Add human validation subset for classifier agreement

[praneethaneelapareddigari]

Add the 100-item human-labeled validation subset described in methodology §5.4, and compute inter-rater agreement (target Cohen's κ ≥ 0.7) between the rule-based classifier and LLM-judge before trusting either at scale.

[praneethaneelapareddigari]

Update: Substantially complete

The 113-record human validation done for #1 satisfies this issue's core goal (target
was ~100 items per methodology.md §5.4) — see docs/human_validation_v1_findings.md
for the full writeup and results/json/human_validation/ for raw judgment data.

Delivered:

• 113 manually-judged records (exceeds the ~100 target)
• Cohen's κ computed overall and per-language
• Disagreement analysis (50 cases, saved to disagreements.json)

Not yet delivered (keeping this open for):

• Inter-rater agreement between the automated LLM-judge and human judgment
  specifically (methodology's original target was κ ≥ 0.7 between rule-based classifier
  and LLM-judge, not human vs. rule-based — we validated the rule-based classifier
  against a human-equivalent judge instead, since the automated judge is still blocked
  on API access per Build LLM-as-judge refusal classifier (hybrid with rule-based) #1)
• Once evaluation/llm_judge_refusal.py runs for real, this issue should be revisited
  to compute human-vs-automated-judge agreement too, which is the originally-scoped
  three-way comparison.