Summary

skills/product-launch-video/scripts/assemble-index.mjs is not idempotent. The first run hoists each approved frame <video> out of its frame file and writes the frame file back with only a placeholder comment (<!-- approved frame video hoisted by assemble-index -->). A second run finds no <video> in the frame files, so it rebuilds index.html with no footage, and it exits cleanly without any error or warning.

Repro (hyperframes 0.8.163)

1. A project with a frame compositions/frames/02-x.html that holds an approved video (data-frame-video="approved", with data-start / data-duration / data-track-index / layout attrs).
2. Run assemble-index.mjs, and index.html has the root-level <video>. ✅
3. Edit any frame (e.g. a caption fix) and run assemble-index.mjs again.
4. index.html has no <video>, the frame file has only the hoist comment, and the render shows empty device frames.

Cause

guardFrame() → hoistApprovedVideos() returns the stripped HTML as repairedHtml, and the caller persists it (writeFileSync(compAbs, guard.repairedHtml), ~line 417). The hoist is meant to happen at assembly time, but it mutates the source of truth, so the operation destroys data on re-run.

Suggested fix

Either of these would work:

• Don't write the hoist back: keep the <video> in the frame source and strip it only in the in-memory copy used for index.html (the root-dimension repair can still be persisted on its own).
• Or make the placeholder carry the full tag (e.g. a data-hoisted <template> or an attribute-bearing comment) so a re-run can re-hoist from it.

A regression test: run assembly twice on the same fixture and assert that the <video> count in index.html is equal across runs.

Impact

Any edit-and-reassemble loop, which the review loop encourages, silently loses all footage. The only signal is a blank render, found after a full render cycle.