feat(engines): GeminiLive/GeminiCascade/InworldRealtime markers + Gemini Live fixes (release 0.7.1)

[nicolotognoni]

Release 0.7.1 — Gemini/Inworld engine markers + Gemini Live fixes

Follow-up to 0.7.0. 0.7.0 shipped the pipeline barge-in fix (#200) but did NOT ship the GeminiLive engine marker (so new GeminiLive({...}) / the engine import threw at runtime — the demo had to vendor an old .tgz), and a live-call Gemini "thinking" leak surfaced. This lands the markers + fixes and prepares 0.7.1. npm publish is intentionally NOT run — it's human-gated (publish happens via the v0.7.1 tag after merge).

Engines (new public API)

• GeminiLive engine marker + richer GeminiLiveAdapter (mulaw8↔PCM telephony transcode, apiVersion auto-detect, GEMINI_LIVE_3_1_FLASH_PREVIEW), GeminiCascade engine, GeminiSTT/GeminiTTS — integrated from the gemini-3.1-live-demo work onto the 0.7.0 baseline (3-way merge; fix(pipeline): stop false barge-in from self-interrupting the user-turn LLM (agent goes mute) #200 barge-in + feat(inworld): complete TTS pricing + config helper + Realtime Router LLM preset #199 Inworld pricing/exports preserved, verified).
• InworldRealtime — native engine marker + InworldRealtimeAdapter that subclasses OpenAIRealtimeAdapter and overrides only the transport (wss://api.inworld.ai/v1/realtime, Bearer/JWT — Inworld's OpenAI-Realtime migration path). Lets the demo drop its OpenAIRealtime monkey-patch.
• Verified at runtime from the built .tgz: require('getpatter') and ESM named imports both resolve GeminiLive, GeminiCascade, InworldRealtime as functions (the exact 0.7.0 bug, now fixed).

Fixes (live-call bugs on the Giulia demo vs 0.7.0)

• Gemini Live spoke the model's chain-of-thought before the reply (every turn, native-audio). The Live setup now always sends thinkingConfig (voice default thinkingBudget: 0, OFF; opt-in thinking/thinkingBudget on GeminiLiveOptions), and the adapter defensively drops any thought===true part from both the audio stream and the text transcript (never logged — no PII).
• gemini-3.1-flash-live-preview → silent "session ready timeout" (call dropped with dead air). It's a native-audio model served only on v1alpha but its id lacks the native-audio token, so the heuristic chose v1beta. New geminiRequiresV1Alpha() selects v1alpha for native-audio AND flash-live-preview (explicit apiVersion overrides); connect failures now throw a clear, actionable error (model id + resolved apiVersion + root causes) instead of dead air.

Changed

• @google/genai peerDep ^0.3.0 → >=2.0.0 (kept optional) for the 2.x unified SDK — npm install getpatter with @google/genai@^2.x no longer needs --legacy-peer-deps.
• create-getpatter launcher realigned to the SDK version (lockstep) — it pins the getpatter version it bootstraps to its own version. It had drifted to 0.6.9 (so npm create getpatter was provisioning 0.6.9); bumped to 0.7.1.

Backward compatibility

OpenAIRealtime / OpenAIRealtime2 / Ultravox / ConvAI unchanged. The instanceof OpenAIRealtimeAdapter gates are extended, not replaced (InworldRealtimeAdapter passes them via subclassing; the Gemini function_call dispatch gate is widened). #200 barge-in fix confirmed intact.

Versioning

All four version files move to 0.7.1 together (Python __init__.py + pyproject.toml, TS package.json, launcher create-getpatter/package.json). Per the "0.7.x" ask this is tagged a patch; note it does add new public API (engine markers), which is strictly a minor — call it 0.8.0 instead if you prefer stricter semver.

Test plan

• npm run lint (tsc --noEmit) green; npm test green — 2334 passed, 9 skipped (+23 authentic tests: gemini-live-thinking, inworld-realtime, …; mock only the provider WebSocket).
• Built .tgz (npm pack → getpatter-0.7.1.tgz): GeminiLive/GeminiCascade/InworldRealtime resolve at runtime (CJS + ESM).
• Adversarial verification pass (independent re-build + re-test + checklist).
• Demo (giulia-three-variants) verifies against the local .tgz, deletes Giulia/vendor/getpatter-0.6.7.tgz + the Inworld monkey-patch, reverts the import type { GeminiLive } workaround to a value import.
• Publish is human-gated — merge → tag v0.7.1 → release.yml publishes npm + PyPI + the realigned launcher.

🤖 Generated with Claude Code

https://claude.ai/code/session_01D52f6rBsvzicXg45C6MKVM

---

Also in this PR: media-stream WebSocket authentication (Closes #204)

Fixes a second security report from the same reporter: the media-stream WS endpoints (/ws/stream/… Twilio, /ws/telnyx/stream/…, /ws/plivo/stream/…) were unauthenticated — the signed carrier webhook validated HTTP, but any peer reaching the public host could open the WS with an attacker-chosen call_id, send a start frame, and drive a full STT→LLM→TTS session on the operator's keys (toll fraud) or extract the agent's system prompt.

• Per-call token: the signature-validated webhook mints a high-entropy token (short-TTL, keyed to the call) and delivers it on each carrier's existing custom channel (Twilio <Parameter>→customParameters, Telnyx URL query, Plivo extra_headers). The WS constant-time-validates it before opening any provider session; unauthenticated → WS 1008, no provider connect, no TTS.
• Fail-closed by default (require_stream_auth / requireStreamAuth), opt-out for custom-TwiML operators (logs a warning). Token never logged. Backward compatible for the standard serve() inbound+outbound path.
• Also added a global concurrent-WS backstop for the forgeable-X-Forwarded-For per-IP cap (documented DoS-cap-only).
• Both SDKs, all three carriers. Tests: tests/security/test_stream_auth.py (30) + TS stream-auth suite. Adversarial security-verify pass (re-ran the Security: media-stream WebSocket endpoints are unauthenticated — toll fraud & system-prompt extraction #204 repro — an unauthenticated peer is provably blocked). Python 2982 / TS green.

Closes #204.