fix(mcp): lazy reconnect on dead session + transport errors for cron jobs

[zccyman]

Summary

Fixes #26042

When cron jobs call MCP tools, the connection may have gone stale since the MCP servers were initialized by the gateway process at startup. The tool handler returned "MCP server X is not connected" without any recovery attempt.

Changes

1. _reconnect_server_sync() (new helper)

Reuses the existing _reconnect_event mechanism (same as auth-recovery and session-expiry paths) to trigger a transport reconnect and wait for readiness. Thread-safe — sets the event on the MCP loop.

2. Lazy reconnect in _make_tool_handler (line ~2204)

When server exists but server.session is None, attempts one _reconnect_server_sync() before returning the "not connected" error. This handles the case where the MCP child process crashed between cron ticks.

3. Transport-level error retry (line ~2342)

After auth-recovery and session-expiry paths, detects transport errors (ECONNREFUSED, ECONNRESET, Connection refused, Connection reset, EPIPE) in the exception message. Triggers a reconnect + single retry before surfacing the error.

Testing

• 184 existing MCP tool tests pass with no regressions

[alt-glitch]

Related: #23540 (lazy reconnect MCP tools on call), #19208 (stale session retry), #21006 (reconnect dead stdio before breaker probe). Umbrella issue: #18165 (unify MCP HTTP recovery).

The cron-specific angle (dead session between ticks) may be novel, but the reconnect mechanism overlaps significantly with existing PRs.

[zccyman]

Thanks @alt-glitch. Closing in favor of #23540 — that lazy reconnect approach is more comprehensive. The cron-specific angle can be addressed as a follow-up on #23540 if needed.