The record is Camille Klein's whistlegraph "loner" regulated onto a 122 floor: her line sung whole six times, the band telling the passes apart. THE OPENING /s/ went missing twice, for two different reasons, and both were structural rather than a matter of level. voice_line places a phrase at at(bar) − leadIn so the VOWEL lands on the beat and the consonant runs 1:1 before it — at bar 0 that is negative time, and emit() drops anything before sample 0, so the 220 ms sibilant was never written to the file. And WORLD systematically under-renders fricatives: the voiced half gets the formant lift, the air shelf and the breath, the unvoiced half is copied from the warped source and gets none of them, widening her own 15 dB vowel-to-/s/ gap to 19. halo3 now finds frames that are unvoiced AND bright — the same dim-but-bright test audit.py uses — and gives them 8 dB back. The record opens on the /s/ at 6 ms: PREROLL is exactly the phrase's own leadIn, so there is no silence and nothing is trimmed. A CHART PER TAKE (bin/takechart.py). The chart was two kinds of number tangled together: the composition (beats, durs, and the melody) and f-'s performance (13 hand-pinned onsets, 5 syllable seams, end — all seconds into one wav). singdub applied both, and never read lead/ants at all, so every other take arrived wearing one singer's phrasing with no consonant runway. Takes are now assembled from their per-word corpus files, which makes the 18 onsets exact, and only the seams need finding — in that take's own voice. Each then runs the whole halo3 pipeline. Three bugs found along the way, all real beyond this lane: · FLOOR was 140 Hz, set for her 283 Hz take. Ten of nineteen takes sing under it, and harvest below its floor returns the second harmonic, not a low note. The bracket is now measured per voice — and only ever widens, so the spine's 140/600 is untouched. · THE HOLD read the source analysis to pick its sustained tone, which silently undid the melody lock on every long note. "of" — the one octave leap in the lyric — came out 16 semitones under the chart. · singdub mapped words to units by index. 18 words against 24 units meant six words sung TWICE on adjacent notes. THE ARRANGEMENT is one voice, no background vocals, and a kit of kick, one offbeat hat and a clap. Swing at 0.575 with correlated per-bar lean rather than per-hit jitter; ghost snares brushed low; 3-against-4 and 5-against-4 running quietly underneath. A hook lifted from her first four notes (7 5 3 2) is stated at the head of each pass and answered by its inversion four bars later, handed to a different instrument each time. Claude-Session: https://claude.ai/code/session_01C7GM9itHCF6nXfnc62Dkdy