benchmarks/idmt-first-run - #335

Summary

Continues #321, which established the first reproducible IDMT baseline and
identified kick recall as the standout weakness (R = 0.285, "missing ~95%" on
acoustic RealDrum).

This PR diagnoses that failure and fixes it with a single constant change:
KICK_LFER_MIN lowered from 0.32 to 0.08. Also adds the diagnostic tooling
used to arrive at that value.

Diagnosis

31 of 95 IDMT tracks scored a kick F-measure of exactly 0.000, while precision
was frequently exactly 1.000 wherever the rule fired. That signature indicates a
rule that is correct whenever it fires but almost never fires.

Stage attribution across a controlled three-track comparison:

Onset detection was not at fault. Every reference kick was detected. All 22
losses occurred at classification, all on the same condition: lfer_low. Peak
frequency passed in every case (43–65 Hz). The lost kicks were labelled as open
hi-hat (12) or closed hi-hat (10).

Root cause

LFER is the proportion of spectral energy below 150 Hz - a ratio, so sensitive
to whatever else occupies the analysis window.

KICK_LFER_MIN was calibrated on isolated drum stems, where the denominator
holds little but the kick. In mixed audio a simultaneous hi-hat adds
high-frequency energy to the denominator while contributing nothing to the
numerator.

Across 5,387 annotated IDMT onsets:

The threshold of 0.32 sat above the median of the class it was meant to
detect.

Threshold selection

Grid sweep over 420 combinations of KICK_FREQ_MIN, KICK_FREQ_MAX and
KICK_LFER_MIN, scored against all 5,387 onsets. Holding the frequency band at
its live 40–140 Hz:

(Onset-level micro-averages, not directly comparable to the track-level
macro-averages reported below.)

The best combination overall (0–180 Hz, LFER 0.08) scored F = 0.855. Rejected.
A lower bound of 0 Hz only scores well because it admits onsets with a measured
peak frequency of 0 Hz, a measurement artefact rather than a drum. Raising the
upper bound to 180 Hz extends the kick band into the snare band (starts at
120 Hz) and the tom buckets, which append labels additively rather than
exclusively. A gain of 0.021 did not justify that coupling.

Frequency bounds unchanged. One constant moved.

Results: IDMT-SMT-Drums V2

Same setup as #321 - 95 #MIX files, 50 ms onset window, macro-average across
items.

Snare and hi-hat are unchanged to four decimal places.

By subset, against the table in #321:

#321 called out acoustic RealDrum as the worst case at KD F = 0.088. It is now
the best subset at 0.907.

Zero-scoring kick tracks: 31 / 95 → 0 / 95. Kick estimates 2,376 against
2,179 references - slight over-detection replaces severe under-detection.

Results: MDB-Drums (23 tracks)

Re-run specifically to test for collateral damage, since the tom rule suppresses
low_tom whenever kick has already been appended to the same event - a
regression IDMT cannot detect, as it does not annotate toms.

No regression occurred, though not for the reason anticipated: the tom rules
produce zero estimates in both runs, so there was nothing to suppress.

MDB kick precision fell slightly, 0.690 → 0.663, while recall rose 0.341 →
0.852. Three previously near-perfect tracks now over-fire: Beatles
(0.990 → 0.847), Hendrix (1.000 → 0.810), Rockabilly (0.960 → 0.684). Five moved
from non-functional to working: Shadows (0.000 → 0.852), FreeJazz
(0.103 → 0.739), Disco (0.033 → 0.696), LatinJazz (0.000 → 0.659), BebopJazz
(0.000 → 0.657).

Tooling added

• analyze_idmt_dataset.py: --diagnose-kick (per-reference stage attribution
  plus which rule condition failed), per-onset feature capture, and a threshold
  grid sweep (--sweep-csv, --top) that scores candidate rules without
  re-processing audio.
• Two guards against sample artefacts: a warning when a sample contains no
  non-kick onsets, and a warning when the best candidate produces zero false
  positives. Both fire on small samples where F = 1.000 is structural rather
  than real.

Not in scope

#321 noted that IDMT lacks crash/ride/tom coverage, which is why MDB-Drums was
added. With MDB now running, the gap is measurable: ride has 835 reference
annotations and has never produced a single estimate; toms have 64 references
and likewise produce none. These are unreachable rules rather than mis-tuned
thresholds, and now represent a larger gap than the one closed here.

Testing

• IDMT full benchmark re-run (95 tracks), compared against the benchmarks/idmt-first-run #321 baseline.
• MDB full benchmark re-run (23 tracks), before/after across all nine annotated
  classes.
• No change to any instrument other than kick on either dataset.
• in progress full unit tests
• in progress E2E test

Housekeeping
in progress - Write up of work forthcoming and I will add results to this PR before merging
[DONE]: Updated CHANGELOG.md and added these changes to the Unreleased section

cc @nanaoto - this builds directly on the adapter you contributed.