fix(llm): add generic opt-in tool-call text policy for all text LLM providers

[martinstiago]

Summary

• Adds an opt-in ToolCallTextPolicy setting (PRESERVE / SUPPRESS_AFTER_TOOL_CALL_DETECTED) that suppresses streamed text after a tool call is detected in the same response.
• The setting lives on base LLMSettings, so every LLM service accepts it uniformly and it's runtime-updatable via LLMUpdateSettingsFrame.
• Suppression is centralized in LLMService._push_llm_text: the base class owns a _tool_call_detected flag (reset per response via LLMFullResponseStartFrame in push_frame) and drops text when the policy is SUPPRESS_AFTER_TOOL_CALL_DETECTED. Providers call _note_tool_call_detected() at their tool-call detection site — one line per streaming loop.
• Covers all text-bearing LLM services: OpenAI-compatible (20+ subclasses via BaseOpenAILLMService), OpenAI Responses (WebSocket + HTTP), Anthropic, Google (Gemini + Vertex), and AWS Bedrock Converse. Realtime speech-to-speech services set tool_call_text_policy=None (unsupported — they don't stream TTS-bound text).
• Default is PRESERVE everywhere — non-breaking; no behavior changes unless users explicitly opt in.

Testing

• uv run pytest tests/test_llm_service_tool_call_text.py tests/test_openai_tool_call_text.py tests/test_bedrock_tool_call_text.py tests/test_anthropic_tool_call_text.py tests/test_google_tool_call_text.py tests/test_openai_responses_tool_call_text.py
• uv run ruff check src/pipecat/services/ tests/test_*tool_call_text.py
• Validated end-to-end with real API calls: OpenAI chat completions, OpenAI Responses, and AWS Bedrock Converse (nova-lite + claude). Function calls execute with correct args in both policies; pre-tool-call text (narration, thinking) is preserved; post-tool-call text is suppressed when the policy is SUPPRESS_AFTER_TOOL_CALL_DETECTED; suppression state resets between responses.

Documentation

• Docs PR #1112

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

Files with missing lines	Coverage Δ
src/pipecat/services/anthropic/llm.py	71.22% <100.00%> (+1.55%)	⬆️
src/pipecat/services/aws/llm.py	70.43% <100.00%> (+0.95%)	⬆️
src/pipecat/services/aws/nova_sonic/llm.py	29.52% <ø> (ø)
src/pipecat/services/google/gemini_live/llm.py	29.30% <ø> (ø)
src/pipecat/services/google/llm.py	72.34% <100.00%> (+0.08%)	⬆️
src/pipecat/services/inworld/realtime/llm.py	30.91% <ø> (ø)
src/pipecat/services/llm_service.py	90.14% <100.00%> (+0.10%)	⬆️
src/pipecat/services/openai/base_llm.py	86.13% <100.00%> (+14.49%)	⬆️
src/pipecat/services/openai/realtime/llm.py	54.88% <ø> (ø)
src/pipecat/services/openai/responses/llm.py	91.05% <100.00%> (+0.39%)	⬆️
... and 3 more

... and 2 files with indirect coverage changes

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[martinstiago]

Renamed the enum member and value to SUPPRESS_AFTER_TOOL_CALL_DETECTED / suppress_after_tool_call_detected per Weitian's naming feedback — clearer that suppression kicks in once a tool call is detected. Docs PR updated to match.