fix(incremental): add stop_at to find_repo_root / find_project_root (#241)

[azizur100389]

Summary

Two tests in tests/test_incremental.py have been failing on any developer machine whose tmp_path has a git-initialized ancestor:

• TestFindRepoRoot::test_returns_none_without_git
• TestFindProjectRoot::test_falls_back_to_start

This PR adds an optional stop_at: Path | None = None parameter to both find_repo_root() and find_project_root(), uses it in the two flaky tests to bound the walk to tmp_path, and adds 3 new positive regression tests to lock in the new parameter's semantics.

Closes #241.

Root cause

pytest's tmp_path lives under $TMPDIR, which on Windows 11 resolves to C:/Users/<user>/AppData/Local/Temp/pytest-of-<user>/.... If the user has ever run git init anywhere under their home directory (dotfiles repo, chezmoi, yadm, a bare git init ~ — all very common on developer machines), then find_repo_root() walking up from tmp_path/no_git finds the ancestor .git and returns C:/Users/<user> instead of None. The same failure mode affects Linux and macOS users with a git-initialized home directory.

find_repo_root previously had no way to bound its ancestor walk — it always climbed to the filesystem root. This is correct production behavior (a nested script inside a git repo should find its repo) but makes the function un-testable in environments where the tmp path sits inside a git-initialized ancestor.

Fix

Add an optional stop_at: Path | None = None parameter to both find_repo_root() and find_project_root():

def find_repo_root(
    start: Path | None = None,
    stop_at: Path | None = None,
) -> Optional[Path]:
    current = start or Path.cwd()
    while current != current.parent:
        if (current / ".git").exists():
            return current
        if stop_at is not None and current == stop_at:
            return None
        current = current.parent
    ...

Semantics:

• stop_at=None (default): existing behavior, walk all the way to the filesystem root
• stop_at=<path>: the walk examines stop_at for .git and then stops — it never crosses above the boundary
• A .git at exactly the boundary is still found (it's examined before the stop check)

find_project_root forwards stop_at to find_repo_root, so both public APIs gain the same testability handle.

Fully backward-compatible. All 7 production callers pass no argument:

• cli.py (4 call sites)
• tools/refactor_tools.py (1 call site)
• tools/_common.py (1 call site)
• internal call in find_project_root itself

stop_at is also a legitimate production API — callers that want to bound the ancestor walk (multi-repo orchestrators, CI containers with bind-mounted volumes, embedded sandboxes, Docker build contexts) can now do so without resorting to monkeypatching.

Tests updated / added

1. TestFindRepoRoot::test_returns_none_without_git — now passes stop_at=tmp_path (the actual fix for the flakiness). Docstring explains why the bound is needed.
2. TestFindRepoRoot::test_stop_at_prevents_escape_to_outer_git [NEW] — positive regression test: when an outer directory has .git, an unbounded walk finds it but a bounded walk with stop_at=inner returns None. Locks in the new parameter's semantics.
3. TestFindRepoRoot::test_stop_at_finds_git_at_boundary [NEW] — regression test: stop_at must not suppress a .git that lives at the boundary itself. The walk examines stop_at for .git before stopping, so a repo rooted exactly at the boundary is still found.
4. TestFindProjectRoot::test_falls_back_to_start — now passes stop_at=tmp_path and monkeypatches CRG_REPO_ROOT via monkeypatch.delenv(...) so a developer env var can no longer shadow the test expectation.
5. TestFindProjectRoot::test_stop_at_forwarded_to_find_repo_root [NEW] — regression test: find_project_root must forward stop_at to find_repo_root, not silently drop it. Also verifies the fall-through path (return start when the bounded walk finds nothing) still works.

Test results

Stage	Result
Stage 1 — TestFindRepoRoot + TestFindProjectRoot	8/8 passed (up from 6/8 on unchanged main)
Stage 2 — tests/test_incremental.py full	48 passed, 1 pre-existing failure (test_default_uses_repo_subdir — UTF-8 gitignore bug, fixed on the parallel branch in PR #240)
Stage 3 — adjacent tests/test_tools.py (find_project_root is used there)	62 passed, 15 pre-existing Windows teardown errors
Stage 4 — full suite	737 passed, 6 pre-existing Windows failures — all 6 are covered by the parallel #239/#240 fix PR; none are regressions from this change. Once both PRs merge, all 8 pre-existing Windows failures will be resolved.
Stage 5 — ruff check on incremental.py + test_incremental.py	clean

Net baseline effect: this PR independently removes 2 failures (8 → 6). Combined with PR #240, the full Windows baseline goes from 8 failures to 0.

Relationship to #240

PR #240 and this PR are independent and touch disjoint code paths:

• fix: three Windows-only bugs in v2.3.1 (#239) #240 fixes product bugs in get_data_dir() UTF-8, parse_bytes CRLF handling, and the stale FastMCP API in test_main.py.
• This PR adds a new stop_at parameter to find_repo_root / find_project_root and fixes the 2 test environment flakiness.

They can be merged in either order — no conflicts, no shared files.

Why this fix is safe

• stop_at default is None, preserving existing behavior exactly.
• All 7 production callers pass no argument for stop_at.
• A new regression test (test_stop_at_finds_git_at_boundary) locks in the "inclusive of boundary" semantics so a future refactor can't accidentally change the behavior.
• No product API removed or renamed.