ActionCheese Sync 0.4.3

Free, open-source multicam sync for Premiere and DaVinci Resolve. Aligns
footage by its sound, so cameras whose clocks were never set still land
together.

This release comes out of a real wedding party that synced badly, and the
footage behind it. Thank you to the videographer who sent the comparison
report — it is the whole reason any of this could be measured.

Music confused the matcher. It no longer does.

Alignment is found by correlating two recordings and taking the lag that
scores best. The check for how much best it scored was effectively off: a
lag no better than the next one along was accepted on the strength of its
score alone.

Dance-floor music repeats, so it produces lags that score alike, and a tie
says nothing about which one is right. Three such matches reached the solver
on that wedding — winning by 0.4%, 0.7% and 0.8% — and one of them put a clip
two and a half hours from where it belonged.

On that footage, clips landing on the exact frame went from 18 of 41 to 30.
A healthy four-camera project is unchanged to the frame: its weakest accepted
match won by 50%, nowhere near the new floor. The floor stays deliberately
low — reverberant rooms give correct matches only a modest margin, and
demanding a tall peak throws away good matches. The rule is that the winner
must not be a coin flip.

"24 in separate synced groups" was not a sync failure, and now says so

Two recordings from one device never overlap, so no amount of sound can
relate them. When that splits a shoot into groups, the clips inside each
group are frame-accurate — it is only where the groups sit relative to each
other that is open, and the report never said which was which.

It does now: each group reports whether its position came from a device's
clock or whether it was simply put after everything else, which makes no
claim about position at all. On the wedding, 28 of 32 clips were already
correct to a frame. That is the difference between "the sync is broken" and
"check these few joins".

The comparison report was grading the wrong moment

Matching runs at a lowered threshold so a second pass has candidates, and the
solver then takes only the confident subset. The report was recording that
first, provisional verdict. One report claimed 49 matches accepted and 33 of
them wrong, when only 32 ever reached the solver and 5 were wrong — it was
blaming the matcher for matches it had already thrown away.

If you have sent a comparison, please send another from this version; it will
tell us something the first one could not.

Requirements: Apple silicon Mac, macOS 11 or later. Intel Macs are not
supported. ffmpeg is bundled, so nothing else to install.

388 tests pass.