 时，把本镜所有允许的发声列成封闭事件集合，每句带口语语言和逐字起止边界，
+事件间及首尾明确回到已声明环境声或静默。必要的叹气、笑声、拟声或重叠也必须有来源并进入集合。
+Brief、剧本或声音方向若已经点名排除外语前导、语气词、额外人声、重复或其他发声，完整时间线末尾
+必须用正文语言把这些已确认边界合并写一次；“允许事件只有这些”不能替代模型需要的直接指令。
+一句“以上为完整声音时间线”闭合并结束声音段落；其后不再重复同一组排除项，也不自行添加上游没有
+提出的声音禁令。
+
+非画内字幕、caption 和对白文字叠加是独立文字层：除非剧本或创作者明确要求，每条可复制正文
+都用项目提示词语言直接声明全镜任何时刻不生成这些内容，不能只写抽象的“文字层为空”，让对白只
+通过声音与口型发生。这个默认不受画内文字影响。旗帜、标牌、
+屏幕等表面确有 `exact_readable` 义务时，正文逐字保留准确字符、语言、承载面和位置，同时继续关闭
+字幕层；没有画内可读文字义务时，才使用全局无文字要求（`VID` 与 `IMG` 的文字策略优先级不变）。
+
+本套件不规定这些排除写成什么字样。字样属于本镜和项目已声明的提示词语言；不要跨项目搬运一句
+固定后缀，也不要把它当成质量分。
+
+## 参考条件方式
+
+| 声明 | 正文承担什么 |
+|---|---|
+| 无参考 | 文生视频：静态视觉锚点必须自足，写清人物、造型、地点、构图与光线中本镜必要的事实 |
+| 首帧 | 只写从该起点开始的变化；静态外观由参考帧与视觉设定共同拥有 |
+| 首尾帧 | 终点帧只能投影 `SHOT-...` 已写明的终点，不得发明新结果 |
+| 多槽参考 | 每张按 `REF-...` 槽位记账，绑定顺序、路径、中文名、用途与控制/不得控制范围；槽位数不超过声明上限，超出时按重要性取舍并写明放弃了哪些 |
+| 多槽参考（创作者自备） | 同上，但图片不在项目里，槽位改用 `PLAN-...`、定位符写 `IMG-...` 或 `SHOT-...`；正文写法与上一行完全相同，`顺序` 就是创作者在生成界面里挂图的次序。这一行与目标模型无关，任何支持多槽参考的执行端都适用 |
+
+无论哪一种，《视觉设定.md》连续性锁的锁面都留在正文里：参考帧只约束它自己画到的部分，
+跨镜不变的颜色、材质与形制仍然要逐镜带走。
+
+参考条件方式由本镜各槽位的 `用途` 读出，不由“想不想用多模态”决定：只有一张 `起始帧` 是首帧，
+`起始帧` 加 `结束帧` 是首尾帧，`起始帧` 与身份/地理/道具图同时存在或只有这些图时是多槽参考。
+不少执行端把首/尾帧输入与多槽参考输入设为互斥，这时整组统一走多槽参考，起始帧也按参考素材送入——
+不要把一次请求拆成“首帧 + 参考图”。正文里的素材标签按同类素材的 `顺序` 编号，与生产端绑定顺序一致；
+文档手写编号和执行端实际编号不一致时，模型收到的是两套互相矛盾的素材说明，而接口通常不会报错。
+
+“上一段实际视频”和“实际尾帧”描述的是两个通用职责，不等于供应商 role。命中版本化方言时再翻译：
+例如 H3 的 full-reference 使用 `reference_video + reference_image`，不能把实际尾帧误写成与参考模式
+互斥的 `first_frame`。生产 adapter 必须执行同一翻译，不能让文档和请求体使用两套模式。
+
+## 未声明时怎么写
+
+按已接受的镜头边界写完整、自足、不依赖任何执行端特性的正文：视频正文语言回退到项目
+`format.prompt_language`；独立任务没有项目配置时沿用用户当前语言，不固定为英文。写声音职责而不是可执行声音指令，
+不做分段，不罗列画幅备选，用可见的画面关系而不是术语描述运镜。这样的正文在任何执行端上都成立，
+只是没有针对性收敛。**未声明不是缺陷**，不能作为审查阻断理由。
+
+## 自检
+
+- 每条能力轴的取值是创作者声明的，还是我推断的？
+- 视频正文语言来自目标模型档案还是项目回退值？逐字人声是否仍保留原口语语言？
+- 声音同轨时，允许的声音事件、口语语言及事件间底声/静默是否组成完整时间线？
+- 收敛写法有没有顺手改掉已接受的时长、起点或终点？
+- 删重复描述时，锁面、逐字对白和已批准文字显示位置留住了吗？

package/resources/skills/drama-production/upstream/skills/short-drama-video-prompts/scripts/container_check.py

@@ -0,0 +1,329 @@
+#!/usr/bin/env python3
+"""Reconcile an episode's delivery containers against its shot set (`VID-15`).
+
+Every container can be correct on its own while the episode is wrong: a shot
+packed into two containers bills its seconds twice, and a shot packed into none
+disappears although its dialogue, bindings, and keyframe prompts are already
+done. Neither error is visible from inside a single container, so the check is
+a set comparison at episode scope.
+
+The script reads accepted creator files and writes nothing.
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+import sys
+from pathlib import Path
+from typing import Any
+
+
+# Creators run these scripts on whatever interpreter their machine provides, so
+# an unsupported version must say so instead of failing inside an import.
+MINIMUM_PYTHON = (3, 9)
+if sys.version_info < MINIMUM_PYTHON:
+    raise SystemExit(
+        "short-drama needs Python {}.{} or newer; this interpreter is {}.{}".format(
+            *MINIMUM_PYTHON, sys.version_info.major, sys.version_info.minor
+        )
+    )
+
+SCHEMA_VERSION = "1.0.0"
+# A .jsonl file opens with a header record declaring the upstream snapshots its
+# references name. The header is a declaration, not one of the file's records.
+SOURCES_RECORD_TYPE = "sources"
+
+
+class CheckError(ValueError):
+    """The inputs cannot be checked at all, as opposed to failing a check."""
+
+
+def _load_jsonl(path: Path) -> list[dict[str, Any]]:
+    records: list[dict[str, Any]] = []
+    try:
+        text = path.read_text(encoding="utf-8")
+    except (OSError, UnicodeError) as error:
+        raise CheckError(f"unreadable JSONL: {path}") from error
+    for number, line in enumerate(text.splitlines(), start=1):
+        if not line.strip():
+            continue
+        try:
+            record = json.loads(line)
+        except json.JSONDecodeError as error:
+            raise CheckError(f"invalid JSONL at {path.name}:{number}") from error
+        if not isinstance(record, dict):
+            raise CheckError(f"JSONL needs one object per line: {path.name}:{number}")
+        records.append(record)
+    if records and records[0].get("record_type") == SOURCES_RECORD_TYPE:
+        del records[0]
+    return records
+
+
+def _finding(code: str, message: str, **detail: Any) -> dict[str, Any]:
+    return {"code": code, "message": message, **detail}
+
+
+def _seconds(value: Any) -> float | None:
+    if isinstance(value, bool) or not isinstance(value, (int, float)):
+        return None
+    return float(value)
+
+
+# The three conclusions `delivery-container.jsonl.md` requires a container to
+# state about why its members belong together.
+MEMBERSHIP_BASIS_KEYS = (
+    "source_order_contiguous",
+    "binding_chain_equal",
+    "scene_boundary_not_crossed",
+)
+
+
+def _member_shot_id(member: Any) -> str | None:
+    if not isinstance(member, dict):
+        return None
+    ref = member.get("shot_ref")
+    if isinstance(ref, dict) and isinstance(ref.get("record_id"), str):
+        return ref["record_id"]
+    return None
+
+
+def reconcile(
+    containers: list[dict[str, Any]],
+    shots: list[dict[str, Any]],
+) -> dict[str, Any]:
+    findings: list[dict[str, Any]] = []
+    durations = {
+        shot["shot_id"]: _seconds(shot.get("duration_seconds"))
+        for shot in shots
+        if isinstance(shot.get("shot_id"), str)
+    }
+    episode_shots = set(durations)
+
+    owner_of: dict[str, str] = {}
+    packed: set[str] = set()
+    container_total = 0.0
+    for container in containers:
+        container_id = container.get("container_id")
+        if not isinstance(container_id, str):
+            findings.append(
+                _finding("VID15_CONTAINER_HAS_NO_ID", "a container record has no id")
+            )
+            continue
+        members = container.get("members")
+        if not isinstance(members, list) or not members:
+            findings.append(
+                _finding(
+                    "VID15_CONTAINER_HAS_NO_MEMBERS",
+                    "a container carries no members",
+                    container_id=container_id,
+                )
+            )
+            continue
+        # Checkpoint 1: `order` unique, contiguous and ascending. A delivery
+        # package is assembled in this order, so a repeated or skipped number is
+        # an ambiguous cut list, not a cosmetic flaw.
+        orders = [
+            member.get("order") for member in members if isinstance(member, dict)
+        ]
+        if not all(
+            isinstance(value, int) and not isinstance(value, bool) for value in orders
+        ) or orders != list(range(1, len(members) + 1)):
+            findings.append(
+                _finding(
+                    "VID15_MEMBER_ORDER_IS_NOT_A_SEQUENCE",
+                    "members must carry order 1..n, ascending and without gaps",
+                    container_id=container_id,
+                    orders=orders,
+                )
+            )
+
+        # Checkpoint 7: the three membership_basis conclusions. Each is a claim
+        # about why these shots belong in one container; leaving one blank makes
+        # the container's own justification unreadable.
+        basis = container.get("membership_basis")
+        if not isinstance(basis, dict):
+            findings.append(
+                _finding(
+                    "VID15_MEMBERSHIP_BASIS_MISSING",
+                    "a container must record why its members belong together",
+                    container_id=container_id,
+                )
+            )
+        else:
+            blank = sorted(
+                key
+                for key in MEMBERSHIP_BASIS_KEYS
+                if not str(basis.get(key) or "").strip()
+            )
+            if blank:
+                findings.append(
+                    _finding(
+                        "VID15_MEMBERSHIP_BASIS_INCOMPLETE",
+                        "every membership_basis conclusion must be stated",
+                        container_id=container_id,
+                        missing=blank,
+                    )
+                )
+
+        member_total = 0.0
+        for member in members:
+            shot_id = _member_shot_id(member)
+            if shot_id is None:
+                findings.append(
+                    _finding(
+                        "VID15_MEMBER_HAS_NO_SHOT_REF",
+                        "a member does not name the shot it packs",
+                        container_id=container_id,
+                    )
+                )
+                continue
+            if shot_id not in episode_shots:
+                findings.append(
+                    _finding(
+                        "VID15_MEMBER_IS_NOT_AN_EPISODE_SHOT",
+                        "a container packs a shot that is not in this episode",
+                        container_id=container_id,
+                        shot_id=shot_id,
+                    )
+                )
+                continue
+            previous = owner_of.get(shot_id)
+            if previous is not None:
+                findings.append(
+                    _finding(
+                        "VID15_SHOT_PACKED_TWICE",
+                        "a shot belongs to more than one container",
+                        shot_id=shot_id,
+                        container_ids=sorted({previous, container_id}),
+                    )
+                )
+                continue
+            owner_of[shot_id] = container_id
+            packed.add(shot_id)
+            seconds = durations.get(shot_id)
+            if seconds is None:
+                findings.append(
+                    _finding(
+                        "VID15_MEMBER_SHOT_HAS_NO_DURATION",
+                        "a packed shot carries no numeric duration",
+                        container_id=container_id,
+                        shot_id=shot_id,
+                    )
+                )
+                continue
+
+            # Checkpoint 3: the member's own `accepted_duration` against the
+            # shot it projects. This field was never read: the total was summed
+            # straight from `shots.jsonl`, so a member could claim 12.0s for a
+            # 6.3s shot and pass, and the execution end would be handed the 12.0.
+            # `motion_timing_check` guards exactly this staleness class; this
+            # file had no equivalent.
+            claimed = _seconds(member.get("accepted_duration"))
+            if claimed is None:
+                findings.append(
+                    _finding(
+                        "VID15_MEMBER_HAS_NO_ACCEPTED_DURATION",
+                        "a member must project the duration it packs",
+                        container_id=container_id,
+                        shot_id=shot_id,
+                    )
+                )
+            elif abs(claimed - seconds) > 1e-6:
+                findings.append(
+                    _finding(
+                        "VID15_MEMBER_DURATION_IS_STALE",
+                        "a member's accepted_duration does not match its shot",
+                        container_id=container_id,
+                        shot_id=shot_id,
+                        claimed=claimed,
+                        accepted=seconds,
+                    )
+                )
+            member_total += seconds
+        stated = _seconds(container.get("container_duration"))
+        if stated is None:
+            findings.append(
+                _finding(
+                    "VID15_CONTAINER_DURATION_MISSING",
+                    "container_duration must be present and a number of seconds",
+                    container_id=container_id,
+                )
+            )
+        elif abs(stated - member_total) > 1e-6:
+            findings.append(
+                _finding(
+                    "VID15_CONTAINER_DURATION_IS_NOT_THE_SUM",
+                    "container_duration does not equal its members' durations",
+                    container_id=container_id,
+                    stated=stated,
+                    computed=round(member_total, 6),
+                )
+            )
+        container_total += member_total
+
+    loose = sorted(episode_shots - packed)
+    # A shot whose duration is still open is a legal state upstream, so it is
+    # held out of the arithmetic instead of being reported as an error. Only a
+    # shot packed into a container must already have one, because the
+    # container's own duration claim depends on it.
+    unmeasured = sorted(
+        shot_id for shot_id in episode_shots if durations.get(shot_id) is None
+    )
+    loose_total = sum(
+        durations[shot_id] or 0.0 for shot_id in loose if durations.get(shot_id) is not None
+    )
+
+    episode_total = sum(value for value in durations.values() if value is not None)
+    if abs((container_total + loose_total) - episode_total) > 1e-6:
+        findings.append(
+            _finding(
+                "VID15_EPISODE_TOTAL_DOES_NOT_RECONCILE",
+                "containers plus loose shots do not add up to the episode total",
+                packed_seconds=round(container_total, 6),
+                loose_seconds=round(loose_total, 6),
+                episode_seconds=round(episode_total, 6),
+            )
+        )
+
+    return {
+        "schema_version": SCHEMA_VERSION,
+        "containers": len(containers),
+        "episode_shots": len(episode_shots),
+        "packed_shots": len(packed),
+        # Loose shots are legal: containers need not cover everything. They are
+        # reported so the count is a decision rather than an oversight.
+        "loose_shots": loose,
+        # Reported, not a finding: these shots are excluded from every total
+        # above, so a caller can tell an incomplete episode from a wrong one.
+        "unmeasured_shots": unmeasured,
+        "packed_seconds": round(container_total, 6),
+        "loose_seconds": round(loose_total, 6),
+        "episode_seconds": round(episode_total, 6),
+        "findings": findings,
+        "status": "pass" if not findings else "fail",
+    }
+
+
+def build_parser() -> argparse.ArgumentParser:
+    parser = argparse.ArgumentParser(
+        description="Reconcile delivery containers with the episode's shot set."
+    )
+    parser.add_argument("containers", type=Path, help="the episode containers JSONL")
+    parser.add_argument("--shots", type=Path, required=True)
+    return parser
+
+
+def main(argv: list[str] | None = None) -> int:
+    args = build_parser().parse_args(argv)
+    try:
+        result = reconcile(_load_jsonl(args.containers), _load_jsonl(args.shots))
+    except CheckError as error:
+        print(f"{type(error).__name__}: {error}", file=sys.stderr)
+        return 2
+    print(json.dumps(result, ensure_ascii=True, sort_keys=True))
+    return 0 if result["status"] == "pass" else 1
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())