Generalize model backends and default to Gemini 3.5 Flash

[eshulman2]

Summary

Make Forge provider-neutral at the product layer and use Gemini 3.5 Flash on Vertex AI as the default model backend.

Changes

• Add explicit LLM_BACKEND selection: vertex-ai, google-genai, or anthropic
• Default new configurations to vertex-ai with gemini-3.5-flash
• Use provider-native credentials:
  • GOOGLE_CLOUD_PROJECT / GOOGLE_CLOUD_LOCATION plus ADC for Vertex AI
  • GOOGLE_API_KEY for the Gemini API
  • ANTHROPIC_API_KEY for the Anthropic API
• Keep existing LLM_API_KEY, VERTEX_*, and ANTHROPIC_VERTEX_* aliases working
• Infer legacy backends for existing Anthropic, Gemini API, and Vertex deployments that do not set LLM_BACKEND
• Route orchestrator and container model construction through the same explicit backend/model contract
• Reject incompatible backend/model combinations with actionable configuration errors
• Replace product-facing Claude-specific language with Deep Agents/LangChain terminology
• Update setup and configuration documentation

Default configuration

LLM_BACKEND=vertex-ai
GOOGLE_CLOUD_PROJECT=your-gcp-project
GOOGLE_CLOUD_LOCATION=global
LLM_MODEL=gemini-3.5-flash

Vertex AI uses Application Default Credentials.

Validation

• uv run ruff check src/forge containers/entrypoint.py
• uv run ruff format --check src/ containers/entrypoint.py
• Backend/configuration/model-factory/container-environment matrix: 23 tests passed
• Python compilation check for configuration, agent factory, and container entrypoint

Compatibility

Existing deployments using only ANTHROPIC_API_KEY continue on the Anthropic backend with the prior direct-compatible model ID. Existing legacy Vertex project variables continue to select Vertex AI. GOOGLE_API_KEY without an explicit backend selects the Gemini API. New deployments default to Gemini 3.5 Flash on Vertex AI.