fix: permanent fd redirect for stdout suppression (issue #42)

[brianlball]

Fixes #42

Problem

Two related bugs corrupt the MCP JSON-RPC stream on stdout:

1. Concurrent tool timeout (issue MCP error -32001: Request timed out on all tools — stdout suppression race condition #42): Global FastMCP middleware held os.dup2() on fd 1 for the entire tool call. Concurrent worker threads interleaved their fd redirects, sending JSON-RPC responses to stderr — clients received nothing → -32001 timeout.

2. Polyhedron stdout leak: OpenStudio's C++ geometry engine writes [utilities.Polyhedron] diagnostics directly to fd 1. These aren't SWIG GC warnings — they fire during read-only queries like get_building_info on complex models. Any per-callsite suppression approach misses them.

Root Cause

Both bugs stem from the same design flaw: toggling fd 1 back and forth between stdout and stderr on a per-call basis. This is inherently racy under concurrency and inherently incomplete against unknown C-level callsites.

Fix

Redirect fd 1 to stderr once at startup, then give Python sys.stdout a private fd to the real MCP client pipe.

saved_fd = os.dup(1)        # copy real stdout pipe
os.dup2(2, 1)               # fd 1 → stderr permanently
sys.stdout = fdopen(saved)  # Python writes to saved fd

After this:

• C code (printf, SWIG, Polyhedron, anything) → fd 1 → stderr. Harmless.
• Python (FastMCP JSON-RPC) → saved fd → MCP client. Clean.
• No toggling, no races, no per-callsite wrappers, no missed callsites.

Files Changed (4)

File	Change
stdout_suppression.py	Replace toggle context manager with one-shot redirect_c_stdout_to_stderr(). Keep suppress_openstudio_warnings() as no-op for backward compat.
server.py	Call redirect_c_stdout_to_stderr() before mcp.run()
tests/test_concurrent_tools.py	Regression test: concurrent slow+fast and fast+fast tool calls
.github/workflows/ci.yml	Add concurrent test to shard 5

Test Results

Test	Before	After
Concurrent slow+fast (baseline_osm + status)	FAIL — 30s timeout	PASS
Concurrent fast+fast (status + status)	FAIL — connection closed	PASS
Stdout purity (44-zone Polyhedron)	FAIL	PASS
Full CI shard 4 (162 tests)	1 failed	0 failed

Supersedes PR #43

PR #43 (from anchapin's fork) used per-callsite targeted suppression with RLock. That approach fixes the race but misses C-level stdout from unexpected callsites (Polyhedron geometry warnings). This PR uses a permanent redirect that catches everything.

🤖 Generated with Claude Code

[brianlball]

Closing — will apply fix directly to optimize branch and merge optimize→develop instead.