test(grep): compare _grep_python against a real grep (closes #63)

[dchaudhari7177]

Closes #63.

What

Eleven differential cases over one fixture corpus — text, nested dirs, a NUL-bearing binary, files in and out of a glob — requiring _grep_python and the system grep to agree on: plain regex, -i, glob filter, no-match, binary skipping, regex metacharacters (a[0-9]b, ^needle, line$, n..dle), and a single-file target.

Output from both sides is reduced to {(relative posix path, lineno, text)} before comparing. grep joins its argument to each entry with / while os.walk uses os.sep, so the raw strings differ on Windows even when the matches are identical — normalising keeps the assertion about which lines matched, which is the thing that can actually drift.

Two tests assert the concrete file set as well as agreement, so "both implementations agree on nothing" can't pass.

One deliberate departure from the issue

The acceptance criterion says "passes on macOS/Linux, skips cleanly on Windows". I gated on grep being present instead of on sys.platform, for two reasons:

• Git for Windows ships GNU grep, so a platform gate skips exactly where _grep_python is the default code path and divergence matters most.
• Gating by platform would have meant I couldn't run these at all — I'm on Windows. As written they run green here (11 passed), and they skip cleanly anywhere grep genuinely is absent, which is the same protection.

Say the word if you'd rather have the platform gate and I'll swap it.

It already found something

Given a single file rather than a directory, GNU grep drops the filename prefix:

grep:          1:needle at the top
_grep_python:  C:\...\top.txt:1:needle at the top

So grep(pattern, path="a.txt") returns a different shape depending on whether a grep binary happens to be installed — precisely the class of drift this issue exists to catch.

I pinned it in test_single_file_output_format_diverges rather than fixing it: aligning the two changes the grep tool's output, which is your call rather than a drive-by in a test-only PR. Adding -H to the tool's flags would do it, and I'm happy to send that separately.

Verification

pytest tests/test_grep_differential.py — 11 passed. ruff check clean.

Full suite on this branch: 63 passed, 2 failed — the 2 are test_tools.py::test_grep / ::test_grep_ignore_case, which fail on a Windows checkout of main for an unrelated reason I've sent as #78. On a Linux runner the whole suite should be green.