Improve Jev synthetic data auditing and selection provenance

[sileod]

Why

The synthetic Jev pipeline already preserves raw generation/Jev responses, but two research-validity gaps remain:

1. scenario quality is checked before Jev labeling, while confident teacher mistakes are not independently surfaced;
2. when downsampling with --n-target, confidence-bucket selection can make the retained distribution conditional on Jev's own uncertainty.

What this PR changes

• adds an optional independent post-Jev teacher audit:

  • auditor sees only the generated state/questions, never Jev probabilities;
  • it independently answers Choice / Score / Noul questions with a confidence;
  • local code compares its answer against the Jev distribution;
  • high-confidence Jev/auditor disagreements are explicitly flagged;
  • raw auditor responses are content-addressed and cached;
  • filtering disagreements is opt-in, while the default keeps flagged rows auditable.

• adds a teacher-independent selection reservoir:

  • a configurable fraction of a downsampled corpus is chosen from state IDs only;
  • the rest follows the requested Jev ambiguity buckets;
  • every selected bundle records selection_reason and selection_bucket.

• exports provenance on flat training rows:

  • target origin / returned teacher model;
  • Jev max probability and entropy;
  • selection arm/bucket;
  • independent-auditor agreement/confidence/disagreement flag.

• adds an example audited config:

  • DeepSeek generation via Albert;
  • pinned Jev 1.13 targets via OpenRouter;
  • independent OpenAI/Luna teacher audit.

Intended research use

This does not replace Jev targets with another LLM's judgments. The independent model is a diagnostic layer. The goal is to make confident teacher failures measurable and to enable matched ablations between unfiltered and confidence-curated synthetic data.

Tests

Adds offline tests for:

• confident teacher-disagreement detection;
• audit caching;
• teacher-independence of the reservoir;
• audited-config provider separation.

CI is expected to run the full suite on Python 3.10 and 3.13.