Sound Pass — the category mixer + five audible improvements

What this pass did

Two deliverables. One: a per-category sound mixer — combat, ambience, radio, and alerts each get an on/off toggle and a volume slider, persisted per device. Two: five improvements to the audio itself, each grounded in either a recorded residual of the 2026-06-07 soundscape campaign or a missing battlefield signature — and each verified by the offline oracle as a number, not a vibe.

1 · The sound mixer

The engine already had four submix buses; the mixer inserts a user gain between each bus and master (bus → category → master, lib/audio/player.ts). That placement is the design decision: the contact-duck system writes absolute values to the bus gains, so putting the user's trim on a separate downstream node means ducking and user preference multiply instead of fighting. Two consequences worth knowing: the shared valley-reverb return is routed into the combat category (only combat sounds feed it — muting combat must also mute its echo, not leave a disembodied tail), and the danger-close klaxon moved from the combat bus to the alerts bus where it semantically belongs (it is a command-channel warning, not battlefield sound) with trims re-balanced so its absolute level is unchanged.

Live-verified end to end: UI toggle → store → engine gain (ambience 0.000, combat 0.500) → localStorage itm-ui-v1 → survives reload. The oracle re-render after the re-routing showed all 24 scenes within 0.15 dB of baseline — the mixer's plumbing is level-neutral by construction.

2 · The five audio improvements

2.1 · The incoming-shell whistle (new battlefield signature)

Why: indirect fire had a tube report ("shot") and an impact ("splash") but nothing in between — yet the 2–3 seconds of descending shriek before a mortar round lands is the single most recognizable indirect-fire experience in every first-hand account. The sim already counts down FireMission.etaS each tick, so the mapper now voices one whistle per mission, deterministically, 2.4 s before the first round (INCOMING_LEAD_S, lib/audio/mapper.ts) — the whistle ends exactly as the blast arrives. The synth is a triangle sweep (~1.3 kHz → 320 Hz) with a flutter that deepens as the round closes, plus a swept-noise whoosh an octave up, under a swelling envelope: the swell is the dread.

2.2 · Calm-bed stereo width (the recorded residual)

Why: the 2026-06-07 campaign shipped the living valley bed but recorded its own gap: the calm scene measured corr 0.997 / width 4% — effectively mono. The fix was less about panning and more about a subtlety of synthesis: all bed voices loop one shared noise buffer, and two in-phase filters of the same source stay correlated no matter how you pan them. Each voice now starts at a different offset into the loop (noise autocorrelation dies in samples, so offset reads are genuinely uncorrelated), and then the panning does real work: the river splits into two band voices (950 / 1900 Hz, the bird-band notch preserved) straddling the river's actual screen bearing, the generator pans to the COP's bearing, rain is Haas-widened (13 ms), and the wind howl spreads ±0.5.

2.3 · Thunder (new — the storm finally sounds like one)

Why: Rain weather had a hiss bed and droplet ticks but no thunder — a Kunar mountain storm without its defining sound. A new Poisson spot layer (λ = 0.022/s, Rain only) builds each roll as a mid-frequency strike head that darkens into a brown-noise rumble with 2–3 swells, answered ~0.6 s later by a quieter roll panned to the opposite quarter — the sound crossing the valley. Deliberately not gated by contact: a firefight silences the birds, never the sky.

2.4 · The adhan gets a voice

Why: the call to prayer is formant-filtered sawtooth with a programmed melisma — but a held note with mathematically flat pitch reads as a slide whistle, not a human voice. A 5.3 Hz vibrato (depth 2.5% of the base pitch) fading in over the first second is the classic synthesis tell of a singing voice, and it does more for realism than any filter change.

2.5 · Ricochet variety

Why: every deflection played the identical zing — the most repetitive sound in a long firefight. Three deterministic timbre families now key off the cue's hash (cue.v): the classic singing zing (~45% of rounds), a short hard fast-tumbling buzz (~30%), and a rarer high thin whine with a long fall (~25%) — plus an impact tick grounding each one in a hit. Same mechanism, three parameterizations; same seed still gives the same firefight.

3 · The numbers, all in one table

Oracle assertions after this pass — all green: no clipping (busiest scene −2.8 dBFS) · calm→combat dynamic spread 25.8 dB (was 23.0) · ambient bed alive (−45.5 dB) · firefight corr 0.48 · occlusion 6.6 dB / HF 30.8% vs 40.9% · calm width 12.4% (was 4%) · storm alive · whistle sustains 2.26 s. Plus scripts/audio-probe.ts green (mapper purity, 1:1 cue mapping, determinism, layer law) — incoming now part of its coverage set.

4 · Honest residuals

5 · How to verify this yourself

• npx tsx scripts/audio-render.ts /tmp/check — re-renders all 27 scenes, prints the table + assertions (expect RENDER OK).
• npx tsx scripts/audio-probe.ts — mapper purity/determinism/coverage (expect AUDIO OK).
• In game: the 🎚 button in the command bar's audio cluster opens the mixer; prefs persist per device.