Use object_id from invariant_receipt as the lookup key instead of equation_id from equation_record. This aligns the predictions JSON with the manifold findMatrix calls. Build passes 3307 jobs.
- generate_predictions.py: reads 278 equations from classifier receipt, builds 8×8 strand adjacency matrices, deduplicates to 250 by equation_id - build_pist_matrices_250.py: emit each matrix as a separate Lean def (match-based findMatrix) to avoid kernel term-size blowup - rrc_pist_predictions_250_v1.json regenerated in Research Stack shared-data/ - Shakeout finding: text-derived matrices have spectral radius < 2.0, so classifyExact returns none (expected — 250 non-zero matrices, weak features)
