fix(mcp_cloud): migrate to MCP Python SDK 2.0

[neoneye]

The mcp 1.28.1 → 2.0.0 bump (#845 / #851) removed mcp.server.fastmcp, so mcp_cloud crashed on import and the container sat in a restart loop:

File "/app/mcp_cloud/server_boot.py", line 30, in <module>
  from mcp.server.fastmcp import FastMCP
ModuleNotFoundError: No module named 'mcp.server.fastmcp'

Four separate API changes had to be handled.

1. FastMCP → MCPServer

Now imported from mcp.server.mcpserver. Transport settings (streamable_http_path, json_response, stateless_http) moved off the constructor onto streamable_http_app(). SERVER_VERSION is now passed to the server, so initialize reports serverInfo.version: "1.0.1" instead of an empty string.

2. DNS rebinding protection — behaviour change worth a look

streamable_http_app() auto-enables DNS rebinding protection when host is a loopback address, restricting Host/Origin to localhost. PLANEXE_MCP_HTTP_HOST defaults to 127.0.0.1, so accepting the new default would reject the LAN clients mcp_cloud/README.md documents (http://192.168.1.10:8001/mcp/).

This PR sets transport_security explicitly to preserve the 1.28.1 behaviour. Access remains controlled by the API-key middleware and the configured CORS origins. If you would rather opt into the protection, that is a one-line change plus an allowed-hosts list — say the word and I will do it.

3. Vestigial lowlevel Server removed

The lowlevel Server in db_setup.py lost its list_tools() / call_tool() decorators in 2.0. Nothing ever served through that instance — the HTTP path goes through MCPServer plus the REST routes — so it is removed. handle_list_tools / handle_call_tool stay as plain functions, which is how the tests and the REST bridge already used them. The stdio entry point in app.main() now builds an MCPServer and calls run_stdio_async() rather than being left broken.

4. snake_case model fields

mcp.types models are snake_case in 2.0 with camelCase aliases. Attribute reads become input_schema / output_schema / is_error / structured_content (construction by alias still works). The REST /mcp/tools payload keeps its camelCase wire keys, so that contract is unchanged.

Verification

Against the running docker-compose stack — container healthy, and over Streamable HTTP:

Call	Result
initialize	protocol 2025-06-18, {'name': 'planexe-mcp-server', 'version': '1.0.1'}
tools/list	11 tools
prompts/list / prompts/get	getting_started, plan_a_project — both render
tools/call	model_profiles, example_prompts return real data, isError: false
REST /mcp/tools, /mcp/tools/call	200; plan_status + plan_file_info still correctly omit the oneOf outputSchema

The mcp_cloud suite was run against a live Postgres on this branch and on main with mcp 1.28.1. Both give an identical 5 failures + 2 errors — same test names, all pre-existing and unrelated to MCP (description-text drift, an annotation mismatch, SSE timing, auth-header parsing). No regressions.

Note: CI did not catch this

test.py ends with runner.run(tests) and never inspects the result, so it always exits 0. The tests job on the merge of #845/#851 printed FAILED (errors=25, skipped=13) and was still reported green. Not fixed here — flipping the exit code alone turns CI red, because the job has no Postgres service and 19 mcp_cloud tests error on connection. Filed separately.