fix(mcp): isolate stdio parsing from process pools

[tirth8205]

Summary

Safe current-main replacement for #615, retaining Minxi Hou as co-author.

• Marks stdio transport explicitly while the in-process FastMCP server is running.
• Uses thread parsing only for active MCP stdio, on every platform.
• Preserves the process-pool default for ordinary Unix CLI, redirected CI, and cron automation.
• Preserves the existing Windows non-TTY compatibility fallback and explicit environment override.
• Restores the prior transport state even when startup, auto-watch, prewarm, or server execution fails.

Source reconciliation

The source head no longer contains its claimed production executor change. Its added test invokes a module without the real entry point, forces thread mode, supplies too few files to enter the parallel path, detects descendants only after the parent exits, uses pgrep on Windows, and accepts crashes or nonzero exit codes.

This replacement uses the real python -m code_review_graph serve entry point and a JSON-RPC initialize plus full-build tool call over stdio. It builds 10 files so parallel parsing is entered, verifies the complete response and database, closes stdin, and requires a clean zero exit.

Overlap resolution

The branch is based exactly on main 4f9f035. Its incremental.py changes auto-merged with #650 NUL-path discovery. The only textual conflict was CHANGELOG ordering; both independent entries were retained.

Verification

• 102 focused incremental, transport, and real stdio shutdown tests passed after the final rebase.
• 1,657 broader non-daemon tests passed, with 1 expected skip and 2 existing xpasses.
• Ruff and git diff checks passed.
• Graph review found no affected execution flows; explicit transport/executor and end-to-end regressions cover the changed boundary.
• Full hosted Python and native Windows CI are the merge gates.

[github-actions]

code-review-graph review

Overall risk: 0.60 (MEDIUM) — 14 changed function(s)/class(es), 0 affected flow(s), 8 test gap(s)

Risk-scored changes

Risk	Level	Symbol	Location	Tested
0.60	medium	code_review_graph/main.py::main	code_review_graph/main.py:1046	no
0.50	medium	tests/test_main.py::TestServeMainTransport.test_http_calls_mcp_run_with_host_port	tests/test_main.py:112	(test)
0.40	medium	code_review_graph/incremental.py::full_build	code_review_graph/incremental.py:873	no
0.40	medium	tests/test_mcp_stdio_shutdown.py::_send	tests/test_mcp_stdio_shutdown.py:16	no
0.40	medium	tests/test_mcp_stdio_shutdown.py::_read_response	tests/test_mcp_stdio_shutdown.py:22	no
0.35	low	code_review_graph/incremental.py::_select_executor_kind	code_review_graph/incremental.py:34	no
0.30	low	tests/test_incremental.py::TestParseExecutorSelection	tests/test_incremental.py:28	no
0.30	low	tests/test_incremental.py::TestParseExecutorSelection.test_stdio_mcp_uses_threads_on_unix	tests/test_incremental.py:29	(test)
0.30	low	tests/test_incremental.py::TestParseExecutorSelection.test_non_mcp_unix_automation_keeps_process_default	tests/test_incremental.py:39	(test)
0.30	low	tests/test_incremental.py::TestParseExecutorSelection.test_explicit_process_override_wins_in_stdio_mcp	tests/test_incremental.py:49	(test)

Test gaps

• code_review_graph/incremental.py::_select_executor_kind (code_review_graph/incremental.py:34)
• code_review_graph/incremental.py::full_build (code_review_graph/incremental.py:873)
• code_review_graph/main.py::main (code_review_graph/main.py:1046)
• tests/test_incremental.py::TestParseExecutorSelection (tests/test_incremental.py:28)
• tests/test_main.py::TestServeMainTransport (tests/test_main.py:92)
• ...and 3 more without direct tests

Token savings: this graph-backed report used ~51,552 fewer tokens (~94%) than reading every changed file in full (estimated, chars/4 approximation).

---

Powered by code-review-graph — local-first analysis; no code leaves the CI runner.