Doesn't work on Windows 11

Win 11, cmd, claude code - opus 4.8

Stops here : zenith - advance_project (MCP)
constant repetitions and the improvement itself has yielded nothing but wasted money and time -- report": "Task report from W1-foundation\ntype: work\ndone: False\nrequest_attention:
False\n\nreport:\nWorker MCP server failed to start (attempt 3/3; port not ready; port 53337; process rc=1;
exe=D:\AI 2025\Zenith\zenith\.venv\Scripts\python.exe). stderr tail: | stdout tail: "

acp_runner.py - doesnt start woker thru harness
if i run thru uv - all Ok, MCP server started

• After 5 sessions, the diagnosis has been stuck because the only diagnostic channel is the child's stdout/stderr
  PIPEs — and those pipes are exactly the suspect. The live child reports "empty output" precisely because whatever is
  wrong happens on the pipe/handle boundary it inherits from the live MCP stdio server.
• Two live-only variables remain unreplicable from bash: (1) the worker inherits the orchestrator's stdin handle,
  which is the live MCP stdio pipe Claude Code actively reads/writes; (2) Claude-Code-injected process state. The fix
  must (a) remove the inherited-stdin variable (stdin=DEVNULL — correct daemon hygiene regardless), and (b) bypass pipes
  for diagnostics by having the worker write a startup trace to a file, so the next live attempt is self-diagnosing no
  matter what the pipes do.

Start with acp_runner.py and compare the worker startup path used by the harness with the successful uv invocation on Windows 11. Reproduce the failure through Claude Code, remove the inherited stdin variable, and add the startup trace described in the issue so the next attempt records the cause; done means the MCP server starts through the harness on port 53337.