[mahdi-shafiei]

Summary

 primitives.py
-  NUM_BOND_FEATURES = 4
+  DEFAULT_BOND_FEATURES = 4          (exported beside DEFAULT_ATOM_FEATURES)
 MolecularPropertyConfig, ADMETConfig, MACCSKeysConfig, CircularFingerprintConfig,
 MolecularFingerprintConfig, MessagePassingLayer, StackedMessagePassing
-  in_features = 4, num_edge_features = 4
+  in_features = DEFAULT_ATOM_FEATURES, num_edge_features = DEFAULT_BOND_FEATURES
 AttentiveFPConfig
-  in_features = 39, edge_dim = 10
+  in_features = DEFAULT_ATOM_FEATURES, edge_dim = DEFAULT_BOND_FEATURES

Closes DB-078. Every drug-discovery operator's default input width is now the graph builder's width, written as its symbol, the way chemprop derives DEFAULT_ATOM_FDIM from its featurizer (chemprop/conf.py, constructor default in nn/message_passing/base.py). Before, the defaults were 4 and 4, so a default config raised a shape error on any smiles_to_graph molecule, and several docstrings documented 34 for the same field.

Two commits:

1. A rename, NUM_BOND_FEATURES to DEFAULT_BOND_FEATURES, with no behavior change.
2. The defaults, their tests, and the callers, fixtures and docs that relied on 4.

This is the foundation for the AttentiveFP featurization PR: once the defaults follow the featurizer, that change touches no config.

Evidence

Before: tests/operators/drug_discovery/test_default_widths.py, 13 of 14 failed. Every default config was 4 (or 39) against the featurizer's 34, and every default operator raised dot_general requires contracting dimensions to have the same shape, got (34,) and (4,) on aspirin.

After: 14 of 14 pass. Each config default equals the symbol, and each operator and both message-passing modules, built with defaults, run on aspirin to finite outputs.

A first draft asserted a num_edge_features attribute that StackedMessagePassing doesn't store. It was rewritten to specify behavior: the module runs on a real molecule, which fails on the old defaults and passes on the new ones.

Fixtures: molecular_graph and batch_molecular_graphs hand-encoded "ethanol" with a 4-feature one-hot (C, N, O, other) matching the old default, not the graph builder. They now come from smiles_to_graph("CCO") and batch_smiles_to_graphs(["CCO", "CC"]). The JIT, gradient and vmap tests in test_drug_discovery.py (jax.jit, nnx.grad, jax.grad, jax.vmap) therefore run on real graphs at the default widths: 56 of 56 pass.

Gate: soft ops, drug discovery, the DTI and MolNet benchmark tests, and integration: 637 passed. pre-commit run --all-files passes all 24 hooks, and mkdocs build --strict passes.

Merge Danger

Door: two-way.

Blast Radius: default widths.

• Default construction: code that built an operator with default widths gets a model sized for the graph builder's 34 and 4 features, where it used to get 4 and 4. That code could only have run on hand-built 4-feature inputs before. AttentiveFP's default goes from 39/10 to the builder's width.
• Speed: no traced code changes. Models built with explicit widths compute exactly as before.
• Siblings: kinase-challenge passes its own NODE_FEATURES and EDGE_FEATURES to StackedMessagePassing (multitask.py:98-99), and receptifex uses only similarity; neither is affected.
• Docs: the user guide's tables now name the symbols, and its examples and scripts/_advanced_plots.py stop hand-passing widths.