fix: orphan-gate pre-GUI mcp-stdio sweep — fix cold-start MCP disconnect (v0.8.2)

[iret77]

Summary

Fixes the 2026-06-08 field report: every cold start of Cowork / Claude Code shows "MCP aiui: Server disconnected" on the first MCP call.

Root cause (from the user's /tmp/aiui-trace.log)

When Cowork spawned aiui --mcp-stdio and no GUI was running, the child cold-started the GUI; the freshly launched GUI then ran kill_mcp_stdio_started_before_self (v0.4.43) and SIGTERM'd its own bootstrapper:

17:27:19.031  ---- mcp-stdio pid=2483 started ----
17:27:19.034  lifetime: gui channel not ready, launching GUI
17:27:19.076  ---- gui pid=2497 started ----
17:27:19.091  housekeeping: killing pre-GUI mcp-stdio child pid=2483 …
17:27:19.097  startup: killed 1 pre-GUI mcp-stdio child(ren);
              Claude Desktop will respawn them against the current binary

Claude Desktop might respawn; Cowork / Claude Code don't. So the bootstrapper just dies on every cold start. Once the GUI is warm, subsequent children attach fine (no new GUI start → no sweep) — which is exactly why the bug presents as "happens sometimes, then works fine for a while".

Fix

Same pattern as v0.4.46's Bug A (kill_orphaned_mcp_stdio_children): add is_orphaned_child as a filter on find_pre_gui_mcp_stdio_to_kill. The sweep still reaps older mcp-stdio children, but only when their parent wrapper is gone (real leak). A live bootstrapper has a live parent → spared.

Test plan

• cargo test --lib — 131/0 (new: pre_gui_kill_spares_bootstrapper_with_live_parent, pre_gui_kill_reaps_orphaned_older_child)
• cargo clippy --lib --all-targets — clean (-D warnings)
• svelte-check — 0 errors
• npm run build — ok
• Field verification after release: cold-start a Cowork session, confirm no "Server disconnected" toast and trace shows the bootstrapper not SIGTERM'd

🤖 Generated with Claude Code