tighten multi-file experiment validation

[CKwin26]

Summary

• make deep experiment validation catch undefined helper calls such as missing local utility functions
• avoid flagging legitimate subclasses that inherit __init__ and only implement one core method
• detect duplicate algorithm-class definitions across multiple generated project files

Why

Generated experiment projects are increasingly multi-file. The validator was missing two important cases:

• it could miss obvious runtime breakage from undefined helper calls
• it could misclassify valid inherited subclasses as low-quality classes

At the same time, duplicated class definitions across files like main.py and models.py are a real failure mode for generated projects and should be surfaced explicitly.

Changes

• add undefined-call checking to deep_validate_files(...)
• relax the "too few methods" class-quality heuristic for classes with explicit base classes
• add a dedicated warning when the same non-trivial algorithm/helper class is defined in multiple files
• avoid emitting the less-useful duplicate-body warning for same-name classes once the cross-file duplicate-class warning is raised

Validation

• python -m pytest -q tests/test_rc_executor.py -k "deep_validation_detects_undefined_helper_calls or deep_validation_allows_inherited_single_core_method_subclass or deep_validation_detects_duplicate_algorithm_classes_across_files"
• python3 -m compileall researchclaw/experiment/validator.py tests/test_rc_executor.py