Merged iron_eval.py (4,533 lines) and comprehensive_eval.py (3,224 lines)
into a unified eval.py (1,999 lines) that achieves:

- 100% tensor coverage: 6,296/6,296 tensors accessed
- 606,202 tests passing (99.9993%)
- All 11 categories verified: alu, arithmetic, boolean, combinational,
control, error_detection, manifest, memory, modular, pattern_recognition,
threshold

Key features from both files:
- TensorRegistry with coverage tracking
- CircuitEvaluator with ~130 test methods
- XOR sublayer access for full adders and multipliers
- Exhaustive 256x256 testing for comparators/adders

TODO: Fix 4 remaining test failures in boolean.xor (2/4) and
arithmetic.halfadder (6/8). These appear to be tensor weight precision
issues in actual circuit execution, not missing coverage. Need to
investigate whether the loaded weights match expected threshold logic
values (e.g., AND should be w=[1,1], b=-2).