test(build-context): write fixture content with the newlines the assertions pin - #505

Problem

Five cases in tests/nodes/test_build_context.py assert cached content exactly, including the
trailing newline:

assert result["file_cache"].get("references/guide.md") == "# Reference guide\n"
assert result["file_cache"].get("scripts/run.py") == "print(1)\n"
assert result["local_file_cache"][".env"] == "TOKEN=not-reported\n"
assert result["raw_file_cache"][".env"] == b"TOKEN=not-reported\n"

but the fixtures are written with Path.write_text and no newline argument:

(root / "references" / "guide.md").write_text("# Reference guide\n", encoding="utf-8")

Text mode translates "\n" to os.linesep, so on Windows the fixture lands on disk as CRLF.
build_context reads it back faithfully and returns CRLF, and the case fails against content
the test itself wrote. The raw_file_cache assertion above makes the intent explicit — it pins
the bytes — so the fixture, not the product, is what is wrong here.

_write_real_oms_signature has the same problem in both directions:

target.write_text(_OMS_FIXTURE.read_text(encoding="utf-8"), encoding="utf-8")

read_text applies universal newlines and write_text applies the reverse translation, so the
pinned signature is newline-translated twice on the way to the temporary directory. Its case
then compares the cache against nested.read_text(...), which normalises again, and the two
never line up.

Fix

• _write_real_oms_signature copies bytes. That is what a pinned signature fixture needs
  regardless of platform, and it removes both translations.
• The writes whose content is asserted exactly pass newline="\n", so the fixture matches what
  the assertion pins. (Path.write_text has accepted newline since 3.10; this project
  requires >=3.12.)

Writes whose content is not asserted verbatim are left alone, so the diff stays at the defect.
On a platform where os.linesep is already "\n" nothing changes, so CI behaviour is
identical.

Reproduction

Windows 11, Python 3.12.10, against unmodified main (704bc95):

$ python -m pytest -p no:randomly tests/nodes/test_build_context.py \
    -k "real_directory_with_skill_md or reads_directory_with_windows_secure_open or \
        stops_at_progressing_shared_deadline or scans_nested_oms_signature or \
        inventories_hidden_file_for_local_analysis" -q --no-header

E       AssertionError: assert '# Reference guide\r\n' == '# Reference guide\n'
E       AssertionError: assert 'print(1)\r\n' == 'print(1)\n'
E       AssertionError: assert {'first.txt': 'first\r\n'} == {'first.txt': 'first\n'}
E       assert '{"mediaType"...d":""}]}}\r\n' == '{"mediaType"...yid":""}]}}\n'
E       AssertionError: assert 'TOKEN=not-reported\r\n' == 'TOKEN=not-reported\n'

With this change, same command:

.....                                                                    [100%]
5 passed, 75 deselected in 5.09s

What must still hold

tests/nodes/test_build_context.py in full — five cases move from failing to passing and
nothing else changes state:

The seven that remain are unrelated to newline handling and are not touched here.

ruff check tests/ — All checks passed!
ruff format --check tests/ — 126 files already formatted

Diff is 1 file changed, 9 insertions(+), 6 deletions(-).