fix(voice): track OAuth login lifecycle and storage readiness

README.md

@@ -237,6 +237,12 @@ _Why cockpit mode exists: you're riding a real aircraft over real terrain — an
 
 > Voice can use an **OpenAI API key** or, on the same machine, an opt-in signed-in **ChatGPT/Codex OAuth** session. The API-key path remains the default. Selecting ChatGPT OAuth in Provider Settings reuses an existing local sign-in or starts the Codex browser sign-in flow and selects OAuth when it completes. The OpenAI key still drives the **AI HUD summary**: a terse, five-word intelligence-style readout of the current view that regenerates as you move.
 
+The OAuth option is experimental. OpenAI documents ChatGPT plan usage for the [Responses API](https://developers.openai.com/siwc/token-sharing-open-source/models-and-inference); a supported third-party Realtime contract for reusing Codex tokens has not been confirmed. A successful token mint does not establish billing or support terms.
+
+GEV reads Codex's file credentials but does not write or refresh them. It checks token shape and expiry, and reports expired sign-in as requiring re-authentication. The login button runs `codex login`, permits one pending attempt, and reports exit failures or a two-minute timeout. `CODEX_HOME` must match the directory that Codex uses. If `CODEX_AUTH_JSON` points elsewhere, sign in separately or fix that mismatch before using the button. Codex credential storage must use `auth.json`; Keychain-only and in-memory storage are not readable by GEV. See [Codex credential storage](https://learn.chatgpt.com/docs/auth#credential-storage).
+
+For a live voice check, run `node scripts/qa-voice-auth.mjs http://127.0.0.1:4189 oauth microphone.wav` against your local dev server. Use a WAV with three seconds of leading silence followed by a short spoken reply request. Replace `oauth` with `api-key` to check the API-key path. The check verifies audio input, a completed spoken reply, usage, resource cleanup, and reconnect. It saves screenshots and credential-free results under `.gev-logs/voice-auth/`.
+
 Click **GEV MIC**, grant the microphone, and just talk. This is more than a voice-controlled remote:
 
 - **🧠 It knows what it's looking at.** The agent pulls live scene context before answering — including coordinates, street names, active layers, and view scale. Ask _"what city is this?"_ mid-flight and it knows.