fix(voice): keep a stopped round's actions and follow-up memory

A reply stopped partway through a round now keeps the actions that ran,
answering the calls it never reached as skipped, so the history stays
paired for both providers. A failed typed turn no longer resets the
pointer, referents and follow-up memory, since its conversation continues.

The server now tells a rejected key (401, or Gemini's 400 with reason
API_KEY_INVALID, matched exactly), denied access (403) and an account out
of credit (OpenAI's 429 insufficient_quota) apart; each cools down for a
minute and gets its own message in the chat.

SECURITY.md

@@ -85,7 +85,7 @@ The data proxies under `server/providers/` are written so the browser cannot tur
 
 Allowing `maps.googleapis.com` in `script-src` also trusts its JSONP endpoints, which can bypass the script allowlist if an HTML-injection bug exists; this is an accepted limitation of loading the Maps JavaScript API, alongside the existing `'unsafe-eval'` allowance.
 - **Full Content-Security-Policy.** The dev and preview servers send a real CSP (`script-src 'self' 'unsafe-eval' blob:` plus the YouTube, Facebook and X scripts the Nepal event's embedded media loads, with a matching `frame-src` for their players and posts: no inline script and no other script origin; `blob:` covers the object URLs Cesium's bundled workers bootstrap through in the built app), alongside the existing `X-Frame-Options: DENY` and `frame-ancestors 'none'`. `'unsafe-eval'` stays because Knockout, bundled inside Cesium's widgets package, resolves the global object with `eval` at load time and the globe does not initialize without it; styles permit `'unsafe-inline'` and Google Fonts because Vite's dev client injects `<style>` elements.
-- **Typed chat spends the provider keys.** `/api/openai/chat` and `/api/gemini/chat` send the conversation to OpenAI Chat Completions or Gemini `generateContent` with the server's `OPENAI_API_KEY` or `GEMINI_API_KEY`. They require a same-site `Origin` (the Gemini request check), cap request bodies at 512 KB, refuse upstream redirects, cut the upstream call off after 60 s and return only sanitized errors. OpenAI chat can only bill `OPENAI_API_KEY`, so while `GEV_PREFER_CODEX_OAUTH` pins OpenAI to ChatGPT OAuth it refuses with `503` (`CODEX_OAUTH_REQUIRED`) rather than spend the key. After an upstream `401` or `403`, a route stops sending that key upstream for a minute, or until the key changes.
+- **Typed chat spends the provider keys.** `/api/openai/chat` and `/api/gemini/chat` send the conversation to OpenAI Chat Completions or Gemini `generateContent` with the server's `OPENAI_API_KEY` or `GEMINI_API_KEY`. They require a same-site `Origin` (the Gemini request check), cap request bodies at 512 KB, refuse upstream redirects, cut the upstream call off after 60 s and return only sanitized errors. OpenAI chat can only bill `OPENAI_API_KEY`, so while `GEV_PREFER_CODEX_OAUTH` pins OpenAI to ChatGPT OAuth it refuses with `503` (`CODEX_OAUTH_REQUIRED`) rather than spend the key. After the provider refuses the key or account (a `401`, a `403`, Gemini's `400` `API_KEY_INVALID` or OpenAI's `429` `insufficient_quota`), a route stops sending that key upstream for a minute, or until the key changes.
 - **ChatGPT OAuth voice reuses the local Codex sign-in.** `/api/realtime/oauth-status`, `/api/realtime/oauth-login` and `/api/realtime/token?auth=oauth` answer only loopback connections, behind the same cross-site gate as the token endpoint; the login route also refuses a foreign `Origin`. None of them returns the access token. The login route runs `codex login` (or the executable named by `CODEX_BIN`) with an allowlisted environment that leaves out GEV's provider keys, keeps one attempt pending at a time and cancels it after two minutes or when the server stops. **Limits:** any program on this machine that can reach the server can mint Realtime sessions on the signed-in ChatGPT account, as it can with the API key; a local reverse proxy that hides the client's address defeats the loopback check; and OpenAI has not documented this use of Codex credentials for Realtime, so it may stop working or fall outside the account's terms.
 
 ## Network exposure — the operator threat model