fix(typescript-sdk/voice): harden ElevenLabs ConvAI post-interrupt / next-turn receive

[drewdrewthis]

Follow-up to #561 (epic #370).

The ElevenLabs hosted ConvAI adapter (adapters/elevenlabs.ts) times out on receiveAudio for the post-interrupt and scripted next-turn cases — the hosted server-VAD does not re-engage for scripted (non-mic) audio. Consequences in #561:

• elevenlabs_interruption demo is gated off (RUN_EL_INTERRUPTION=1 to attempt) — 4 honest attempts timed out on the post-interrupt receive. Not faked.
• elevenlabs_hosted demo is limited to one live exchange (a scripted 2nd user turn times out the next agent receive). EL multi-turn is covered instead by the in-process elevenlabs_branded (composable) demo.

Scope

• Harden the EL ConvAI receive path so a scripted user turn after an agent turn (and after an interrupt) reliably re-engages a response — e.g. an explicit response trigger / commit, or a documented silence() nudge — without relying on mic-style server VAD.
• Then un-gate elevenlabs_interruption and extend elevenlabs_hosted to ≥2 exchanges.

References

• PR feat(typescript-sdk): voice agent testing — consolidated clean stack #561 (consolidated TS voice stack), EDR docs(#372): voice internal design record + ADR-002 (per-run provider state) #560, epic Voice Agents #370.

[drewdrewthis]

Fix pushed on branch `fix/567-el-convai-next-turn` (commit b2a01a1), branched off #561's head.

Root cause (confirmed live): EL ConvAI 2.0 uses a hybrid VAD + DL turn-detector, not a silence threshold — so the fixed ~333ms zero-byte tail doesn't deterministically end a scripted (non-mic) turn. Greeting→turn-1 worked; scripted turn-2+ intermittently never re-engaged → `receiveAudio` timed out. (This is the same blocker that made a recent TS voice demo attempt abandon hosted EL ConvAI and pivot to OpenAI Realtime — it's fixable, not an EL-side dead end.)

Fix: verified against both official EL SDKs that ConvAI exposes no audio-flush/commit event; `user_message` (`{"type":"user_message","text":...}`) is the only client event that deterministically forces an agent response without mic-VAD. Added `turnCommitMode: "text" | "silence"` (default `text`) + configurable `silenceTailBytes`. Text mode sends only the `user_message` commit (audio still recorded locally; EL echoes `user_transcript` so observability is intact). Legacy silence path preserved as fallback.

Verified BOTH levels:

• Unit (`elevenlabs-turn-commit.test.ts`, fake-socket): 7 pass; full voice suite 179 pass.
• Live, real EL socket, ≥2 exchanges, 3 consecutive clean runs no retries (greeting→user→agent→user→agent→judge, 5 segments / 14.24s). The previously-timing-out 2nd turn now answers. (First attempt streaming audio+text together flaked 2/3 — drove the text-only revision; honest negative on record.)

Python parity: `python/.../elevenlabs.py` has the identical bug (same 16000-byte tail, no commit) — NOT fixed here (can't live-verify Python in this worktree; shipping unverified would be the false-green this issue warns against). Follow-up: port `user_message` commit + live-verify `python/examples/voice/elevenlabs_hosted.py`.

Not done (separate asks in this issue): un-gating `elevenlabs_interruption` — post-interrupt path is covered by the unit test and the same commit should fix it, but I didn't live-run it. Worth a follow-up.

[drewdrewthis]

PR #596 opened with the fix (stacked on #561 voice/372-refactor, since the voice stack isn't on main yet). Live-verified ≥2 exchanges, 3 consecutive clean runs.