Rebuilding lesson sources from repaired transcripts

Use the repair before requesting the same import again#

Some lessons had usable manual transcripts while automated caption imports were still failing. The rebuild queued another remote import before looking for manual or managed text, and could return without rebuilding the segments and chunks needed downstream.

The job now selects a preferred ready transcript version and checks that it contains usable text or nonempty chunks. If usable material exists, canonical ingestion and the guide and embedding jobs can proceed. If not, transcript acquisition is still required.

A ready flag doesn't establish usable, current text#

A source marked ready can still be empty. One regression creates that state and expects transcript repair, with no study-guide generation. The status alone must not skip acquisition.

A transcript can also belong to media that has since been replaced. The shared material service compares media-derived sources with the lesson's current video or uploaded file. Guide assembly and assistant retrieval both use that check to exclude stale chunks.

Manual transcripts are explicitly supplied lesson material rather than automatically linked to a remote-media ID. Reusing one doesn't verify that it accurately describes the video; that remains part of the source review.

What the development record established#

The June report records failing-then-passing tests for manual transcript reuse and empty ready sources, followed by regressions for media replacement and stale chunks in both consumers. Those are dated local results. The repair changes the pipeline's ordering and source checks; it doesn't create missing transcripts or establish a fresh run here.