layers_split uses Gemini even when provider=mock in MCP

[iamliuhaitian-coder]

Bug: layers_split ignores provider=mock and still calls Gemini during analyze stage

Summary

When calling layers_split via MCP with provider="mock", the tool still attempts a Gemini API call in layer analysis, causing failure in offline/no-key environments.

Environment

• OS: macOS Darwin 21.1.0 (arm64)
• Vulca MCP command: /Users/liuhaitian/Documents/PHD/2026/upgrade/vulca/.venv/bin/vulca-mcp
• FastMCP client: fastmcp 3.2.4
• Invocation mode: MCP stdio

Reproduction

1. Configure MCP server vulca.
2. Run:
   • view_image(image_path=".../theory-materiality-practice-diagram.png") (works)
3. Run:
   • layers_split(image_path=".../theory-materiality-practice-diagram.png", output_dir=".../vulca_mcp_demo_layers", mode="sam", provider="mock", tradition="default")

Expected behavior

• With provider="mock", the whole layers_split pipeline should avoid Gemini dependencies.
• In no-network/no-key setups, either:
  • complete successfully using mock/local path, or
  • fail fast with a clear message that provider=mock is not supported for analyze stage.

Actual behavior

• Tool retries Gemini calls and fails with:
  • litellm.APIConnectionError (offline)
  • litellm.AuthenticationError (API_KEY_INVALID) when network is available but key invalid
• Output indicates failures in layers.analyze before split output is produced.

Evidence (trimmed)

Layer analysis failed after 3 attempts: litellm.AuthenticationError: GeminiException ... API_KEY_INVALID

Impact

• Breaks minimal/local-first MCP workflow for decomposition.
• provider semantics are unclear/inconsistent from user perspective.

Suggested fix

• Ensure layers_split(..., provider="mock") passes provider choice through analysis stage.
• Add explicit provider capability matrix in MCP tool docs/response.
• Return structured error with actionable remediation (missing_api_key, provider_not_supported, etc.) instead of generic pipeline failure.