Replace 46 blind `except Exception` handlers with the exception types the
guarded code can actually raise, so an unforeseen failure surfaces instead
of being silently swallowed. The types were derived per site from the try
block: file reads get (OSError, UnicodeDecodeError), TOML and YAML parsing
add their decode errors, int() guards get ValueError, imports get
ImportError, subprocess calls get (OSError, subprocess.SubprocessError),
git queries get GitRunError, and importlib.metadata lookups get
PackageNotFoundError.
Three handlers stay blind, now with a stated reason: the per-repository
loops in install and update are batch boundaries where one failing
repository must never abort the run.
Three imports come along because the narrowing needs them: GitRunError in
the changelog and version commands, and http.client in token validation --
urllib raises HTTPException, which is not an OSError subclass and would
otherwise have escaped.
Add tests/integration/test_error_path_degradation.py, which drives 44 of
the 46 handlers into their except branch and asserts the documented
degradation. Faults are injected for real rather than mocked wherever
possible: undecodable bytes, malformed YAML, digit strings past
sys.get_int_max_str_digits(), a real HTTPError with an unreadable body.
That approach paid for itself immediately -- tomllib.load() decodes
internally and raises UnicodeDecodeError, which was missing from the
pyproject handlers and would have turned a malformed pyproject.toml into
a traceback.
The two handlers left uncovered are the branch-close path in _release_impl,
reachable only through a full release, and the jinja2 import guard.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
7764a2946c