Windows Desktop: managed node_repl helper fanout after project/chat activation

Environment:

• Windows
• Codex Desktop version: exact local version not captured in the fanout sample
• Browser / Chrome / Computer Use runtime surface present
• Browser plugin disabled in config during the recurrence
• Clean Codex profile rebuild had been performed; old SQLite/log pressure was not restored

Related issues:

• #24727
• #25488
• #24963

Observed:

• After a clean Codex profile rebuild, DB/WAL pressure was low and old chat/profile state was not restored.
• Disabling the bundled Browser plugin persisted across restart.
• Codex Desktop still rehydrated the managed [mcp_servers.node_repl] runtime entry.
• Opening/enabling additional project chats caused repeated helper fanout:
• 10 node_repl.exe
• 10 child codex.exe app-server --listen stdio://
• The main Codex Desktop process and main app-server process stayed alive.
• Manually terminating only those helper pairs restored responsiveness.

Why this may be distinct from the existing setup-refresh reports:

• The public Windows reports already show node_repl failing with windows sandbox failed: spawn setup refresh and config args not being respected.
• This report adds a lifecycle symptom: Desktop can repeatedly launch managed node_repl helper pairs after project/chat activation, even with Browser disabled and without local DB/WAL growth.
• I am opening this separately because the main symptom here is process fanout after project/chat activation, not only the setup-refresh startup failure already covered in the linked issues.

What I can provide:

• Repro/evidence from the local fanout sample.
• Process-tree sampling commands and redacted logs if helpful.
• A narrow public-repo PR for Windows sandbox diagnostics or helper-resolution behavior if maintainers confirm that path is relevant.

Boundary:

• I do not have the closed Desktop/plugin runtime source, so I cannot provide the full Desktop lifecycle patch unless maintainers point to the relevant public code or ask for a targeted change in the open repo.

Expected:

• Desktop should dedupe or reuse the managed node_repl runtime per relevant scope.
• Disabling Browser should not leave Browser/Computer Use helper launch loops active.
• Thread/project activation should shut down stale managed MCP/runtime helpers.
• User mcp_servers.node_repl disable/config should either be respected or clearly documented as not controlling the Desktop-managed runtime.

Additional note:
Manually terminating only those helper pairs restored responsiveness without killing the main Desktop app. This suggests the runaway process set is narrower than the whole Desktop process group. Local containment is not a recommended product workaround.