Add Krisp VIVA Turn as a user turn-stop strategy

[weiz9]

Summary

Adds Krisp VIVA Turn (pipecat's KrispVivaTurn, VIVA SDK turn detection v3 / Tt API) as a selectable turn_stop_strategy in the cascade pipeline, alongside the existing LocalSmartTurnAnalyzerV3. Unlike Smart Turn (which analyzes audio in batches when VAD detects a pause), KrispVivaTurn processes audio frame-by-frame in real time, using the VAD's speech flag as an external signal. This PR also bakes the licensed SDK into the Docker image so the strategy can run on Toolkit.

Changes

• turn_config.py — new krisp_viva_turn branch in create_turn_stop_strategy(). Forwards threshold / frame_duration_ms to KrispTurnParams and model_path / api_key / sample_rate to the KrispVivaTurn constructor (all fall back to env vars). The pipecat module is imported lazily because it hard-requires the proprietary krisp_audio SDK — environments without it are unaffected until the strategy is selected.
• config.py + .env.example — document the new strategy and required env vars (KRISP_VIVA_TURN_MODEL_PATH, KRISP_VIVA_API_KEY).
• Dockerfile + vendor/krisp/ + .gitignore — optionally bake the Krisp wheel + turn model into the runtime image (see Deployment below).
• scripts/docker_build.sh — helper to build the linux/amd64 image with git-provenance build args (and optionally push). Usage: scripts/docker_build.sh <image:tag> [--push].
• test_turn_config.py — 4 new tests covering the branch (happy path, case-insensitivity, tuning-param routing, constructor-param routing) via an injected fake krisp module.
• Bumped simulation_version 2.0.1 → 2.0.2 (affects benchmark outputs).

Enabling it locally

Krisp is a licensed SDK, not a PyPI package, so it is intentionally not in pyproject.toml. To run this strategy you provision it at runtime:

1. Download from the Krisp developer portal (developers.krisp.ai): the VIVA UAR Python SDK zip and the Turn-Taking models zip. Unzip both.
2. Install the wheel matching your interpreter's Python version and architecture from the SDK's dist/ folder — e.g. macOS x86_64 / Py3.11:

   uv pip install .../krisp-viva-uar-python-sdk-<ver>/dist/krisp_audio-<ver>-cp311-cp311-macosx_11_0_x86_64.whl

   (Pick arm64 vs x86_64 to match python -c "import platform; print(platform.machine())".)
3. Point env vars at the model + your license key, then select the strategy:

   export KRISP_VIVA_TURN_MODEL_PATH=/path/to/krisp-viva-tt-models/krisp-viva-tp-v3.kef
   export KRISP_VIVA_API_KEY=<license key>
   eva --domain airline --model.llm gpt-5-mini --model.turn-stop-strategy krisp_viva_turn

Optional tuning via EVA_MODEL__TURN_STOP_STRATEGY_PARAMS — threshold (default 0.5) and frame_duration_ms (default 20; one of 10/15/20/30/32).

Network: the SDK validates its license against sdkapi.krisp.ai and reports usage to analytics.krisp.ai. Both must be reachable; a WAF/proxy that blocks them will fail license validation. Eval keys are time-limited and metered per minute. (Locally, this requires CloudVPN.)

Deployment (Toolkit / Docker)

The image is linux/amd64, so it needs the Linux wheel from the same SDK download —
krisp_audio-<ver>-cp311-cp311-linux_x86_64.whl (ships alongside the macOS wheels).

The Dockerfile runtime stage optionally installs the SDK: it copies from vendor/krisp/
and, when a wheel is present, installs it into the venv (uv pip install --python /opt/venv/bin/python, matching the builder stage — a bare pip would miss the uv-created
venv) and copies the .kef model to /opt/krisp/models/, setting KRISP_VIVA_TURN_MODEL_PATH.
If vendor/krisp/ has no wheel, the build still succeeds and krisp_viva_turn is simply
unavailable — so existing image builds are unaffected.

• The wheel + .kef are proprietary (not public, not on PyPI) so they are git-ignored;
  only vendor/krisp/README.md (setup instructions) is committed. Populate vendor/krisp/
  from your Krisp account before building.
• KRISP_VIVA_API_KEY is not baked in — supply it as a runtime env var/secret in the job.
• Runtime egress to sdkapi.krisp.ai + analytics.krisp.ai is required (confirmed reachable
  from Toolkit — see Testing).

Testing

• 27/27 test_turn_config.py unit tests pass; ruff + version-bump checks pass; no new mypy errors.
• Live end-to-end validated locally (macOS, cascade pipeline): a full multi-turn airline
  conversation ran with turn_stop_strategy=krisp_viva_turn; the SDK initialized, loaded the
  model, and produced 90 real end-of-turn predictions (probabilities 51–99%, e2e latencies
  ~0.1 ms–1.5 s, mostly ~100–300 ms). Turn-taking was driven entirely by Krisp.
• Validated on Toolkit with the baked image (flux-general-en + gpt-4.1 + sonic-3.5):
  Krisp turn detection ran correctly ([KrispVivaTurn] TURN: COMPLETE, e.g. probability 88–96%,
  e2e ~99–282 ms) with no license errors — confirming Toolkit's egress reaches Krisp's
  licensing endpoints. Also verified krisp_audio imports from /opt/venv in the built image.
• Confirmed the lazy-import boundary: when krisp_audio is absent, selecting the strategy fails
  with a clean, actionable error and no other strategy is affected.

[raghavm243512]

docker build.sh should not be in the PR