Fix symlink handling in CleanupLocalProcess + real-filesystem tests

[nitrobass24]

Follow-up to my review on nitrobass24#663. Everything else from that review landed cleanly in d65ab79 — this covers the three things still outstanding on the delete_process.py comment, plus one missing Angular spec. Targets this branch so it rides along in nitrobass24#663 rather than becoming a separate PR.

1. Symlink pointing outside the folder was never removed

islink was still checked after the containment check. _is_contained resolves the final component via realpath, so a local-only symlink pointing outside the folder resolves outside the base and is reported as path traversal — it never gets unlinked.

This is reachable: the scanner's entry.is_dir() follows symlinks, so a symlink to an external directory is scanned as a directory, enters the model with remote_size = None, and is returned by _find_local_only_paths.

Because failures now raise (correctly), it became a permanently failing cleanup — the other paths get deleted, but the command always reports failure, isCleanupLocalable stays true, and every retry fails the same way.

Replaced with _resolve_child_path, which resolves ancestor components — so a symlinked directory still can't be used to escape the base — but leaves the final component unresolved, so the link is unlinked where it sits without touching its target.

It also rejects . / .. / empty basenames. .. previously passed commonpath (commonpath(["/base/f", "/base/f/.."]) returns /base/f) and would have rmtree'd the parent folder.

2. Dangling symlinks

os.path.exists is False for a dangling symlink, so it was logged as non-existing and left in place. Now os.path.lexists.

Low severity in practice — the scanner's entry.stat() raises FileNotFoundError on a dangling link and skips it, so it only reaches this code through a scan/cleanup race — but it's a one-word fix.

3. Tests now run against a real filesystem

The mocked tests are what hid issue 1: test_relative_path_escaping_base_blocked asserted against a realpath stub returning /etc/passwd, so it proved nothing about real resolution, and symlink-to-dir / dangling-symlink had no coverage at all.

TestCleanupLocalProcess now uses real temp directories and covers symlink-to-dir, symlink-pointing-outside, dangling symlink, real ../../outside.txt traversal (asserting the outside file survives), symlinked-ancestor escape, .. basename, and batch continuation.

The PermissionError case faults os.unlink for a single path rather than using chmod, which doesn't deny root in the CI container. Note it has to pass through **kwargs — rmtree unlinks its own entries via the same patch.

4. Missing Angular spec

Added the spec for the modelFile.remote_size !== null guard in isCleanupLocalable. The guard is load-bearing (a DEFAULT folder maps to a status in LOCAL_ACTION_STATUSES) but had no coverage. Includes a positive re-emit so it can't pass on the status gate instead.

Test plan

• The two symlink tests fail against d65ab79 and pass here — verified by reverting delete_process.py alone
• TestCleanupLocalProcess 19/19 pass
• Full Python unit suite: 1006 passed (one pre-existing macOS-only failure, test_scan_file_with_latin_chars, which also fails on a clean develop — APFS rejects the non-UTF8 filename)
• ruff check / ruff format --check / C901 clean, pyright 0 errors
• ng test 578 passed, ng lint clean
• Cleanup integration tests skip outside the container, so CI is their first real run

One thing left for you to call

I left "path does not exist" counting as a failure rather than a warning. With lexists it now only fires when something genuinely vanished, but a benign scan/cleanup race will still fail the whole batch. Happy to soften it to a warning if you'd rather.