learn: trained classifier arm (logistic and gradient boosting on embeddings + signals, per-fold leak screening); fit-eval record

On the 501 exp-005 units, repo-held-out folds: logistic pooled within-pair AUC
0.721 vs the LLM programs 0.762 (diff -0.041, CI [-0.116, 0.032], a tie);
gradient boosting 0.597 (significantly worse). Per family all differences span
0. Implemented with Codex; fit run by the coordinator (4.2 h on the VM).

Co-Authored-By: Claude Opus 5.5 <[email protected]>

pyproject.toml

@@ -13,6 +13,7 @@ requires-python = ">=3.11"
 dependencies = ["pyyaml>=6.0"]
 
 [project.optional-dependencies]
+train = ["numpy>=1.26,<3", "scikit-learn>=1.4,<2"]
 # The documentation site (mkdocs.yml, docs/). Mermaid renders client-side through Material's superfences fence.
 docs = [
     "mkdocs>=1.6,<2",
@@ -68,7 +69,7 @@ warn_return_any = true
 warn_unused_configs = true
 
 [[tool.mypy.overrides]]
-module = "tree_sitter"
+module = ["tree_sitter", "sklearn.*", "numpy"]
 ignore_missing_imports = true
 
 # boto3 is an optional extra without inline types, absent in the type-check environment (plane/memory.py imports