Classification Provenance

Read-only diagnostic: entities whose industry_sector was stamped by a given classification method — default 'llm_classify', the paid LLM second pass. Built for vendor bake-offs: pull the exact records a prior LLM model already classified, replay them through a candidate model, and measure agreement against what's on file — instead of eyeballing a fresh sample with no reference answer. Modeled on agents/services/llm_distill.py's _pull_llm_wins query (same join, same method filter).

Caveat: entity_attribute.method records WHICH STAGE produced a result ('llm_classify' vs 'classification_service'), not which specific model version was live when it ran — that isn't persisted per-row. This identifies "the LLM stage decided this", not a pinned model ID.

No vendor spend, no writes — gated to ADMIN/READ_ONLY like every other full-fidelity read (see backend/auth.py::require_export_key).