Windows: full-suite debt of 194 failing tests (measured 2026-10-10), fix order by root cause #692

Summary

Cortex runs about 9000 tests on Linux and only 20 hand-picked test files on Windows. A one-off measurement of the full suite on Windows found 194 failing tests (188 failed, 6 errors) against 0 on the Linux control and 3 on macOS. This issue tracks the Windows debt as a sequence of root causes; the verified product defects and the bash question have their own issues.

Evidence

Measured on 2026-10-10 at main 18bd25b, SQLite backend, Python 3.12, dependencies from uv.lock, the process-leak guard active, no -x. Two identical runs gave the identical set of failing ids.

• Run 1: https://github.com/cdeust/Cortex/actions/runs/38008915302, run 2: https://github.com/cdeust/Cortex/actions/runs/38011318496, diagnostics: https://github.com/cdeust/Cortex/actions/runs/38012683041.
• Totals: ubuntu 9197 passed, 0 failed; macos 9194 passed, 3 failed; windows 8933 passed, 194 failing, 78 more skips than the control. The Windows full suite takes about 32 minutes against about 5 for Linux.
• By directory: tests_py/scripts 104, benchmarks 40, hooks 23, infrastructure 16, handlers 6, core 2, shared 2, invariants 1.

Failure classes (mutually exclusive, they sum to 194):

1. 50 tests: the empty_env fixture of tests_py/scripts/test_launcher_platform_coverage.py (line 92) returns Scripts/python without .exe. Once fixed, uv's cross-platform resolution runs on Windows for the first time and may expose more.
2. 42 tests: bare bash resolves to the WSL launcher (tracked in its own issue).
3. 16 tests: default text encoding is cp1252; 4 of them are a product defect in auto_recall (tracked in its own issue), 12 are tests that call read_text() without an encoding or compare mojibake.
4. 13 tests: benchmarks/lib/machine_load_snapshot.py:54 calls the POSIX-only os.getloadavg.
5. 21 tests assume a POSIX path or branch: /a.py is not absolute on Windows (test_preemptive_context 9), os.getppid patched while _window_claude_pid takes the Windows branch (session_registry.py:210-212, test_session_registry 5; PR fix(session-registry): resolve the window's claude.exe on Windows and probe liveness without os.kill (#665) #678 made its own tests platform aware but not this file), ~/state under USERPROFILE, is_supported() false so dispatch takes the spool branch (test_capture_dispatch), a path regex over tmp_path (test_validate_memory 4).
6. 4 tests use a POSIX primitive without a platform guard (os.O_NOFOLLOW, multiprocessing.get_context("fork"), select.select on a pipe).
7. 2 tests: tests_py/_hermetic_hook_env.py:91 sets HOME only, and Path.home() reads USERPROFILE on Windows.
8. 23 tests, suspected CRLF translation of fixtures written with write_text (test_migrate_decision_blocks 9, test_wiki_rebucket_file_docs 9, test_wiki_bulk_migrate 5); inferred from the code, not run.
9. 23 tests not isolated (first error lines recorded): installer rerun fixtures 6, file priming aliases 6 (heat 0.5 against 0.6: a symlinked project alias is scoped differently, possibly a product defect in the ADR-1086 scoping), mcp client spawn 2, a mocked Path.resolve 2, a TTL on a coarse Windows clock 1, test_capture_drain 1 (tracked in its own issue), and 5 singles.

Expected

Fix the classes in this order, one pull request each, re-measuring on Windows after each one with the scratch workflow used for the measurement: 1 (50 tests), then 3 (the product defect first), 2, 4, 8, then 5, 6 and 7 together, then class 9 one by one. After the first four steps 73 ids remain, after 5 to 7 about 20. The required Windows job then grows one directory at a time (not straight to the full suite), and the per-system skip counts (Windows 78, macOS 6 more than the control) are pinned in CI so that a new platform skip is a visible decision.