Caught on glass during the WP2 baseline, on the first stream the metric ever
measured: judder=0 mode=0 cadN=0 disorder=119, every second, on a perfectly
healthy 118 fps stream with the panel period correctly learned at 8.33 ms.
Every single present was scoring as disordered. The cause was the vendor quirk
decode/display.rs already documents — Android's render callback can deliver a
garbage far-future system_nano on a session's first frames. The rule "hold the
later instant so one reordered delivery cannot corrupt the following spacings"
then latched onto that stamp permanently: every real timestamp afterwards was
behind it, so nothing was ever scored again for the rest of the session.
The rule was right for what it was written for and wrong past a bound. A step
backwards of a few refreshes IS a reordered delivery and the later instant
should win; a step backwards of an hour is a bogus stamp and the run must
re-anchor onto the new sample. One bad sample now costs one sample.
Both implementations get the bound and the regression test, since the two must
agree; the Swift port would otherwise have shipped the same latch-up.
Also makes the statistic self-diagnosing, which is what turned a puzzling
result into a five-minute diagnosis: summary() returning None was
indistinguishable from a window of perfectly smooth zeros in a log line, so
"no cadence is being scored at all" looked exactly like "no judder". The
pf.present line now carries the raw sample/stall/disorder counts and the period
the run is quantising against, whatever the evidence bar.
Gates: punktfunk-core 190 tests green (22 in phase, incl. the new
a_garbage_far_future_stamp_does_not_wedge_the_run); swift test
--filter PresentIntervalsTests 11/11 green; fmt clean; on-glass re-run against
.173 now reports judder 8-36permille with cadN ~119/s.
enricobuehler
1099c94ca3