docs(video): narrated script with short-form captions

The video now has a voice. Reading blocks of subtitles off a silent screen for
three minutes asks more of a judge than it returns, and the words a person says
are not the words a person reads.

docs/demo-script.json replaces demo-subtitles.json as the only source. Each line
carries the spoken English, a Korean gloss for whoever edits, the timing, and
the shot. scripts/make-video-assets.mjs turns it into the narration text, the
short-form caption file, two subtitle tracks and the shot table in this doc.

The script is rewritten for the ear: contractions, direct address, sentences
that vary in length. 29 lines, 341 words, exactly 3:00.

demo.short.ass carries the caption style rather than leaving it to be rebuilt by
hand: three or four words a burst, Arial Black at 96px, white on a heavy black
outline, each one overshooting to 112% for two frames so it lands. 96 bursts,
3.6 words and 1.9 seconds each on average.

The generator refuses output that would not work, and each refusal is there
because the first draft tripped it: a line the voice cannot fit, over three and
a half seconds of dead air after a line, a caption on screen for under half a
second, overlapping cues, a dash a voice reads as an unplanned pause. Setting
one word per caption makes it exit non-zero, so the check is not decorative.

The old demo-subtitles.json and make-srt.mjs are deleted. Two sources for one
script is how the shot table and the cue list got out of step the first time.