fix: resolve mutable defaults, narrow exception handling, and sys.modules leak

[zurfjereluhmie]

• Mutable default arguments — Replace shared list defaults in DatasetUploader.__init__ with None sentinels to prevent shared-state bugs (B006)
• Overly broad exception handling — Narrow except Exception to except ValueError in uploader skip/append paths, matching Phoenix's actual not-found exception; non-recoverable errors (network, auth) now propagate correctly
• sys.modules leak — Clean up partially-initialized modules from sys.modules when exec_module fails in runner._load_attribute
• contains evaluator docs — Add prominent note clarifying that only the first expected item is checked

Tests

• Added test_non_not_found_error_propagates to verify RuntimeError from Phoenix is not swallowed
• Added test_cleans_sys_modules_on_import_error to verify broken modules are removed from sys.modules
• Updated all existing tests to use ValueError instead of generic Exception for not-found scenarios
• 191 tests pass, 2 skipped (langgraph extra)

Files changed

File	Change
src/evalwire/uploader.py	Mutable defaults fix + narrowed exception catch
src/evalwire/runner.py	sys.modules cleanup on import failure
src/evalwire/evaluators/contains.py	Improved docstring
tests/test_uploader.py	Updated exception types + new propagation test
tests/test_runner.py	New sys.modules cleanup test
tests/test_cli.py	Updated exception type in mock