Bug: MCP-v2 HTTP calls fail after an idle period with "the server went away"

Issue details

With FX_MCP_ENGINE=v2, the first call to a remote HTTP MCP server after a few idle minutes can fail with:

MCP server '<name>' ended the call: the server went away. The tool may have run.

The server never receives the call.

Steps to reproduce:

1. Add a remote MCP server hosted on Vercel to ~/.fx/mcp.json with "type": "http".
2. Start fx with FX_MCP_ENGINE=v2 and call one of its tools. The call succeeds.
3. Wait six minutes and call the tool again.

Expected: the call reaches the server, as it does with the default engine.

Actual: the call fails with the error above.

Client.run and Client.fetchRun in src/mcp/io/http.zig use the std.http.Client keep-alive pool, which hands back an idle connection without checking it. Vercel's edge closes idle connections after about 340 seconds, so the next request goes out on a closed socket. The v1 engine sets keep_alive = false on every request. #1366 applies the same setting to v2 and adds a deterministic test that fails on main.

Reproduced with fx 0.0.14 on macOS arm64.

Trace

No response

Sensitive information

• I reviewed the information above and removed secrets and other sensitive data.