Recover cleanly when stdio MCP transport fails; prefer stable local HTTP bridge

[Trace824]

Summary

stdio MCP clients sometimes fail with transport/unreachable errors while the local FSB HTTP endpoint (127.0.0.1:7226) remains healthy. Recovery should prefer the stable transport without manual guesswork.

Repro

• Context: MCP client configured for stdio to FSB; hub may still be up.
• Steps:
  1. Induce stdio disconnect / unreachable client error.
  2. Hit http://127.0.0.1:7226/health and /mcp.
• Observed: stdio path dead; HTTP path still works; switching is undocumented / ad hoc.

Code pointers (verified against /workspace/FSB)

• mcp/README.md — http://127.0.0.1:7226/mcp, /health; hub 7225.
• mcp/src/native-host/constants.ts — port 7226.
• mcp/src/errors.ts — bridge disconnect / port 7225 heuristics (~L183).
• mcp/src/bridge.ts — hub/relay lifecycle.

Proposed fix

Document and, where possible, auto-fallback or clear error text: if stdio is unreachable, use Streamable HTTP on :7226. Improve error messages to cite health URL and hub port.

Acceptance / regression

• Unreachable stdio yields actionable error pointing at :7226/health.
• Documented recovery without deleting profiles.

Owner

mcp-server + docs

---

Priority: P2 · Owner: mcp-server