0006. Time references resolve against an asset manifest

Extended by 0014: captions now resolve word references to where the word plays, and generate cues from the transcript.

Context

The model never sees pixels or hears audio. It edits against what an ingest step knows: a transcript with word timings, shot boundaries, silences. If the file stored seconds, a re-transcription or a re-ingest would silently shift every cut.

Decision

in, out, at and caption times accept TimeRefs: "word:12", "word:23.end", "shot:1", "silence:2", "marker:intro", resolved by the resolver against the manifest. A gesture that changes such a prop writes seconds. The manifest is data the editor and the player receive with the file.

Consequences

• The model trims by meaning ("the sentence from word 12 to 23"), which is how it reasons anyway.
• The lint checks every src against the manifest (missing-asset).
• Captions take composition seconds today; generating <Caption> children from the transcript is a known gap.