Error messages, docstrings, CI and test cleanups found in the third pre-4.0 audit #618

This was generated by AI during a third pre-4.0 audit of main at f6e343d.

Non-breaking improvements found in the same audit. None block 4.0.0.

Error messages

1. Subclassing a provider fails at first resolve with a bare TypeError: no compiled resolver for provider type MyF, and validate() says nothing. Reject at definition time (__init_subclass__) or raise a ModernDIError. tests/test_container.py:224 subclasses Factory.
2. DuplicateProviderTypeError does not name the two clashing providers, though ProvidersRegistry._add has both; add them with their definition sites. Its second hint ("pass dependencies via kwargs") does not fix a duplicate.
3. FinalizerError prints [ValueError('boom'), ...] without saying whose finalizer failed. add_note each inner error with the instance type, and use the one-line summary style of ValidationFailedError.

Readability

4. No docstrings on Factory, Alias, Group, AbstractProvider, CacheSettings, Group.get_providers, get_named_providers; IDE hover shows nothing.
5. resolver_compiler.py: [*edges][arg_lines[exc.__traceback__.tb_lineno]] needs re-reading; rename arg_lines to arg_index_by_line and spell out f, a, cp. wiring.find_dep_provider and Edge.dep keep the "dep" abbreviation feat!: make exception attributes consistent and ValidationFailedError an ExceptionGroup (#604) #607 removed from the exceptions.

CI and tooling

6. Action versions: actions/cache@v4 (Node 20 deprecation warning in the benchmarks job), setup-uv v8.2.0, actions/checkout v6. Add dependabot for GitHub Actions.
7. Benchmarks: the C-series never runs in CI; a --benchmark-disable smoke run (33 tests, about 1 s) in scheduled.yml would catch breakage. The G-series baseline cache key ignores the Python version. Nothing alerts on pushes to main (comment-on-alert: false).
8. Run the Python code blocks in docs/ and README (190 blocks): ast.parse all, execute those importing only modern-di.
9. --cov-report xml writes a file nothing uploads. fix = true with unsafe-fixes = true makes a bare ruff check apply unsafe fixes. benchmarks/comparative is neither linted nor type-checked.

Tests

10. test_container.py:755 assert gc.collect() == 0 depends on global state and adds nothing. test_types_parser.py:252 asserts only provider is not None. ty flags tests/test_types_parser.py:402 on 3.12+.
11. test_container.py (1200 lines) and test_context_provider.py (1287 lines) mix topics; move export checks to test_packaging.py and context-leak tests next to the container tests.
12. Consider a scheduled mutmut run over cache.py, container.py, resolver_compiler.py.