D23: refuse the E-mu whole-extent "no loop" at both ends, not only at frame 0

[bmxcode]

Closes #41.

What this does

sample/emu3.py refuses a loop spanning a sample's whole declared extent, because that span is the E-mu format's "no loop": the sampler fills the four loop pointers with the sample's own bounds when nothing set them, and emitting it writes a smpl chunk telling a DAW to loop the entire file. The guard was a == 0 and b >= extent − FULL_EXTENT_SLACK — slack at the end, an exact zero required at the start.

The discs do not write those bounds at frame 0. They write them inset by a small fixed amount at both ends — loop_start = start + C1, loop_end = end − C2 for a per-disc constant of a handful of bytes: ditto-drums (12, 12), the EIIIX discs (4, 4), esi32-gm/protozoa/eiv-vitous/eiv-analogia (12, 10). So the frame-0 guard never fired, and ten discs shipped a loop over the whole file — 934 of ditto-drums's 948 records most visibly.

This widens the guard to refuse a loop within FULL_EXTENT_SLACK frames of both ends:

if a <= FULL_EXTENT_SLACK and b >= extent - FULL_EXTENT_SLACK:
    continue

What the discs showed

Widening was not obviously right — a sustained organ or string can legitimately loop over nearly its whole length, and a rule that deletes those to catch drum hits would destroy real loop points. It was measured against all ten reference discs before the guard was written.

The shape/join correlation is the wrong instrument here, exactly as the issue warned: a whole-extent loop starts within a few frames of 0 and ends within a few of the last, so there is almost no audio before its start for the windowed correlation to score. Two measurements that do apply separate the populations on every disc:

• End energy — the loop end sits below 15 % of the sample's peak (ends in silence) on 70–100 % of the inset population, where a real loop does so on only 13–33 %.
• Uniqueness — where a record is loud enough at both ends to score, the join splices where a random start against the same end does not (a chosen loop point) on 0–11 % of the inset population, against 33–56 % of the loops this project already ships.

Disc	refused	inset ends quiet	inset uniquely-splices	real-loop control
esi32-gm	428	70 %	11 %	37 %
protozoa	1 762	72 %	10 %	37 %
eiiix-1	472	98 %	1 %	49 %
eiiix-2	372	97 %	2 %	41 %
emu-classics	302	88 %	3 %	36 %
vintage	107	82 %	5 %	46 %
ditto-drums	934	100 %	0 %	43 %
eiv-analogia	443	97 %	1 %	33 %
eiv-studio	337	87 %	4 %	41 %
eiv-vitous	628	100 %	—	56 %

The right-hand column is that same uniqueness measured on the disc's real loops — the calibration — and the inset population sits far below it every time. Structure agrees with content: the inset bounds are the record's own extent inset by a fixed constant (an auto-filled field), where a hand-set loop start is at an arbitrary musical position (44, 136, 11 204, 52 536, never a fixed few).

eiv-vitous and eiv-studio are the check that matters most: ADR-0025 validated their loops by the shape test at +0.68 and +0.86, and those validated loops are the real-loop column — they are kept. What is refused there is a separate whole-extent population that ends in silence. eiv-analogia's loops never had independent evidence (ADR-0025: 34 scored, showed nothing), so refusing 443 of its 449 takes nothing that was ever established.

Effect

Loop counts move on every disc; no audio moves. read_file and the offset arithmetic are untouched, so every per-disc payload digest, and the sample and stereo counts, are unchanged — only the loop-count column moved, and the disc-backed suite pins the new numbers.

disc	loops before → after	disc	loops before → after
esi32-gm	1 778 → 1 350	ditto-drums	948 → 14
protozoa	5 244 → 3 482	eiv-analogia	449 → 6
eiiix-1	1 215 → 743	eiv-studio	2 551 → 2 214
eiiix-2	1 264 → 892	eiv-vitous	826 → 198
emu-classics	1 435 → 1 133	vintage	953 → 846

End-to-end: ditto-drums now writes 948 WAVs of which 14 carry a loop, down from 948.

Anyone who extracted an E-mu disc before this got a smpl chunk looping the whole file on a large fraction of its samples — harmless where a DAW ignores smpl, wrong where it does not. The audio was always right.

Alternatives rejected

• Leave the guard at a == 0 — ships the bug on ten discs.
• A content guard on end-silence — keeps the loud-ended inset loops, but puts an RMS measurement of the audio into the parser, the coupling ADR-0025's "watch for" warns against; and the loud-ended remainder is not clearly real (it splices at random starts nearly as often as at the bound — a sustained tone, not a chosen loop).
• Match each disc's fixed (C1, C2) inset — needs a per-disc constant in the parser and buys nothing the simple both-ends slack does not.
• Widen the end slack instead — misreads the bug; it was the start that was pinned to 0.

What it deliberately does not claim

It catches the whole-extent-with-inset form the ten discs write, not every "no loop" a future generation might encode. And on esi32-gm and protozoa about 10 % of the refused records splice uniquely and could conceivably be a near-whole-extent loop the disc intended; they are refused with the rest because they are structurally the record's own bounds inset by the same fixed constant, and separating them would need the content test this project keeps out of the parser. That cost is recorded rather than hidden (ADR-0030).

🤖 Generated with Claude Code