Updates

docs/qvac-runtime.md

@@ -63,6 +63,10 @@ cache directory, never under a system-wide writable location.
 ## GPU policy
 
 `JARVIS_GPU_REQUIRED=1` is set by the service. The master requests GPU device
-and maximum GPU layer offload, then rejects a result QVAC reports as CPU. Run
+and maximum GPU layer offload. An explicit GPU request (or
+`JARVIS_GPU_REQUIRED=1`) fails directly on a GPU load error rather than loading
+a CPU copy that the master would reject. Optional `device: 'auto'` callers outside
+that policy can still fall back, and runtime status labels that backend as CPU.
+Run
 `npm run gpu-doctor` and consult [hardware compatibility](hardware-compatibility.md)
 before changing a model profile.

docs/voice-pipeline.md

@@ -10,9 +10,10 @@ flowchart LR
   R --> W[Wake engine]
   W -->|phrase| V[VAD + QVAC ASR]
   V -->|final utterance| H[Harness bridge]
-  H -->|streamed reply| T[Sentence buffer]
+  H -->|final reply| T[Sentence queue]
   T --> Q[QVAC TTS]
   Q --> OUT[PipeWire playback]
+  OUT -. prepare next sentence concurrently .-> Q
   OUT -. anti-feedback gate .-> V
 ```
 
@@ -44,7 +45,30 @@ very short fragments are discarded.
 
 Run `npm run voice-doctor`, then follow [voice acceptance](voice-acceptance.md).
 The daemon metrics cover wake accepts/rejects, feedback drops, utterances, and
-replies. Raw audio is not persisted by default.
+replies. Speech metrics also report `synthesisCount`, `synthesisMs`, `audioMs`,
+and `synthesisRealtimeFactor` (total synthesis time / generated audio duration;
+less than 1 means synthesis is faster than playback). Raw audio is not persisted
+by default.
+
+## Inference and playback concurrency
+
+Whisper explicitly requests `contextParams.use_gpu: true`; it otherwise defaults
+to CPU independently of the LLM's GPU policy. TTS honors the saved `ttsUseGpu`
+setting. These engines use their own configuration fields, not llama.cpp's
+`gpu_layers`. Concurrent requests to load the same auxiliary model share the
+in-flight load, avoiding duplicate allocations. The shared model stays loaded
+until its last auxiliary user releases it.
+
+Speech synthesis prepares one sentence ahead while PipeWire plays the current
+sentence. TTS requests and playback remain ordered, with at most one prefetched
+sentence. The feedback gate stays active across sentence boundaries. Interrupts
+discard prefetched audio; shutdown drains any in-flight synthesis before unloading
+models. Synthesis and playback failures release the gate and allow the next reply.
+
+Independent harness tool calls already run concurrently. Conversation inference
+stays serialized to preserve history ordering and the single active chat model.
+Increasing llama.cpp `parallel` also divides its configured context across slots;
+it is not a free speed increase for a single conversational stream.
 
 ## Release voice readiness