[Fix] Voice Sessions invent codebase answers before checking Fast

[roomote-roomote]

​Opened on behalf of @mrubens. Follow up by mentioning @roomote-roomote, in the web UI, or in Slack.

What changed

• Route every completed voice utterance through the Fast Session, including an ordered fallback-only mode after GPT-Live misses a delegation so delayed identifiers cannot attach to later speech.
• Restrict GPT-Live's independent speech to brief acknowledgements and wait updates, then have it report Fast commentary instead of originating answers.
• Apply grounding safeguards to any product, repository, or connected tool discussed in the Session, while preserving Roomote-specific handling when the platform itself is the topic.
• Mark any direct voice-layer transcript output as unverified in later Fast history so it is not reused as established evidence.
• Update voice documentation and add a patch changeset.

Why this change was made

GPT-Live could classify a substantive product discussion as small talk and answer from its own unsupported assumptions before Fast, repository inspection, or tools were involved. That allowed plausible but false claims about the Roomote codebase to enter the Session transcript.

Impact

Voice conversations now consistently hand user utterances to the Session's selected Fast model while preserving existing interruption behavior. Mocked event-ordering regressions cover fallback on utterance A, a delayed delegation before and after utterance B begins, delays beyond the former three-second window, duplicate prevention, and normal delegation after restarting the call. Once a call misses one delegation, later turns use the 1.5-second fallback rather than accepting uncorrelatable delegation IDs. No live GPT-Live call was run. Because the WebRTC audio track still autoplays, prompt-disobedient direct model audio remains a residual risk until output is provenance-gated at the transport layer.

[roomote-community]

No code issues found. See task

• Late GPT-Live delegations can be attached to the next spoken utterance and route the Fast response to the wrong turn. (apps/web/src/hooks/useLiveVoice.ts:248)

_{Reviewed 20a41b6}

[roomote-roomote]

Updated the voice safeguard to cover any product, repository, or connected tool discussed in the Session, while retaining Roomote-specific context when the platform itself is the topic. Focused prompt tests, fast web typechecking, lint, formatting, and pre-push checks pass in f32d8f5. The separate late-delegation turn-attribution review finding remains unresolved.

[roomote-roomote]

Resolved all current actionable feedback in 20a41b6. After a missed delegation, the call now ignores uncorrelatable later delegation IDs and uses ordered 1.5-second fallback delivery until restart. Regressions cover A fallback, B beginning before A’s delayed delegation, delays beyond the former three-second window, duplicate prevention, and reset on a new call. The focused client suites pass 72 tests; full web typechecking and pre-push checks pass. Browser proof is not applicable to event timing. No live GPT-Live call was run, and fallback-only mode adds up to 1.5 seconds to later turns in a degraded call.