GitIgnoreSpec re-includes files under an excluded directory (violates git's parent-exclusion rule)

[eeshsaxena]

GitIgnoreSpec re-includes a file through a negation pattern even when a parent directory is excluded, which git forbids ("It is not possible to re-include a file if a parent directory of that file is excluded").

Repro (pathspec 1.1.1)

from pathspec import GitIgnoreSpec

spec = GitIgnoreSpec.from_lines(["build", "!keep.log"])
spec.match_file("build/keep.log")   # -> False (treated as NOT ignored)

build/keep.log should be ignored: build excludes the whole build directory, and !keep.log cannot re-include a file whose parent directory is excluded.

Verified against real git

With .gitignore:

build
!keep.log

and a real build/keep.log on disk:

$ git add build/keep.log
The following paths are ignored by one of your .gitignore files:
build

$ git status --porcelain --ignored
!! build/

git treats build/keep.log as ignored (the entire build/ is excluded); git check-ignore build/keep.log agrees.

The distinction it misses

The bug is the difference between excluding a directory and excluding a directory's contents:

.gitignore	path	git	GitIgnoreSpec.match_file
build + !keep.log	build/keep.log	ignored	not ignored ❌
a + !keep.log	a/keep.log	ignored	not ignored ❌
build/* + !build/keep.log	build/keep.log	re-included	re-included ✅

build/* excludes only the contents, so re-inclusion is allowed and pathspec gets it right. build excludes the directory itself, so re-inclusion must fail.

Root cause (as far as I can tell)

GitIgnoreSpec matching is flat: match_file matches the full path against every pattern and takes the last match, so a file-level negation (!keep.log, no directory marker) overrides a directory-level exclusion (build, matched with a directory marker). git evaluates the path hierarchically, so once a directory is excluded nothing beneath it can be re-included. Telling the two cases apart seems to need ancestor-directory awareness rather than a flat last-match.

Environment

• pathspec 1.1.1
• Python 3.14.3
• git 2.55.0

Happy to help with tests or a fix if you have a preferred approach across the backends.