fix(tests): finish the Windows portability tail — path separators, signal handlers, skills paths

[mjerris]

Finishes the Windows portability work: the 15 TEST failures left after PR #76 took the count from 36 → 15. Measured from Multi-OS run 30261304144 (windows-latest).

Three defect classes. One is a product bug, not a test bug.

1. PRODUCT — RelayClient.run() was broken for every Windows user (7 failures)

_run_forever() called loop.add_signal_handler(SIGINT, ...) unguarded, as its first statement:

signalwire\signalwire\relay\client.py:684: in _run_forever
    loop.add_signal_handler(
...\asyncio\events.py:582: in add_signal_handler
    raise NotImplementedError
E   NotImplementedError

Loop-level signal handling is a Unix-only asyncio capability — both Windows event loops raise NotImplementedError unconditionally. So the exception escaped before connect() was ever reached: RelayClient.run() could not establish a RELAY connection on Windows at all. This is a genuine platform limitation of asyncio, but calling it unguarded from product code is our defect — a Windows user hits it, not just the test harness.

Fixed with a NotImplementedError guard that degrades gracefully: Ctrl+C still stops the client (asyncio.run() surfaces KeyboardInterrupt, which run() already suppresses); only the graceful _shutdown() handshake is lost. Also replaced the odd __import__("signal") inline with a module-level import.

Per the bar set by #76 — where it skipped, it also added platform-independent coverage — this ships a regression test that forces the exact Windows condition on every OS (patching the loop method to raise what Windows raises) instead of a skipif(win32) that would leave Windows uncovered. Proven on macOS: against the unfixed product the new test fails with the same NotImplementedError at client.py:684 as the Windows traceback.

The 7th relay failure was a different mechanism (Request timeout for signalwire.connect). The ping-loop test patched _EXECUTE_TIMEOUT=0.01 around client.connect(), putting the auth round-trip under a 10 ms deadline. The handshake needs several event-loop turns (the mock queues the reply, then _recv_task must be scheduled to drain it), and 10 ms is at/below Windows' asyncio timer granularity (~15.6 ms clock tick), so connect() could expire before the loop ever ran the recv task. Scoped the patch to the pings the test is actually about. Reproduced on POSIX with a deadline sweep — the debug log shows Response for unknown/expired request, i.e. the reply does arrive but the deadline already passed:

_EXECUTE_TIMEOUT=1.0      failures= 0/20  connected OK
_EXECUTE_TIMEOUT=0.01     failures= 0/20  connected OK
_EXECUTE_TIMEOUT=0.0      failures=20/20  RELAY error -1: Request timeout for signalwire.connect

2. TESTS — POSIX-separator expectations (3 failures)

test_schema_utils (2) and test_registry (1) compared product output against hardcoded POSIX literals. The product is separator-correct in all three cases — it returns str(Path(...)), composes with os.path.join(), and splits on os.pathsep. Only the expectations were POSIX-only, so these are test bugs, and a skip would be illegitimate here.

Each expectation is now built the way the product builds the value (compare Path objects / computed joins). Verified under Windows path semantics (ntpath / PureWindowsPath) from POSIX — the old literals match only on POSIX, the new expectations match on both, so Windows keeps real coverage:

== manual_search: product composes os.path.join(os.getcwd(), 'schema.json')
  posix    product='/current/schema.json'   old_literal_match=True  new_computed_match=True
  windows  product='\current\schema.json'   old_literal_match=False new_computed_match=True

3. TESTS — claude_skills paths (5 failures)

• WinError 267 (invalid directory) — the test passed Path("/tmp") as the subprocess cwd. /tmp doesn't exist on Windows, so the spawn failed before the timeout could fire (assert '[command timed out:' in '[command error: sleep 10]'). Switched to pytest's tmp_path. The command was also sleep 10, which isn't a Windows shell builtin and exits immediately, so the timeout contract wouldn't be exercised even with a valid cwd — replaced with a portable python -c sleep. The test now takes 1.01 s, confirming it genuinely blocks and hits the timeout rather than passing on a spawn error.
• 8.3 short-path mismatches (3) — setup() canonicalizes skills_path via .resolve(), which expands the Windows 8.3 temp dir to its long form (RUNNER~1 → runneradmin). The tests compared an unresolved tempfile path to resolved product output. Both sides now resolve. Correct product behavior; test-side bug, as the short-path shape suggested.
• Removed all 15 hardcoded /tmp uses from this file per the standing project rule; apart from the cwd above they were inert placeholders (_make_skill never calls setup(), so skills_path was unvalidated).

Disposition summary

No test was blanket-skipped, and no skipif(win32) was added. Every one of the 15 is either a product fix (7) or an expectation corrected to match platform-correct product behavior (8).

Verification

bash scripts/run-ci.sh → exit 0, all gates PASS:

[TEST] python -m pytest tests/unit/ (STRICT-MOCKS: MOCK_RELAY_STRICT=1) ... PASS
[FMT] ruff format (local: apply; CI: --check) ... PASS
[LINT] ruff check zero findings ... PASS
[TYPECHECK] mypy zero findings ... PASS
[DRIFT] python_signatures.json unchanged after regen ... PASS
[SPEC-PARITY] implemented routes == canonical spec ... PASS
[REST-COVERAGE] every implemented REST route covered success+error ... PASS

tests/unit → 5710 passed, 3 skipped, 0 failed (5699 before; +1 new regression test). Working tree clean after the FMT gate, so no auto-applied formatting is left uncommitted.

Windows itself still needs CI confirmation: Multi-OS runs on cron + workflow_dispatch only, so it does not run on this PR. A dispatch against this branch is noted in the thread.

Sequencing

• Land after fix(tests): Windows portability — sqlite handles, path assertions, permission bits, UTF-8 encoding #76 (fix/win-test-portability) — its fixes are the other half of the 36 → 15 → 0 count, and this PR's description is written against that baseline. No file overlap, so the order is about the count, not conflicts.
• Independent of fix(ci): declare build — multi-OS PACKAGE-SMOKE never had the module it needs #75 (fix/multi-os-python-interpreter), which declares the missing build dev-dep. Python's macOS PACKAGE-SMOKE failure (No module named build) is known, pre-existing, and unrelated to this diff — it clears when fix(ci): declare build — multi-OS PACKAGE-SMOKE never had the module it needs #75 merges.

Branched from origin/main (d782db8); the diff is 5 files and contains no content from #75 or #76.

🤖 Generated with Claude Code

https://claude.ai/code/session_01GFKJhLvfV8yGrASwqxdgaf