The tooling suite asserts machine properties with no platform-conditional convention, so it is red on one box and wrong-way-green on the other nf-0178 ← Beads

Twinned from https://github.com/cfsh/ninefold/issues/565 by tools/beads/import_github.py's reconcile pass.

The tooling suite asserts machine properties in several places, and has no platform-conditional convention at all: grep -c 'skipIf\|skipUnless' tools/tests/*.py is 0 across the whole suite. That fails in both directions, and neither is detectable from the box that causes it.

Direction 1 — the office box's gate is permanently red

./presubmit.py --suite python on the WSL office box: 3 failures, every run, on a clean checkout of main. All three assert something about the *Windows* machine:

| test | what it asserts | |---|---| | test_nflib_proc.test_a_child_that_imports_nflib_can_print_these_glyphs | that a bare python -c "print('✓')" fails — i.e. that stdout is cp1252. On Linux it succeeds, so assertNotEqual(code, 0) fails | | test_presubmit.test_doctor_detects_a_broken_interpreter_by_RUNNING_it | the Microsoft Store stub's does NOT execute output | | test_rebuild_f5.test_registration_matching_is_case_insensitive_on_windows | normcase folding a drive letter |

Each is a *good* test — the glyphs one pins the cp1252 blocker nflib exists to fix, and its docstring is explicit that going green means the blocker is gone and the design note wants revisiting. The problem is that it runs unconditionally on a box where the premise is false.

Consequences, in order of cost:

1. presubmit.py prints FAILED: python on the office box always. The gate is red before a session has written a line. To tell its own breakage from the baseline a session must git stash -u, re-run, compare counts, and pop — which I did on #558, and which nothing tells you to do.
2. The baseline is tribal knowledge. It is documented in exactly one place: session rows on the board (#398), phrased as "3 documented Linux-only failures". grep -rn 'known.*fail\|Linux-only' over notes/, CLAUDE.md and presubmit.py returns nothing. A session spawned on this box without that line in its kickoff sees a red gate and no explanation.
3. A gate that is always red teaches sessions to ignore it. Same argument as #560, and the same shape as #389 — the failure presubmit.py's own docstring says it exists to prevent is "a gate that prints green while doing less than the reader assumes"; this is its mirror.

Direction 2 — Windows-red, and invisible from where it is written

The reverse is worse because nothing on the office box can see it. PR #558 shipped three defects of this shape, and it is the whole reason that PR took six review rounds:

| | passed on office box, failed on Windows | |---|---| | F1 | str(Path) uses the host separator, so a replace("/", "-") was a no-op | | F7 | as_posix() left a C: prefix, and PureWindowsPath treats a segment starting C: as drive-relative — introduced by F1's fix | | F12 | a MainTest assertion depended on tmux being on PATH — introduced by #563.1's fix |

Two of the three were introduced *by the fix for the previous one*. The only detector was ninefold-reviewer checking the branch out on the sandbox box and running the suite there — six times. That is not a repeatable process; it happened because that PR drew unusual review attention.

⚠ This direction cannot be fixed by skipping. A skip marker helps a test that *knows* it is platform-specific. These were tests nobody knew were platform-specific.

What might close it

Directions, not a decision — parts of this are needs:direction:

• Mark the machine-property tests. @unittest.skipUnless(os.name == "nt", "asserts a Windows machine property") on the three, so the office box's suite is green and a real regression is visible on the line it happens. Cheapest, closes direction 1, and establishes the convention the suite currently lacks. ⚠ Do it as *skips*, not as a presubmit-level expected-failures list — a list in the runner is a second place to keep in sync, and the reason belongs next to the assertion.
• Give presubmit.py --doctor a baseline line — "this box cannot run: dotnet, godot; N tests are skipped as Windows-only" — so a session's first command tells it what its gate does and does not cover.
• For direction 2, seams beat vigilance. #558's fix for F12 is the pattern that worked: main grew a which= seam so a test states which box it is pretending to be, and the rule went in the test module docstring rather than in the author's head. Applying that deliberately — no test asserting on Path string form or on what is installed — is a repo-wide sweep, not a rule one file can hold.
• Or run the suite on both boxes. That is #417 (hosted vs self-hosted runner) and this issue is a concrete argument for it: three defects in one PR that only a second box could see.

Why file rather than fix

Direction 1 is ~4 lines and I would happily have folded it into #558 — but it touches test_nflib_proc, test_presubmit and test_rebuild_f5, which are the tests for the merge path, and #558 is a spawn tool. Genuinely unrelated files, which is CLAUDE.md's bar for a second PR. Direction 2 is a chunk, not a fix.

Filed unclaimed by nf-dev-office-spawn while working #557 (PR #558). Not on my chunk. Related: #560 (a wake that cannot be cleared trains sessions to ignore it), #417 (where CI would run), #541 (five pieces rediscovering the same environment traps).