Marrow Fen

• Every entry and exit lands on the barline — proven by the mid-bar-flip test.
• Stingers duck the music bus and release automatically.
• The ambience follows menace both ways: the marsh floor rises — and the frogs go quiet. When the fen stops singing, something is close. Measured, not just asserted.

• 100% sampled voices — an oscillator spy wraps createOscillator across full playback and every SFX; it must read zero, and the ambience stems are prefetched so the engine's procedural fallback never fires.
• Form verified on the event stream — per-section voice sets, register medians, borrowed-pitch-class discriminators (A♭/D♭ only in the lull, G♭ pads only in A′, E natural only in the pursuit), and the 5/4 bar period measured from event timestamps.
• The lull's echo and the shadow proven by pairing — every celesta note matched to a music-box note exactly two beats earlier; every choir note matched 3–4 semitones under an english-horn note at the same instant.
• Layers verified — menace 0 vs 1 event capture plus the mid-bar-flip test (entries join at the NEXT barline).
• The frogs-inverse binding measured — ambience-only captures at menace 0 vs 1: high-frequency frog energy must fall while the marsh floor holds.
• SFX variation verified — 20 triggers per effect: every trigger ≥2 voices, unique voice signatures, round-robin never repeats; onset spectra checked against the thin-transient defect class.
• Explicit zero-404 gate — the audit's HTTP server counts misses; a silently-missing sample cannot pass (the defect class that survived two prior cycles' voice-count gates).
• Player checks — draw-cost budget, keyboard operation, visible play state, persisted volume + mute, first-gesture unlock in a strict-autoplay context.

Six audit rounds to the clean pair, honestly split: TWO were piece-side — three SFX effects varied too little (9–13/20 unique voice signatures; fixed with independent per-voice detune jitter on top of the master drift → 18–20/20) and the visualizer cost 17.4 ms a frame (a per-frame scaled halo blit, a full-screen reed overlay and per-frame text → pre-rendered and cropped, 0.41 ms). The other rounds fixed the MEASUREMENT: the ambience-balance capture raced the engine's 1.2 s bed ramp; the frogs check first listened in a band the frog stem doesn't occupy (its energy lives at 800–2500 Hz, measured), then raced the 26 s stem loop's own non-stationarity, then was contaminated by the reverb return, which bypasses the muted music bus — the final gate pauses the transport, drives the score's own exposed binding on an isolated handle, and verifies the layer gains before each full-loop capture. One further finding came from the rubric, not the numbers: the bed sat 17.5 dB under the score — too buried for a piece whose signature is the fen going quiet — and was raised. A late visual polish (a gradient edge vignette replacing hard-seamed rects) then reintroduced the draw-cost defect in a new costume — a fractional globalAlpha on one full-screen blit costs 13–17 ms in the software rasterizer while the same sprite at alpha 1 costs 0.25 ms — fixed by baking the strength into the sprite per menace bucket. The measurement lessons and the alpha trap shipped to platform guidance alongside the piece lessons.