Video Batch Runner

# Video Batch Runner

## Use When
- Approved image, script, voice, or prompt-plan inputs already exist and the next step is a hosted talking-head, Seedance, or reference-motion render.
- The output must preserve a local render manifest, source basis, hosted handles,
  output URLs, downloaded video files when the host can fetch them, and a
  pollable checkpoint.

## Do Not Use When
- The task belongs to ideation, QA, or another released skill listed in the handoff section.
- Required inputs are missing and guessing would change the result.
- Task class, hook logic, storyboard, or reference policy is still unresolved.
  Use `video-generation` and `video-request-architect` first.

## Execution Boundary
- This runner validates and executes normalized video requests. It must not make
  creative strategy, task-classification, or reference-policy decisions.
- Default to the highest practical render quality for realism-sensitive human video; step down only for an explicit cheap draft, latency test, or provider limit, and persist the choice.
- Default creative format is short-form vertical `9:16`; use
  `creativeFormat: "instagram_meta_ads"` or explicit `aspectRatio: "3:4"` when
  the target is Instagram Meta Ads, except Wanx 2.1 endpoints which currently
  accept only `9:16` or `16:9`.
- Released provider is `hosted-media` only. Direct provider routes, ad hoc
  structured motion-control fields, and unbound media roles are not released.
- `video-kling-v2-6-pro-motion-control` is only reference-motion transfer with
  a reference image plus reference motion video.
- Wanx 2.1 Turbo endpoints are `video-wanx2-1-t2v-turbo` and
  `video-wanx2-1-i2v-turbo`; duration is fixed at 5 seconds, resolution is
  `480p` or `720p`, and image-to-video input must already be a remote HTTP(S)
  image URL.

## Source And Route
- Source from the active project/client manifests first. Do not reuse another
  client directory as the default source basis.
- Required for all routes: hosted capability request, `jobId`, `assetPurpose`,
  `sourceBasis`, `localOutputDir`, `provider: "hosted-media"`, and `model`.
- Talking head requires approved `image`, approved `audio`, and script/concept
  source. Seedance requires intentional `final_prompt` or `prompt_summary` plus
  `promptPlan.prompt_storyline`, and required media for the selected mode.
- `promptPlan.camera`, `promptPlan.shotType`, and `promptPlan.motion` constrain
  prompt text only; do not map them to provider-native trajectory fields.

## Request Boundary
- `--request` must point to a hosted capability request JSON with explicit
  `capability`, `operation`, `operationId`, and normalized video `input`.
- The agent writes normalized `request.json` and `poll-request.json` under
  `.postplus`; polling must use the poll request built from `output.data.id`.
- Keep internal requests/responses under `.postplus` when they are not final
  handoff artifacts; keep final renders/manifests in the active render folder.

## Seedance Prompt Boundary
- Prefer `prompt_summary` plus `promptPlan.prompt_storyline` over a dense
  paragraph. Each segment should carry subject, storyline, scene/style, camera
  language, sound intent, continuity constraints, and explicit `[image 1]` /
  `[audio 1]` / `[video 1]` bindings as applicable.
- Do not use deprecated `promptPlan.storyboardTimeline`; use
  `promptPlan.prompt_storyline`.
- Choose duration by spoken density. Use a shorter supported bucket or rewrite
  the segment instead of padding silent holds.

## Rev