fix(pipelines): make the screen-demo manifest load

[bbudaedu]

The defect

pipeline_defs/screen-demo.yaml carries a top-level production_modes block the manifest schema does not allow, so load_pipeline("screen-demo") raises. The pipeline is marked production stability in AGENT_GUIDE.md's table.

lib/checkpoint.py degrades gracefully when a manifest cannot be parsed. That is right for a corrupt user manifest, but it means a manifest shipped in this repo is silently ignored rather than loudly rejected — and both fallbacks then misfire.

1. The pipeline cannot start

The stage-order fallback returns the canonical list, which inserts research and proposal ahead of screen-demo's real first stage:

get_pipeline_stages("screen-demo")
  ->  [research, proposal, idea, script, scene_plan, assets, edit, compose, publish]
manifest declares
  ->  [idea, script, scene_plan, assets, edit, compose, publish]

So the opening checkpoint of every screen-demo run is rejected:

write_checkpoint(..., stage="idea", status="completed", human_approved=True)

CheckpointValidationError: PREREQUISITE VIOLATION: stage 'idea' cannot advance;
  incomplete or missing: ['research', 'proposal']

The other five idea-first pipelines (talking-head, hybrid, clip-factory, …) clear that same check, which isolates the cause to the unloadable manifest rather than the prerequisite rule.

2. Its approval gates fail open

_stage_requires_approval("screen-demo", "idea")  ->  None      # manifest declares True

None means the caller's own human_approval_required flag wins. All five declared gates — idea, script, scene_plan, assets, publish — stop being binding, against AGENT_GUIDE.md:

Read human_approval_default from the pipeline manifest per stage. The manifest value is binding — never re-judge it.

The repair

production_modes is a documented, first-class feature rather than stray YAML:

• skills/pipelines/screen-demo/idea-director.md builds its entire real_capture vs synthetic_terminal decision table on it, down to which render runtimes each mode permits.
• AGENT_GUIDE.md routes agents to this manifest to make that choice: "Picking a screen-recording mode (real capture vs synthetic terminal)? → pipeline_defs/screen-demo.yaml + skills/pipelines/screen-demo/idea-director.md."
• Nothing in Python reads it — it is agent-facing declarative content, like best_for and agent_skills elsewhere in the manifest.

So the schema is what is stale, and it gains a typed definition for the block, following the shape of the existing stages entries.

extensions was considered as a home and rejected: it is a fixed set of four capability booleans (custom_scripts, custom_playbooks, custom_skills, custom_tools), not a free-form bucket.

Coverage

Existing manifest tests load talking-head, framework-smoke, animated-explainer and documentary-montage by name, so the one broken manifest was never touched — the same gap that hides this class of bug elsewhere in the repo.

tests/contracts/test_pipeline_catalog.py iterates pipeline_defs/ and asserts three things, one per symptom above:

• every shipped manifest validates
• the stage order the checkpoint writer actually uses equals the one the manifest declares
• every declared human_approval_default is what _stage_requires_approval resolves

Verification

• The 3 screen-demo cases fail on the unfixed tree; all 40 pass after.
• Full suite: 964 → 1004 passed, 10 skipped, no regressions.
• make lint: passed.

Independent of #468 and #469 — no overlapping files.

🤖 Generated with Claude Code