Prevent cold LocalVQE startup from overflowing capture pipes

AGENTS.md

@@ -388,6 +388,8 @@ Frontend and shell:
   Preserve the 16 kHz 256-sample hop adapter, one-hop delay compensation,
   bounded shared-timeline queue, and Stop/Pause/EOF tail flush. Model construction
   must emit no stdout because that channel carries the sidecar JSON protocol.
+  Initialize the model and connect output consumers before opening upstream
+  capture pipes; opening those pipes starts bounded producer writes.
   The token-protected Meeting device test must reuse this path, remain explicit
   and local-only, return only bounded level/activity statistics, and always stop
   its ephemeral sidecar capture without persisting or uploading PCM. Product

Frontend/package.json

@@ -1,6 +1,6 @@
 {
   "name": "scriber",
-  "version": "0.5.126",
+  "version": "0.5.127",
   "type": "module",
   "license": "MIT",
   "engines": {

Frontend/src-tauri/src/audio_sidecar.rs

@@ -3754,6 +3754,20 @@ fn run_meeting_aec_relay(
         .map(|handle| *handle as HANDLE)
         .collect();
     let result = (|| -> Result<(), String> {
+        // Connecting upstream pipes starts bounded capture writes. Prepare the
+        // model and all consumers first so a cold model cannot fill those pipes.
+        let mut aec = if aec_enabled {
+            // Force the real pipe regression beyond its buffer/write deadline
+            // in tests; production performs no artificial wait.
+            #[cfg(test)]
+            thread::sleep(Duration::from_millis(1500));
+            Some(MeetingEnhancer::new()?)
+        } else {
+            None
+        };
+        for handle in &outputs {
+            wait_for_meeting_output_pipe_client(*handle, &stop_rx)?;
+        }
         let microphone = open_frame_pipe_reader(&microphone_pipe)?;
         let system = match open_frame_pipe_reader(&system_pipe) {
             Ok(handle) => handle,
@@ -3765,14 +3779,6 @@ fn run_meeting_aec_relay(
             }
         };
         let processing = (|| -> Result<(), String> {
-            for handle in &outputs {
-                wait_for_meeting_output_pipe_client(*handle, &stop_rx)?;
-            }
-            let mut aec = if aec_enabled {
-                Some(MeetingEnhancer::new()?)
-            } else {
-                None
-            };
             let mut relay_sequence = 0u64;
             let mut microphone_frame = Some(read_meeting_frame(microphone, &stop_rx)?);
             let mut system_frame = Some(read_meeting_frame(system, &stop_rx)?);
@@ -6732,10 +6738,13 @@ mod tests {
                 })
             })
             .collect();
+        let capture_id = response["payload"]["meetingCaptureId"].as_str().unwrap();
         for _ in 0..3 {
-            ready_rx.recv_timeout(Duration::from_secs(10)).unwrap();
+            if let Err(error) = ready_rx.recv_timeout(Duration::from_secs(10)) {
+                let stopped = state.stop_meeting_capture(capture_id, "testFailureCleanup");
+                panic!("Meeting frames not ready: {error}; stopped={stopped}");
+            }
         }
-        let capture_id = response["payload"]["meetingCaptureId"].as_str().unwrap();
         let stopped = state.stop_meeting_capture(capture_id, "meetingCaptureStop");
         assert_eq!(stopped["stopped"], true);
         assert_eq!(stopped["relay"]["relayError"], Value::Null, "{stopped}");

docs/ARCHITECTURE.md

@@ -218,6 +218,8 @@ Meetings:
    loopback sources at 48 kHz in shared mode; it never opens a camera/video
    device. The microphone and loopback frames are downsampled to 16 kHz before
    the single pinned LocalVQE v1.3 model reduces echo, noise, and reverberation.
+   The relay initializes the model and connects consumers before opening its
+   upstream capture pipes, so cold model startup cannot overflow producer buffers.
    Its 256-sample inference hops are adapted to the 160-sample transport frames;
    the relay holds at most four frames, removes the native one-hop pre-roll,
    and flushes the final partial hop before EOF. All three tracks retain the

native/scriber-audio-sidecar/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "scriber-audio-sidecar"
-version = "0.2.0"
+version = "0.2.1"
 description = "Scriber audio capture worker"
 authors = ["Scriber"]
 edition = "2021"

src/version.py

@@ -3,7 +3,7 @@
 import os
 import re
 
-__version__ = "0.5.126"
+__version__ = "0.5.127"
 
 _SEMVER_RE = re.compile(r"^\d+\.\d+\.\d+(?:[-+][0-9A-Za-z.-]+)?$")
 

tests/test_audio_worker_build_contract.py

@@ -93,5 +93,8 @@ def test_audio_cache_only_validation_does_not_require_a_local_rust_compiler(tmp_
         check=False,
     )
     assert result.returncode != 0
-    assert "cache is missing, stale, or does not match worker version 0.2.0" in result.stderr
+    worker_version = tomllib.loads((ROOT / "native/scriber-audio-sidecar/Cargo.toml").read_text(encoding="utf-8"))[
+        "package"
+    ]["version"]
+    assert f"cache is missing, stale, or does not match worker version {worker_version}" in result.stderr
     assert "rustc" not in result.stderr