Follow-up windows support tasks

[pschanely]

Windows CI: remaining follow-ups after enabling the Windows job

Windows CI is green as of #466 (first-ever passing Windows job). That PR
deliberately deferred a few things to keep the initial job green. This issue
tracks them.

Context: the suite is sound on Windows — no crashes, no correctness/soundness
bugs. The blocker was runtime (heavy z3 queries overrunning their soft
wall-clock timeout), addressed with an opt-in deterministic z3 work budget
(CROSSHAIR_SMT_RLIMIT, set to 200000000 for the Windows job only).

Deferred items

• Un-skip fuzz_core_test on Windows and triage the real gaps.
  Currently skipped via a module-level pytest.mark.skipif(sys.platform == "win32", ...) in crosshair/fuzz_core_test.py. A partial Windows run surfaced genuine failures in three buckets:

  • Windows-only ops not modeled/classified: msvcrt.open_osfhandle, msvcrt.SetErrorMode, ctypes.set_last_error (these don't exist on Linux, so CI never exercised them).
  • Platform-divergent ops: select.select, os.waitstatus_to_exitcode, turtle.back (needs a display).
  • Cross-platform, need verification (may also fail on Linux, or be timing-flaky): operator.pow, operator.ipow, statistics.linear_regression, bz2.decompress, ast.parse.
    Fix path: classify the Windows-only/OS-handle ops into the appropriate override tables (SIDE_EFFECT_OVERRIDES / _OS_HANDLE_PARAMS / PROBE_HAZARD_OVERRIDES), and root-cause the cross-platform divergences.

• Run the inputgen probe guard on Windows.
  inputgen_test.py::test_uncategorized_ops_probe_cleanly is skipped on Windows because probe_side_effect_isolated (crosshair/inputgen.py) uses multiprocessing.get_context("fork"), and fork is unavailable on Windows. Its docstring says this guard should run per-platform (it's what would force-classify the msvcrt/winreg/winsound ops above). A spawn fallback isn't a drop-in — the probed call holds live, often-unpicklable objects. Needs a fork-free probe design. (Bucket with the item above.)

• Normalize --unblock matching across platforms.
  Python's subprocess.Popen audit event has a different arg shape on Windows — (None, "<joined cmdline>", None, None) vs Linux (argv0, [argv], cwd, env). The auditwall matches positionally over map(str, args) (crosshair/auditwall.py), so --unblock=subprocess.Popen:echo doesn't match on Windows — a real (if minor) user-facing bug. Two tests are pinned-skipped on win32: auditwall_test.py::test_popen_allowed_if_prefix_allowed and watcher_test.py::test_auditwall_unblock[...echo...CONFIRMED]. Fix: canonicalize the subprocess.Popen args to a cross-platform shape before matching (design decision on the canonical form + a docs update), then un-skip.

• One-time git add --renormalize . commit.
  Enable Windows CI: bound z3 work with rlimit; fix test Unix-isms #466 adds .gitattributes (* text=auto eol=lf). Do a dedicated renormalization commit so files currently committed with CRLF get normalized and future diffs stay clean.

• Tune the rlimit budget.
  CROSSHAIR_SMT_RLIMIT=200000000 is validated but generous. Lower it for speed while watching for tests flipping pass → CANNOT_CONFIRM (that's the signal the budget is too tight). Also consider whether an rlimit belongs on more platforms as a general determinism win, and whether the underlying wall-clock-timeout overrun is worth addressing at the source.

• Expand the Windows matrix.
  The Windows job currently runs Python 3.13 only. Add more versions once it's proven stable.