fix(git): fix canonical_root for worktrees and subdirectory projects

[anivaryam]

Summary

Fixes #659 — detect_changes returns empty impacted_symbols for projects indexed inside git worktrees or as repo subdirectories because canonical_root was computed incorrectly.

Root cause: git rev-parse --git-common-dir outputs a path relative to the directory passed via -C (the indexed path, input_path), not to worktree_root. Two errors in derive_canonical_root():

1. The relative git_common_dir (e.g. "../.git") was joined with worktree_root instead of input_path. For a project at workspace/scripts/, worktree_root is workspace/ but the relative path is relative to scripts/ — joining it with workspace/ produced workspace/../.git, which strips to workspace/...
2. Even if joined correctly (scripts/../.git), stripping the /.git suffix leaves scripts/.. — a path with unresolved .. components. Without normalization this is stored as-is and never matches the project's stored root path.

Fix (src/git/git_context.c):

• Pass input_path (the original indexed directory) as the join base for relative git_common_dir paths.
• Call realpath() (POSIX) / _fullpath() (Windows) on the joined .git path before stripping the suffix. The .git directory always exists for a valid repository, so the call succeeds and collapses any .. components.

Tests (tests/test_git_context.c):
Three new tests in suite_git_context:

• canonical_root_repo_root — normal repo indexed from its root (regression guard)
• canonical_root_subdir — project indexed from a subdirectory of the repo (issue detect_changes: impacted_symbols always empty for projects inside git worktrees (canonical_root miscalculated) #659)
• canonical_root_linked_worktree — project in a linked git worktree (issue detect_changes: impacted_symbols always empty for projects inside git worktrees (canonical_root miscalculated) #659 primary case)

Test plan

• canonical_root_repo_root passes
• canonical_root_subdir passes (was failing before fix)
• canonical_root_linked_worktree passes (was failing before fix)
• Full make test suite has no regressions

[DeusData]

Huge thanks for opening this PR and for the work you put into it.

The maintainer shop is currently full, so this may sit for a bit before it gets a proper review. We will come back to this as soon as possible with real feedback; I wanted to make sure it did not sit unacknowledged in the meantime.

[DeusData]

Thanks, this looks focused on #659. Before merge, please add DCO sign-off to the non-merge commit. Also, because the regression test shells out to git, please confirm the command inputs are fixed/test-controlled and note expected CI platform coverage.

[DeusData]

Thank you @anivaryam — this is a correct and well-diagnosed fix. You nailed the root cause: git rev-parse --git-common-dir is relative to the -C dir, and the old code joined it against worktree_root and stripped /.git textually, leaving an unresolved .. that never matched the stored root.

I've taken your git_context.c fix verbatim into a distilled follow-up, #749 (354befe), co-authored with you. The only change is the test suite: your canonical_root_repo_root / canonical_root_subdir tests called make_git_repo() (which shells out to git) unconditionally, so the Windows CI leg failed with "failed to init git repo" (ci-ok red) — while the linked-worktree test was already SKIPed there. In the follow-up all three git-based tests SKIP_PLATFORM on Windows (matching that existing convention), so the suite is green cross-platform, and the linked-worktree case is kept as the genuine reproduce-first guard (it's the one that reliably produces a relative --git-common-dir and fails on the unfixed code).

Closing this as superseded by the distilled PR, with full credit to you. Really appreciate the careful work — the fix ships in the next release. 🙏