feat(voice): migrate to push2talk and serialize websocket text sends

README.md

@@ -7,12 +7,12 @@ ESP32-S3 works as a Wi-Fi REST controller and replaces Android App logic:
 ## Features
 
 - Wi-Fi STA-only mode (configured SSID/password; no SoftAP fallback)
-- Built-in minimal web UI at `/` for voice tap2talk
+- Built-in minimal web UI at `/` for voice status and MIC_KEY guidance
 - Direct thermal printer control with sensor precheck
 - Async print queue with jobs (`queued/running/success/failed/canceled`)
 - Printer precheck (paper / battery / temperature)
 - Built-in UTF-8 text rendering with Chinese support (GB2312 character set, 16x16 bitmap)
-- Tap2talk multimodal voice interaction over DashScope WebSocket:
+- Push2talk multimodal voice interaction over DashScope WebSocket:
   - ES8311 microphone/speaker via `esp_codec_dev`
   - raw-opus uplink/downlink via `esp_audio_codec`
   - WebSocket transport via `esp_websocket_client`
@@ -125,7 +125,7 @@ http://<esp-ip>/
 ```
 
 The page provides a minimal voice console:
-- single-round tap2talk button
+- MIC_KEY(IO14) trigger guide (press to talk, release to process)
 - voice status display
 
 ## REST Examples
@@ -205,20 +205,20 @@ curl http://<esp-ip>/v1/jobs
 curl http://<esp-ip>/v1/jobs/1
 ```
 
-### Voice tap2talk
+### Voice push2talk
 Start session:
 
 ```bash
 curl -X POST http://<esp-ip>/v1/voice/session/start
 ```
 
-Start one tap2talk round:
+Press-to-talk start (sends `SendSpeech` when entering `Listening`):
 
 ```bash
 curl -X POST http://<esp-ip>/v1/voice/tap/start
 ```
 
-Cancel current tap2talk round:
+Release-to-stop (sends `StopSpeech`):
 
 ```bash
 curl -X POST http://<esp-ip>/v1/voice/tap/cancel
@@ -244,7 +244,8 @@ curl -X POST http://<esp-ip>/v1/voice/session/stop
 - Characters outside embedded glyph set are rendered as square fallback boxes.
 - Large buffers for image upload/decode prefer PSRAM first, then fallback to internal RAM.
 - Partition table uses `partitions.csv` with a 4MB `factory` app partition on 16MB flash modules.
-- Voice flow is now single-round: audio uplink stops after `SpeechEnded`, then `LocalRespondingEnded` and `Stop` are sent after local playback drain.
+- Voice flow uses `push2talk`: client sends `SendSpeech` to start uplink and `StopSpeech` on release to end input.
+- Voice flow is now single-round: after recognition/response completes, `LocalRespondingEnded` and `Stop` are sent after local playback drain.
 - Voice flow blocks starting the next round while voice is Thinking/Responding, marker image generation is running, or printer queue is still busy.
 
 ## Font Assets