diff --git a/docs/architecture/sdd.md b/docs/architecture/sdd.md
index b47ae90..022cbc3 100644
--- a/docs/architecture/sdd.md
+++ b/docs/architecture/sdd.md
@@ -63,9 +63,12 @@ Design rules:
 | Codec | Opus encode/decode lives in `opus_voice.rs` and associated audio modules |
 | DSP chain | High-pass filter, noise suppression, echo cancellation, and AGC are represented by audio processing modules/backends |
 | Transmit control | `TransmitMode` supports `Ptt`, `Continuous`, and reserved `VoiceActivity`; `VoiceActivity` has no active MVP implementation |
+| VoiceActivity gate (capture-side) | `voice_activity::VoiceActivityStateMachine` is the 10 ms-cadence gate for `TransmitMode::VoiceActivity`; open-after 40 ms (debounce), hangover 500 ms (anti-chatter), min-tx 200 ms (anti-flicker), weak-hold 30-100 frames (anti-stale-VAD); live `configure()` re-clamps existing timers on settings change without resetting state; 9 unit tests cover the main paths |
 | PTT | Desktop/mobile backends expose capability level and active backend; missed-key-up watchdog prevents stuck transmit |
 | Release tail | Tail handling prevents abrupt cutoffs after PTT release where configured |
 | Render peak limiter | `voice_render::limit_peak_inplace` is a single-pass, allocation-free per-frame peak scaler applied in both the macOS and iOS render callbacks before the i16 downmix; default threshold 0.99 prevents hard clipping on multi-client mixes that sum past 0 dBFS while remaining transparent for normal voice levels (allocation-free, lock-free, safe on the realtime audio thread) |
+| macOS render cadence (producer + ring) | `ios_voice_unit.rs:851-961` runs a 20 ms tokio producer task that calls `AudioHandler::fill_buffer(1920)` and `force_push`es each sample into a `crossbeam ArrayQueue<f32>` (SPSC-effective, MPMC-but-wait-free-per-end); ring capacity 12000 samples ≈ 6.25× pull quantum; 100 ms prebuffer (`PREBUFFER_SAMPLES = 9600` stereo f32) before the VPIO render callback starts draining, matching Mumble's playout margin and WebRTC's kStartDelayMs order of magnitude |
+| iOS render cadence (direct-fill) | `ios_voice_unit.rs:968-1034` does `AudioHandler::fill_buffer` directly in the VPIO render callback (VPIO on iOS requests 480-frame ≈ 10 ms slices that align with tsclientlib's 20 ms Opus frame); scratch buffer preallocated to 4096×2 f32 at setup time so the realtime callback never `resize()`s; `try_lock` (not `lock`) on the AudioHandler mutex so contention never stalls the realtime IO thread; on `WouldBlock` the callback emits silence and increments `callback_xrun` |
 | VPIO ducking config (macOS 14+) | `ios_voice_unit.rs` writes an 8-byte `AuVoiceIoOtherAudioDuckingConfiguration` struct (`m_enable_advanced_ducking = 0` disables dynamic voice-activity-driven ducking; `m_ducking_level = kAUVoiceIOOtherAudioDuckingLevelMin = 10`) to selector `kAUVoiceIOProperty_OtherAudioDuckingConfiguration` (= 2108) on the VoiceProcessingIO AudioUnit at startup, minimising the ducking of other apps' audio during a voice session; on macOS 13 the property is silently ignored (VPIO returns the default ducking behaviour) and the code logs a debug message and continues |
 | Benchmarks | Realtime capture, Opus, and resampler benchmarks provide advisory baseline evidence |