fix(asset_based): handle scenes with empty narrations list

[jonathanzhan1975]

Summary

The asset_based pipeline (custom_media in the UI) crashes with IndexError: list index out of range when a scene has an empty narrations list. This is a legitimate LLM output for cover slides, chapter dividers, or any purely-visual scene — it is reliably reproducible by uploading a title slide as the first asset.

Reproduction

1. Open the Custom Media (自定义素材) pipeline.
2. Upload 9 PPT slide images. The first slide is a cover with no spoken narration.
3. Provide an article / intent and click Generate.

Observed log:

INFO  asset_based:initialize_storyboard:530 - ✅ Created storyboard with 9 scenes
INFO  asset_based:produce_assets:544 - 🎬 Producing scene videos...
INFO  asset_based:produce_assets:555 - Producing scene 1/9...
INFO  asset_based:produce_assets:574 - Scene 1 has 0 narration(s)
ERROR linear:handle_exception:161 - Pipeline execution failed: list index out of range

The pipeline aborts before any video is produced.

Root cause

The same defensive read pattern appears in four places:

narrations = scene.get("narrations", [scene.get("narration", "")])
if isinstance(narrations, str):
    narrations = [narrations]

It guards against missing keys and string values, but does not handle narrations being an empty list (which happens when LLM legitimately decides a scene has no narration). All four call sites silently degrade or crash:

Location	Behavior on empty narrations
L382 (log preview)	Empty preview line, hides the gap
L448-453 (all_narrations.extend)	Scene silently disappears from context.narrations
L495-505 (StoryboardFrame.narration)	Empty subtitle string
L570-635 (TTS production)	💥 IndexError at narration_audios[0] (L635) — user-facing crash

Only the last one crashes, but the other three create silent data drift that is harder to diagnose later.

Fix

1. Centralize the read with a single _coerce_narrations(scene) helper that normalizes any of {missing, None, str, list, list-with-empty-strings} into a non-empty list of stripped, non-empty strings, falling back to [""] when the scene is genuinely silent. The fallback preserves the existing length-based assumptions (len(narrations) >= 1) so callers can still index safely; callers that care about content can detect the gap via the empty string.

2. Replace all four call sites with narrations = _coerce_narrations(scene).

3. In the TTS loop (produce_assets):

   • Skip empty narration text (don't waste a TTS call).
   • When a scene produces zero audio segments, synthesize a 3-second silent track with ffmpeg (anullsrc) so downstream composition still has a valid audio_path. This lets cover slides render as a 3-second silent shot instead of crashing.

The single-narration and multi-narration concat branches are unchanged.

Diff size

pixelle_video/pipelines/asset_based.py: +43 / −14 lines, single file.

Verification

• Reproduced the crash with a 9-slide PPT (slide 1 = cover with no narration).
• After the fix, the pipeline completes; slide 1 plays as a 3-second silent shot, scenes 2–9 narrate normally.
• Single-narration and multi-narration scenes are unaffected (verified via the existing branches at L623-664).

Out of scope

• Whether the cover slide should get LLM-generated narration in the first place is a separate question (a cover-detection heuristic in generate_content would be one path). This PR only ensures the pipeline does not crash when LLM legitimately decides a scene has no narration.
• The 3-second silent fallback duration is currently hardcoded; making it configurable (e.g. config.cover_silent_duration) is a follow-up.

Why no tests

The pipeline depends on LLM, ffmpeg, and a runtime task directory; there is no existing unit-test scaffold for produce_assets. _coerce_narrations itself is a pure function and trivially testable if the project later adopts a test harness — happy to add tests in a follow-up if the maintainer wants.

🤖 Generated with Claude Code

[CLAassistant]

All committers have signed the CLA.