Phase 7.3: LLMService facade (#644)

[ShaerWare]

Summary

• New LLMServiceImpl in modules/llm/facade.py — unified facade implementing the LLMService Protocol from modules/llm/protocols.py
• Wraps CloudLLMService / VLLMLLMService (generation) and CloudProviderService (provider CRUD) behind a single interface
• Lazy access to container.llm_service — handles runtime backend switches by InternetMonitor
• generate() — sync provider call wrapped via asyncio.to_thread
• stream() — sync generator consumed in thread, yields normalized StreamChunk TypedDicts
• _to_stream_chunk() normalizes both plain str and dict (tool_calls) from providers
• _to_provider_info() converts provider dicts to ProviderInfo TypedDict
• Registered in ServiceContainer.llm_service_facade during orchestrator startup
• Module-level singleton llm_service_facade set during startup for direct imports

Scope note: This PR creates the facade layer only. Prompt assembly, RAG injection, agentic tool loops, and token accounting remain in callers (chat router, widget router). Migration of callers to use the facade is a separate step.

Closes #644

Files changed

File	Change
modules/llm/facade.py	New — LLMServiceImpl + converters
orchestrator.py	Create facade during startup
app/dependencies.py	Add llm_service_facade field to ServiceContainer
CLAUDE.md	Document Phase 7.3 completion
tests/unit/test_llm_service.py	New — 13 tests (protocol, signatures, converters)

Test plan

• 13 new unit tests pass (protocol compliance, signatures, converters)
• All 183 existing unit tests pass
• Ruff lint + format clean
• CI green

🤖 Generated with Claude Code