#4119 @swamp/video: follow-ups from the initial review (card length, input validation, OCR frames on disk)

Non-blocking findings from the final verify-reviews run on the initial @swamp/video PR (swamp-club #4103):

• medium: an intro or outro card shorter than the 0.7s crossfade gives negative xfade/adelay/apad offsets; ffmpeg fails after the loudness pass with an error about the filter graph. Require seconds > 0.7 in the cards schemas (render CardSchema and CardsResourceSchema) or fail early naming the card.
• low: sampled OCR frames in .video/frames/ stay on disk unpixelated; delete them after analysis or pixelate them.
• low: edit build ignores a custom edit.txt path the take was last edited through and may skip the backup of the default file.
• low: screen redactions are cast, not schema-validated, before render puts them in the filter graph.
• low: video-recut has no early title check for simple cards (video-edit has one).
• low: a malformed max-pause header in edit.txt yields NaN settings; validate it.
• low: the loudness measurement passes the whole select expression inline in argv; very long takes could hit the per-argument size limit. Use a filter script file like render does.
• low: two misplaced doc comments in edit.ts.