[P1.5-S3-ARI-22] `llm.provider` config field + DI layer routing

[bjridicodes]

Ticket: P1.5-S3-06

Type: Feature | Est: 0.5d

Goal: Operators switch LLM provider by changing one line in conf.yaml — zero code changes needed to deploy ARIA with Vertex AI instead of Claude Code.

Scope:

• core/config.py: llm_provider() -> str reads llm.provider from conf (values: claude_code, anthropic, vertex_ai; default: claude_code)
• api/dependencies.py: new get_llm_client(agent_num: str) -> LLMClientInterface factory function:
  • claude_code → ClaudeCodeLLMClient(model=resolve_model(agent_num))
  • anthropic → AnthropicLLMClient(model=resolve_model(agent_num)) (requires ANTHROPIC_API_KEY)
  • vertex_ai → VertexAILLMClient(model=resolve_model(agent_num), project_id=gcp_project_id(), location=gcp_region())
  • Unknown value → ValueError at startup with clear message
• Replace direct LLMClient(model=...) construction in get_agent1(), get_agent3(), get_agent4(), get_pipeline() with get_llm_client(agent_num)

Acceptance criteria:

• llm.provider: vertex_ai routes all agent LLM calls to VertexAILLMClient
• llm.provider: anthropic routes to AnthropicLLMClient
• llm.provider: claude_code (or unset) routes to ClaudeCodeLLMClient
• Unsupported provider value raises ValueError at startup with clear message naming the valid options

[bjridicodes]

Closed by PR #91 (feat/p1.5-s3-docker-config-llm) — merged 2026-06-16.