fix: reconnect stale MCP tool wrappers

[gitgrooves]

Summary

• attempt in-process MCP rediscovery when a generated MCP tool wrapper has lost its server session
• preserve the existing MCP server '<name>' is not connected error when rediscovery cannot recover the session
• add regression coverage for stale wrapper recovery and failed reconnect behavior

Test Plan

• uv run pytest tests/cron/test_scheduler_mcp_init.py tests/tools/test_mcp_stale_wrapper_reconnect.py -q

Context

A long-running Hermes gateway can keep generated MCP tool wrappers after the underlying client session has gone stale/disconnected. Raw MCP endpoint health can be green while the active wrapper still returns not connected. This adds a best-effort reconnect path at wrapper invocation time before surfacing the error.

[alt-glitch]

Duplicate of #23540 — both add the same lazy/best-effort in-process MCP reconnect at the _handler tool-invocation site in tools/mcp_tool.py. When not server or not server.session, both introduce a helper (here _try_reconnect_disconnected_server; #23540's _get_connected_server_for_call) that pops the stale _servers entry and calls discover_mcp_tools() to reconnect before surfacing the MCP server '<name>' is not connected error. Same file, same edit site, same mechanism. #23540 is the earlier still-open version, so consolidating there.

[gitgrooves]

Closed: opened in error; withdrawing this patch from public review.