[albertsikkema-elix]

Closes #138.

• Adds google (Gemini API / AI Studio) and google-vertex
  (Vertex AI) as native providers, using the google-genai SDK
  directly instead of routing through the OpenAI-compatible shim
• Fixes Gemini 3.x compatibility: thought signatures are
  round-tripped correctly via pydantic-ai's GoogleModel
• Vertex AI is ~50x cheaper than Opus for doc generation ($0.05 vs
  $0.32-$22.75 per repo in a four-repo trial)

Changes

New module: codewiki/src/be/google_provider.py

• create_client(): builds a genai.Client for either provider
  (API key for google, ADC for google-vertex)
• create_google_model(): wraps the client in pydantic-ai's
  GoogleModel via GoogleProvider(client=...) (works on both
  pydantic-ai 1.x and 2.x)
• call_google(): sync one-shot completions (callers run inside an
  active event loop, so async model_request_sync would deadlock)
• check_access(): credential validation against the model metadata
  endpoint
• Cost tracking via genai-prices through pydantic-ai's
  ModelResponse.cost()

Config changes:

• codewiki/src/config.py: google_project, google_location
  fields
• codewiki/cli/models/config.py: serialization, validation,
  is_complete() updates
• codewiki/cli/config_manager.py: Google-aware validation gating
  (no API key required for Vertex, no base URL for either)
• codewiki/cli/commands/config.py: --provider google|google-vertex,
  --google-project, --google-location options; config show and
  config validate branches for Google providers

Integration:

• codewiki/src/be/backend.py: GOOGLE_PROVIDERS set,
  is_google_provider(), needs_api_key(), needs_base_url()
• codewiki/src/be/llm_services.py: Google branch in
  create_main_model / create_fallback_model / call_llm
• codewiki/src/be/pydantic_ai_backend.py: _run_usage_with_cost()
  for Google agent-loop cost tracking
• pyproject.toml: google-genai dependency

Docs: guides/providers.md, guides/cli-reference.md,
guides/development.md, README.md, CHANGELOG.md

Test results

17 new tests across two files:

• tests/test_google_config.py (11 tests): provider config,
  validation, serialization, is_complete(), needs_api_key(),
  needs_base_url()
• tests/test_google_provider.py (6 tests): model normalisation,
  client creation, call_google usage extraction, cost tracking

All 332 tests pass (existing + new). No changes to existing test
files.

Live tested on Vertex AI (google-vertex, gemini-3.8-flash,
eu and global locations) against real repositories.

Breaking changes

None. Existing providers (openai, anthropic, bedrock) are unaffected.
The google-genai dependency is new but optional at runtime for
non-Google providers.

Notes for reviewers

• GoogleProvider(client=genai.Client(...)) is used instead of
  GoogleProvider(vertexai=True) because the latter only exists in
  pydantic-ai 1.x; the client form works on both 1.x and 2.x.
• Vertex AI never receives an API key in create_client() to
  prevent the SDK from switching to Express Mode.
• The genai client must be assigned to a variable before use (not
  created inline) to prevent Python GC from closing the underlying
  httpx client mid-request.
• trial.sh is included for testing convenience but can be excluded
  from the merge if preferred.

[anhnh2002]

Thanks @albertsikkema-elix, nice work! I pushed two small commits to your branch to fix minor points.

[albertsikkema-elix]

Thanks, glad to help!