Add frozen shared corpus builder

[DearKarl]

Research question

Can the sealed synthetic-data smoke be scaled into a reproducible, role-isolated 40,000-record shared corpus without weakening determinism or overlap safeguards?

Changed variable

This change adds the frozen shared-corpus-2026.08.12 configuration, schema, coverage matrix, streaming generator, scalable overlap audit, and independent verification entry point. It does not add model training.

Data and license boundary

All records are project-authored synthetic content with rebuild-only runtime assets. Terminal benchmark and Customer50 inputs are explicitly absent. Train/development/holdout family isolation and exact/near duplicate checks are hard gates.

Acceptance

• 10,000 records in each of four categories
• 28,000/8,000/4,000 train/development/holdout
• 10,000 same-role counterfactual pairs
• deterministic semantic and normalized-pixel replay
• exact and supported near-overlap audit PASS
• independent SHA256 verification

Validation

pytest -q — 43 passed.