formats: a music bank's third sub-wave was its own header

The port hit `sound_bank_riffs("BGM_103.slb")` returning three against a census
that says two, and refused to guess which to drop. It was our reader.

`to_xma_riffs`'s hybrid branch derives a leading packet stream's start as
`first_riff % XMA1_PACKET`. That is right only when the bank header is smaller
than one 2048-byte packet -- true of the voice banks the branch was written for
(1392/1468/1600/1728), false of a music bank, whose header is exactly five
packets. The modulus returned 0 and the whole 10 240-byte header was emitted as
sub-wave 0.

The header states its own length, so the guard needs no threshold: BE u32 0x800
at +0x18 with the bank id repeated at +0x00 and +0x20, header length in blocks at
+0x24. Disc-wide over sound.pak's 9 519 entries, 28 match at offset 0 -- every
music bank, ids 1001-1023 and 1101-1105 -- and on 28/28 the declared header ends
EXACTLY at the first RIFF. Zero have a gap, so a header and a leading packet
stream never coexist here; zero false positives among the other 9 491.

Controlled rather than argued: decoding the emitted region through the same
chain, on the same bank, in the same run gives 0.009 s of PCM where the bank's
real wave 0 gives 87.744 s against a declared 87.75. The region is also 99.1%
zero bytes. And the oracle had already said two -- the XMA probe at the main menu
saw exactly two streams, at BGM_103's two declared wave sizes.

BGM_106-109 are deliberately NOT in the 28: their entries start mid-bank, so they
have no header at offset 0 and their leading region is real audio. The
VOICE_D_453 recovery is untouched and its tests still pass, 10/10 green.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014voBspJ6kFncNErZJuZcLw

docs/re/INDEX.md

@@ -163,5 +163,6 @@ files, which is how the same ground got covered twice.
 | [`weapon-datasheet-runtime.md`](weapon-datasheet-runtime.md) | Weapon DATA SHEET — runtime capture (Route B) |  🟡 first dynamic capture, 2026-07-28. The Arsenal's Gallery Mode panel is a |
 | [`xpr2-colour-check.md`](xpr2-colour-check.md) | XPR2 colours: channel order ✅ confirmed against the running game | — |
 | [`focus-ring-spin-measured.md`](focus-ring-spin-measured.md) | The main menu's focus ring spins continuously — and how fast | ✅ **measured**: period **2.177 s** over 9 revolutions (8 evenly spaced autocorrelation peaks) = 120 units = 60 frames = 2.00 s at 30 Hz. A pulse is excluded — annulus total conserved to 0.4 % while per-bin brightness swings by 24. ✅ the ring is the **only** moving thing on the settled main menu (std exactly 0.000 elsewhere). 🔴 no angle is quoted: the angular estimator FAILED its own control (30° → 0°) |
+| [`structures/slb-bank-header-not-a-wave.md`](structures/slb-bank-header-not-a-wave.md) | Why a music bank read as THREE sub-waves when the census says two | ✅ **decoded**: the third is the **bank header**, emitted by our own reader. `to_xma_riffs`'s hybrid branch derives a leading packet stream's start as `first_riff % 2048`, which is right only for a header shorter than one packet; a music bank's header is exactly **5 packets (10 240 B)**, so the modulus gave 0 and the whole header came back as sub-wave 0. The header states its own length at `+0x24` in blocks. Disc-wide over 9 519 `sound.pak` entries: **28** match the header signature at offset 0 (ids 1001–1023, 1101–1105), **28/28** end exactly at the first `RIFF`, **0** have a gap, **0** false positives — so a header at offset 0 and a leading packet stream never coexist. Decode control, same chain, same bank: the emitted region gives **0.009 s** against **87.744 s** for the real wave 0. Corroborated by the runtime XMA probe, which saw exactly two streams at the main menu. Fixed + 2 regression tests; the `VOICE_D_453` recovery is untouched (10/10 green) |
 | [`title-plate-delay-measured.md`](title-plate-delay-measured.md) | How long the boot title shows build 4 before the `PRESS Ⓐ` plate | ✅ **measured**, two independent boots: **2.138 s** and **2.132 s** from the frame build 4 settles (glyph = its no-plate 154, motion → 0). Agreeing to **6 ms**. So the boot title's end state is **not** plate-free and a compositor must draw **two builds at once**. ⚠️ Measure from *settled*, not from first pixels — "first drawn → plate" is 3.78 s vs 4.26 s across the same two runs, because the build-in animation's own duration varies with emulator frame pacing. Plate pulse re-measured at 2.12/2.19/2.34/2.31 s (mean 2.24), replicating the corpus's ≈2.3 s. ✅ black hold between screens bracketed at **0.14–0.30 s**, consistent with the declared 12 units. 🔴 the Ⓐ→menu latency is still **not** available: both runs freeze one frame for ~1.4 s at surface mean **26.626** — agreeing between runs to 1e-6, and reproduced with stream restarts disabled — which is a guest **load stall**, not the capture path. Probe: 8.7 ms/frame, 7.97/7.98 fps against a requested 8, controls 9/9 + 4/4 |
 | [`menu-idle-and-b-2026-08-29.md`](menu-idle-and-b-2026-08-29.md) | The main menu does not idle back to the title — and four durations that were a pipeline | ✅ **refuted**: no self-return in **≥ 60 s** untouched; the ~8–10 s idle belongs to the **title**. 🟡 Ⓑ→title ordering measured, latency not. 🔴 `classify_array` at **1503 ms/frame** drained an 8 fps stream at 0.64 fps and manufactured four latencies (24.66 s / 15.58 s / 25.60 s / 20.26 s) — all withdrawn; a backlog preserves ordering and destroys durations |