PLN downloaded the CosmicFest capture and called it "an interesting test". It is:
a 63-minute handheld stage-mic MP3, and all three boundary lenses the toolbox
owns need something it does not have. Orbit-activity needs stems, there are
none. The gig-log `track` events are the real tracklist, but the 1 Hz recorder
never ran that night — newest session is 2026-08-20. Gap detection assumes a
set stops between tracks, and a livecoded set does not.

What is left is audio, plus one large advantage: we know what was prepped, so
this is matching over ~23 candidates rather than open discovery. PLN warns "the
set was slightly altered", so candidates are treated as candidates — presence
is strong evidence, absence is weak.

Two lenses, because neither is sufficient. Tempo finds seams but cannot name
them: five candidates sit at 120 BPM and three at 124, and tempo from audio is
only ever determined up to a factor of two. Sample signature names a segment,
and PLN pointed at exactly the right prior art — sample_tfidf.py already
computes which sounds IDENTIFY a track. take5 and love_parade are df=1;
jungle_breaks is in 80 tracks and worthless. 15 of 23 candidates carry a sound
rare enough (df<=6) to fingerprint; 19 such banks, 533k reference hashes.

Constellation fingerprints rather than correlation, because the reference is a
dry one-shot off disk and the query is that sound through a PA, a room, a
limiter, "ppl messed with it", and a 320 kbps encode. Peak constellations encode
only which time-frequency points stick up locally, which survives all of that.

My first attempt at extracting banks was wrong and the output said so: 9 tokens
across 23 tracks, every count 1, and a label reading "shares every bank" that
actually meant "nothing parsed". The regex looked for `s "name"` while the
ParVagues idiom applies a bare string — `$ "meth_bass" # n "..."`. The fix was
not a better regex, it was using the parser that already existed.

The docstring carries the limits because they are easy to launder: a detection
means the SOUND occurred, not that the TRACK played, since banks get reused
live. Recall is asymmetric — a track whose signature never sounded is invisible.
And vote counts are not comparable between banks, because a 582-file bank offers
far more chances to match than a 4-file one.

Also adds `lcxl3.py sniff`: capture raw SysEx in full. The custom-mode format is
undocumented, and Components writes modes to the device over SysEx — capturing
those bytes is the difference between a layout authored in lcxl_grid.py and one
merely stored in a web app.