Add human-labeled failure taxonomy

[Danny1218]

Create a small human review layer for sampled paired failures.

Goal:

• Add semantic texture to automated exact-match, regex, and code-scoring failures.

Acceptance criteria:

• Define a compact taxonomy for failure types.
• Sample paired FP16-correct/quant-wrong examples from existing artifacts.
• Store labels separately from raw generations and document inter-review assumptions if multiple reviewers are used.