[MustBeSimo]

The bake recipe could leave a paused video waiting for a decoded-frame callback before initiating the seek that produces that frame. The advertised 1080p preset also retained the input resolution. Correct those instructions and the integration/licensing errors found in the motion-stack references.

• Schedule seeks through requestAnimationFrame; use video-frame callbacks only to observe decoding, and verify byte-range serving. Prefer a measured short GOP over mandatory all-intra encoding; bound the image cache and avoid upscaling small frame sequences.
• Make the landscape export actually fit/pad to 1920×1080 and preserve H.264, yuv420p and faststart when adding captions. Distinguish timed browser recording from frame-deterministic rendering.
• Correct Theatre core versus studio usage and licences, Draco/KTX licensing, GSAP restrictions and Anime.js's release date; link primary sources.
• Clarify Lenis's RAF units, fixed-step physics, splat asset validation/readiness, and repository-only video documentation links for installed skills.

Validation:

• Full npm test, skill validator, link and whitespace checks passed.
• Executed all eight FFmpeg commands using generated fixtures. Verified dimensions, H.264/AAC, pixel format, MP4 header order, 96-frame MP4/WebM/WebP outputs and keyframe spacing.
• Chrome 156 reproduced a paused rVFC-driven seek stalling at time 0. RAF-driven seeks decoded 1.4 → 0.4 → 1.4 seconds with distinct/repeatable frame pixels.
• On the synthetic 3.2-second, 320×180 fixture, GOP 15 produced 148,305 bytes versus 418,212 bytes for GOP 1 at the same CRF (65% smaller). This is a fixture result, not a mobile performance benchmark.
• All three skills remain text-only and under 100 KB combined.

Reference-only change: this does not claim new end-to-end tests of every third-party engine or measured LLM routing accuracy. No release/version bump is included.