Transferability study PR edits

[plstcharles-saifh]

Summary of improvements I did while doing my own review of lawzero-org#2:

• Isolated transferability smoke/limited outputs so validation runs cannot contaminate full-sweep resume state;
• Fixed CueFlip correctness issues around GSM8K operation-flip cache validation, HumanEval cue-uptake scoring, and analysis denominators;
• Added CueFlip rendered_chat prompt mode, which builds neutral system/user messages, renders them locally with the HF chat template, and still calls /v1/completions (should be more aligned with Qwen3 usage expectations);
• Added regression tests for the cache validator, HumanEval scoring, CueFlip analysis, prompt rendering, and prompt-format resume filtering.

Next step on my end: test the monster, and take a look at the eval results...