fix(integrations): don't abort uninstall when the manifest file can't be deleted

[jawwad-ali]

Problem

IntegrationManifest.uninstall() (src/specify_cli/integrations/manifest.py) guards every tracked-file delete:

try:
    path.unlink()
except OSError:
    skipped.append(path)
    continue

…but the manifest's own delete is bare:

if remove_manifest and manifest.exists():
    manifest.unlink()          # <-- unguarded

The manifest is deleted last, after every tracked file is already gone. So an undeletable manifest — a read-only file, a directory left at the path, a Windows lock, an ACL — raises out of uninstall() when the destructive work has already happened.

Why this matters more than a lost return value

The caller doesn't just lose (removed, skipped). Everything after the call is skipped (integrations/_install_commands.py:282):

• reassigning the default integration when the removed one was default
• _write_integration_json(...) / _remove_integration_json(...)
• _clear_init_options_for_integration(...)

So the files are gone but integration.json still records the integration as installed — and possibly still as the default. The project is left in a state the CLI can't reach again through uninstall.

Reproduction on current main (2e44ed6)

Leaving a directory at the manifest path (portable, no chmod, no monkeypatch):

manifest saved: True | tracked file: True
manifest path is now a directory: True
uninstall() RAISED PermissionError: [WinError 5] Access is denied: ...test.manifest.json
  tracked file deleted anyway? True   <-- work done, but caller lost (removed, skipped)

After the fix, same harness:

uninstall() -> removed=1 skipped=1  (no crash)

Fix

+13/-1 in one source file — guard it like its sibling:

try:
    manifest.unlink()
except OSError:
    skipped.append(manifest)

The empty-parent cleanup below is deliberately left unconditional rather than moved into an else:. With the manifest still on disk its parent is non-empty, so the first rmdir() raises and breaks immediately — an else: would be behaviourally identical while re-indenting 8 lines for no reason. I mention it so the omission reads as a decision, not an oversight.

Precedent

• The tracked-file path.unlink() guard in this same function is the pattern being mirrored — the _sha256 guard added by merged fix(integrations): guard _sha256 against unreadable managed files #3376 points at it explicitly: "(mirrors the path.unlink() OSError guard below)". fix(integrations): guard _sha256 against unreadable managed files #3376 hardened this function against OSError from managed files but left the manifest unlink bare.
• integrations/kimi/__init__.py:151 already does except OSError: skipped.append(legacy_dir) for an undeletable legacy directory.

Anticipating one review question

The CLI renders skipped under a header along the lines of "N modified file(s) were preserved", and an undeletable manifest isn't "modified". I still think skipped is right: the docstring commits only to `(removed, skipped)` — absolute paths, and kimi already puts undeletable directories there. The alternative — a third return value — would be a signature change to a public method for a rare I/O error. Happy to adjust the CLI wording in a follow-up if you'd rather the header were broader.

Verification

• Fail-before / pass-after: the new test raises PermissionError: [WinError 5] on unpatched src and passes with the fix.
• tests/integrations/test_manifest.py → 48 passed, 6 failed; the same 6 fail on clean main (2e44ed6) with 47 passed — all six are Windows symlink-privilege tests (test_rejects_symlink_target, test_symlink_removed_with_force, …), unrelated to this change. Delta is exactly my one new passing test.
• tests/integrations/ → 2253 passed, 19 skipped; failures identical to baseline.
• uvx ruff@0.15.0 check → clean

The new test needs no chmod and no monkeypatch: Path.unlink() on a directory raises IsADirectoryError on Linux and PermissionError on Windows/macOS — both OSError subclasses, so it exercises the guard on every CI platform.

---

Written with assistance from Claude Code. Bug found, reproduced, and verified by me on current main; every number above is from output I ran.