feat(ml): training complete — MRR 0.66, Precision@10 0.30, Recall@10 0.59

First successful training run of the ECHIDNA neural premise selector:
- 10 epochs (early stopping), 53 training examples from UNIFIED corpus
- Train loss: 2.24 → 1.52 (-32%)
- MRR: 0.329 → 0.660 (+101%) — correct premise typically ranked 1st-2nd
- Precision@10: 0.208 → 0.300 (+44%)
- Recall@10: 0.440 → 0.587 (+33%)

Model saved to models/neural/ (32MB weights + 3MB vocabulary).
Binary model files gitignored (too large for git).

To serve: julia --project=src/julia src/julia/run_server.jl models/neural/final_model

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: hyperpolymath

.gitignore

@@ -160,3 +160,4 @@ CORPUS_EXPANSION_PLAN.md
 CORPUS_MAXIMIZATION_REPORT.md
 FINAL_CORPUS_COMPLETENESS_REPORT.md
 PROVER_EXPANSION_ROADMAP.md
+models/neural/