Candidate-evaluation trap: selected-result wrappers can reject genuine improvements

While preparing COR-3791 Reference LoRA evaluation, I found that the historical selected-adapter service-replay wrapper asserts the exact selected 34/42 confusion matrix. Reusing it for a new candidate would therefore fail even if the candidate improves. Candidate evaluation should instead call the generic, hash-pinned qwen_reference_repeatability path with model_role=reference_candidate, while independently binding the unchanged holdout, runtime, base model, and adapter hashes. Provenance: e3-solutions/ender-inference, training/turn_detection/qwen_reference_repeatability.py, COR-3791; observed 2026-09-04.