Record measured end-to-end turn latency

Full speech-to-speech verified on live 0G mainnet: microphone audio through
0G whisper-large-v3, 0G glm-5.2 and local synthesis.

A turn takes about thirteen seconds from end of speech to first audio, roughly
5 s for transcription and 8 s for the reply, which corrects the seven seconds
recorded earlier from the chat call alone. The two calls are inherently serial.

README.md

@@ -134,9 +134,12 @@ set off by a television in the background. Reading one REST value cannot.
 
 ## Known characteristics
 
-- **A turn takes roughly seven seconds before Ante starts speaking.** This is the
-  0G providers' fixed round-trip, not generation speed: measured time-to-first-token and
-  total time are within 0.1 s of each other, so the response arrives in one burst.
+- **A turn takes about thirteen seconds end to end**, measured from the moment you stop
+  speaking to the moment Ante starts. It splits roughly 5 s for transcription and 8 s for
+  the reply. This is the providers' fixed round-trip rather than generation speed:
+  time-to-first-token and total time are within 0.1 s of each other, so a response
+  arrives in one burst and streaming buys nothing. The two calls are inherently serial,
+  since there is nothing to think about until the audio has been transcribed.
   Ante's personality is written for short answers partly for this reason.
 - **`glm-5.2` is a reasoning model.** Ante sends `enable_thinking: false`, without which
   roughly a hundred tokens and several seconds go to hidden reasoning every turn.

plan.md

@@ -219,7 +219,28 @@ The installed `reachy-mini` 1.9.0 scaffolder has drifted from the current templa
 
 These are worth reporting upstream; item 1 makes the documented command unusable.
 
-## Outstanding: funding speech-to-text
+## Closed: full speech-to-speech verified on live 0G mainnet
+
+Verified 2026-08-24 once the Whisper sub-account was funded. Two complete turns,
+microphone audio in and synthesized speech out, every stage on the live network:
+
+| Heard (0G whisper-large-v3) | Replied (0G glm-5.2) |
+|---|---|
+| "Hello robot, what is the capital of France?" | "Paris." |
+| "what are you actually running on, and keep it brief." | "A model called GLM, served over the 0G Compute Network. No single company owns the machine that's thinking right now." |
+
+**Measured turn latency: ~13 s** from end of speech to first audio — about 5 s for
+transcription and 8 s for the reply. The two calls are inherently serial. This is the
+providers' fixed round-trip, not generation speed.
+
+**Funding gotcha found in the process**: a provider requires
+`minimum reserve + unsettled fees`, and the reserve is a locked floor that cannot be
+spent from. Funding a sub-account with exactly 1.0 0G bought exactly one request, then
+refused everything after with `required minimum is 1.003071 0G`. Fund with headroom
+above the reserve. The proxy auto-funds the difference once the ledger can cover it, but
+only picks up a new balance on restart.
+
+## Superseded: funding speech-to-text
 
 Transcription is the one thing not yet exercised end to end. Provider
 `0x36aCff…` requires **1.0 0G locked** in its sub-account and currently holds zero;