[P1.5-S3-ARI-21] `VertexAILLMClient` — new `LLMClientInterface` implementation

[bjridicodes]

Ticket: P1.5-S3-05

Type: Feature | Est: 1.5d

Goal: ARIA runs LLM inference via Vertex AI (Gemini or Claude-on-Vertex) when deployed on GCP — no Anthropic API key or Claude Code subscription required in the container.

Scope:

• implementations/llm/vertex_ai/llm_client.py (new): VertexAILLMClient implements LLMClientInterface
  • Constructor: model: str, project_id: str, location: str = "europe-west1"
  • Auth via ADC (google.auth.default()) — no credentials in constructor
  • complete(messages, max_tokens, temperature, system):
    • Gemini models: use google.cloud.aiplatform.gapic.PredictionServiceClient or vertexai.generative_models.GenerativeModel
    • Claude-on-Vertex: use Anthropic Messages API via Vertex AI endpoint
  • Raises LLMAuthError, LLMUnavailableError, LLMResponseError matching the existing interface contract
• requirements.txt: add google-cloud-aiplatform >= 1.60.0
• implementations/llm/vertex_ai/__init__.py (new, empty)

Supported models: gemini-2.0-flash, gemini-2.5-pro, claude-sonnet@20250201 (Claude-on-Vertex)

Acceptance criteria:

• VertexAILLMClient fully implements LLMClientInterface (passes interface contract test in tests/unit/test_interfaces.py)
• complete() returns a string for both Gemini and Claude-on-Vertex model identifiers
• Unit test with mocked aiplatform SDK; assert correct API endpoint and payload shape
• LLMAuthError raised on ADC failure (mocked)

[bjridicodes]

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