ChemPropModel.aggregation is ignored when from_foundation is set, and "attentive" isn't a valid value

[smcolby]

Summary

ChemPropModel.build() hardcodes MeanAggregation whenever from_foundation is set, silently ignoring the configured aggregation field. Separately, validate_aggregation only accepts "mean" or "norm", even though chemprop.nn also provides AttentiveAggregation.

Where

openadmet/models/architecture/chemprop.py:

• build(), from_foundation branch (~line 649): always constructs aggr = nn.MeanAggregation(), regardless of self.aggregation.
• build(), else branch (~lines 657-660): this is the only place self.aggregation is actually consulted, and only reachable when from_foundation is falsy.
• validate_aggregation (~lines 524-542): raises unless value is "mean" or "norm"; chemprop.nn.AttentiveAggregation exists but is unreachable through this field.

Why this is a bug, not intentional behavior

Aggregation pools the message-passing body's already-computed per-atom hidden states into a molecule-level embedding. MeanAggregation/NormAggregation have no learned parameters, and AttentiveAggregation adds a small learned layer on top of the aggregation step, independent of the message-passing body's own weights. There's no architectural reason a foundation-initialized encoder (CheMeleon or a custom checkpoint) can't use a different aggregation strategy — the current hardcoding looks like it was never revisited after from_foundation was added, not a deliberate constraint.

The logger.warning at that call site ("Using a foundation model overrides settings for depth, message_hidden_dim, messages, and aggregation") documents the current behavior accurately, but it means the aggregation recipe field is a silent no-op for anyone using from_foundation.

Suggested fix

• In the from_foundation branch, build aggr from self.aggregation the same way the else branch does, instead of hardcoding MeanAggregation.
• Extend validate_aggregation to accept "attentive", and extend the aggregation-class dispatch to include nn.AttentiveAggregation.
• Update the warning message to drop "and aggregation" once it's no longer overridden.

Repro

Any anvil recipe with model.params.from_foundation: chemeleon and model.params.aggregation: norm (or any non-"mean" value) trains with mean aggregation regardless, with no error or additional warning tied to the ignored value.