The first scan reported raw hash collisions and I then analysed the noise. Every
bank occupied 375 of 378 time bins. Three things were wrong, each traceable to
prior art rather than to tuning:

Offset consistency was missing entirely, and it is half the algorithm. Wang's
2003 Shazam paper is explicit: real matches agree on a time offset, so the
(query_t, ref_t) scatterplot shows a diagonal and the offset histogram spikes.
Counting collisions without that step counts coincidences.

Fixed offset is wrong anyway for this material. Sonnleitner, Arzt & Widmer
(ISMIR 2016, DJ mix monitoring) allow the offset line a free slope, because a DJ
plays material at a different rate — and a livecoder changing setcps does the
same, across 80 to 170 BPM in one set.

Raw counts measure loudness. Measured, not assumed: 8 of 18 banks had their
global peak in the final minute, at 1.3-2.5x density. The finale is loud and
broadband, so it makes more spectral peaks, so it matches everything better.
Scores are now normalised by each window's own query-hash count and by each
bank's reference size.

But the real problem was the reference unit, and PLN named the fix: "comparing
to prior recs, of each track of the opal songs for example (almost the same
lineup in the end!)". A 0.4s dry one-shot cannot survive a PA and a 320 kbps
encode — too few hashes, below the detection floor by construction. Meanwhile
Prod/Opal26_master/tracks holds FIFTEEN ear-verified per-track renders, 154 to
524 seconds each, same performer, same sample library, same patches.

That changes the task from sample detection to VERSION identification, where the
literature is unambiguous: when audio is related but not identical you stop
matching exact time-frequency points and match sequences of features instead
(Serra et al. on covers, McFee & Ellis on structure). So `crossmatch` embeds
15s windows as MFCC mean+std, chroma and spectral contrast, matches by cosine
against every reference window, and median-filters over 45s because a track
lasts minutes and a one-window flicker is noise by definition.

Standardising before the cosine is not cosmetic: raw MFCC coefficients differ in
scale by orders of magnitude, so without it the first two — level and spectral
tilt — decide everything, and level is exactly what a room mic gets wrong.

The docstring says what it cannot do. The three cafe_* tracks are the same
instruments at nearly the same tempo and no timbral profile will separate them,
so the output reports a runner-up margin and flags thin ones: a win by 0.01 is
not a win.

Also included, from the custom-mode thread: the generated Components entry table
(40 of 48 controls need changing, all of row A already matches), and the sniff
that caught Components driving the device with documented feature controls —
9F 0B 7F then CC 30 = 6 — plus an undocumented cmd 0x05 whose read/write forms
match the audiocontrol project's reverse-engineered protocol.