tests: fix seven Windows-only failures, one POSIX assumption at a time

[jothimani-rajendran]

Seven tests fail on a clean main on Windows and pass on the CI runners. Each rests on a
POSIX assumption that was never true off POSIX. None is skipped, deleted, or weakened; one
commit per failure, each named for the assumption it removes.

Full suite on Windows: 1534 passed, 4 skipped, 0 failed (was 7 failed).

The three assumptions

os.path.expanduser reads HOME — five failures. It does on POSIX. ntpath.expanduser
reads USERPROFILE, then HOMEDRIVE + HOMEPATH, and never HOME. conftest.py's
autouse isolated_home fixture already sets all four, and its docstring says exactly this.
The five tests set HOME again by hand and then asserted against a path derived from that
one variable, so on Windows they measured a directory nothing had written to.

The failure modes were worth reading rather than just fixing:

test	what it actually reported
test_a_user_scoped_config_path_is_not_nested_under_the_repo	install wrote the right file; the assertion looked one level up
test_install_never_destroys_a_config_it_cannot_parse	the BOM case passed vacuously; the corrupt cases raised JSONDecodeError from the test's own json.loads, not ConfigUnreadableError from install
test_a_query_never_raises_on_an_unparseable_config	installed() returned False because nothing was there — right answer, wrong reason — and uninstall() had no file to fail on, hence DID NOT RAISE
test_install_appends_a_block_...untouched	install wrote a correct config; the test read the untouched original it had seeded and reported guard.py missing
test_reinstalling_replaces_our_block_...	FileNotFoundError before any assertion about stacked blocks ran

Each now takes isolated_home as a parameter. tmp_path stays where it means the repo
root, which is a different argument and still belongs there.

The platform default text encoding is UTF-8 — one failure. The generated pages carry em
dashes; both readers that compare a committed page to a freshly built one used a bare
open(). Under cp1252 every em dash read back as three characters, so all twelve pages
reported stale — and could not be made to pass, because regenerating wrote the same bytes
it had just failed to read. Fixed in the test and in examples/generate.py's --check
path, which the pre-commit hook and the examples CI job both run and which carried the
identical latent bug. The writer now pins encoding="utf-8", newline="\n" too: a Windows
regeneration was rewriting all thirteen pages with CRLF, so the same run that fixed one
page dirtied twelve others.

The filesystem carries a POSIX execute bit — one failure. st_mode & 0o111 is 0 for
every file on Windows, including one git records as 100755 and checks out executable on
Linux and macOS. The test now asserts on the mode git records for the path: that is what
travels with a clone, it decides whether the hook runs at all, and it is the same answer on
every platform. Confirmed non-vacuous — git currently records 100755, and the assertion
names the mode it found when it is anything else. An untracked hook now fails with its own
message rather than an index error.

One extra commit, not one of the seven

test_a_query_never_raises_on_an_undecodable_toml_config carries the same hand-set HOME
as the three install tests, but it passed on Windows. That is the worse outcome:
installed() returned False because the utf-16 file was not on the path it reads, not
because the TOML branch had decoded it safely. The test asserted its own setup and never
entered the branch it is named for.

Fixed the same way, in its own commit, and verified non-vacuous by writing a valid
installed config to that path (installed() → True), then overwriting it with the utf-16
body (installed() → False, no raise). Drop that commit if you would rather keep this PR
to the seven.

Checks

ruff check ., ruff format --check . and tests/check_stdlib_only.py clean. No blanket
ruff --fix was run; no import reordering was needed.

🤖 Generated with Claude Code