Eyevinn / go-608 / 30572119609

84%

push

github

fix(schedule): pre-roll the pop-on build so a cue's time is when it appears

A pop-on caption is two transmissions -- a build into non-displayed memory, then
the EOC that flips it on screen -- and both drain at one byte pair per frame. The
build was transmitted starting AT the cue's time, so the caption only became
visible a build later: measured 367 ms for "HELLO" and 433 ms for "SECOND ONE"
at 30 fps, halving at 60 fps. Longer text appeared later.

schedule now backdates the build so its EOC lands on the pushed time. That is
FlipTiming's new FlipOnTime default; FlipAfterBuild restores the old behaviour,
reachable from the CLI as -no-preroll on go608-inject and go608-extract.

WebVTT -> 608 -> WebVTT is now frame-exact where it used to slide:

    in         1.000->3.000  3.000->5.000  5.000->6.000
    before     1.433->3.467  3.467->5.367  5.367->6.000
    after      1.000->3.000  3.000->5.000  5.000->6.000

The old timing also compounded -- re-converting shifted it again, 1.000 -> 1.367
-> 1.667 -- and is now stable across passes. Ends were always exact, because the
clearing EDM is itself the visible change; a batch not ending in an EOC is
likewise never backdated.

This reconciles the containers' timing conventions as a side effect. An SCC
entry's timecode is when its first pair is transmitted; a WebVTT cue's start is
when the caption is displayed; they differ by exactly the build. So SCC -> text
-> SCC now returns the original timecodes (00:00:01:00, was 00:00:01:12), the
only asymmetry left being the terminating EDM cue.Compile appends for a dangling
final cue. 608 -> text is unchanged: it always reported actual display time.

generate is unaffected -- Generator and BuildUnitCues already pushed build and
EOC as separate batches, which is this pre-roll done by hand. Verified: the
split serializes to byte-identical output, so only placement changes.

The inject test previously allowed a 500 ms window around a 1 s cue, which is
how this stayed invi... (continued)

53 of 57 new or added lines in 4 files covered. (92.98%)

3412 of 4049 relevant lines covered (84.27%)

0.97 hits per line