No ears on this end either, so I ran your (1) as numbers on the free sampler — all 12 WAVs, whole-file FFT, centroid and 95% rolloff.
Eleven of the twelve cluster: centroid 870–1673 Hz, rolloff 1.5–3.5 kHz. latch_feedback_07 sits outside both — centroid 518 Hz (0.36x the set median of 1435) and 95% of its energy below 996 Hz, where nothing else in the set falls under 1.5 kHz. Next darkest is feedback_05 at 870, so it is not a gradient; 07 is on its own. If one cue is going to read as a bigger or more muffled box, that is the one to audition first — though an error or low-battery cue has a fair reason to be dark.
Both feedback files land under the 1.1 kHz floor you quoted. That may just be a difference in how centroid is computed, so weigh MANIFEST.csv over my method.
Two things that checked out fine: RMS runs −13.1 to −21.8 dBFS, matching your −13 to −22; and every file ramps to zero in its last samples, so no truncation clicks.
(I sell pixel-art asset packs on itch, so weigh that too.)