Document fine_tune NotImplementedError on Sei + LegNet

The two oracles raise ``NotImplementedError`` for ``fine_tune()`` — a
deliberate design decision for Sei (21,907-class head is tied to
training vocabulary) and LegNet (single-output MPRA head trained on
one cell type). Replaced the stale ``# TODO: for now we decided not
to implement`` comment with a real docstring that explains *why* and
points users at AlphaGenome / Borzoi for track-adaptive workflows.

Behavior unchanged — same exception, clearer error message +
docstring.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: lucapinello

chorus/oracles/legnet.py

@@ -312,9 +312,16 @@ def _predict_direct(self,
         return preds
 
     def fine_tune(self, tracks: List[OraclePredictionTrack], track_names: List[str], **kwargs) -> None:
-        """Fine-tune LegNet on new tracks."""
-        # TODO: for now we decided not to implement this functionality
-        raise NotImplementedError("LegNet fine-tuning not yet implemented")
+        """Fine-tuning is not supported for LegNet.
+
+        LegNet's single-output MPRA head is trained on a specific cell
+        type (HepG2) and assay. Fine-tuning on arbitrary tracks would
+        require retraining from scratch — outside Chorus's scope.
+        """
+        raise NotImplementedError(
+            "LegNet fine-tuning is not supported. Train a new LegNet "
+            "model externally if you need a different cell type / assay."
+        )
 
     def _get_context_size(self) -> int:
         """Return the required context size for the model."""

chorus/oracles/sei.py

@@ -492,9 +492,17 @@ def _predict_direct(self,
         return selected_preds, selected_classes 
 
     def fine_tune(self, tracks: List[Track], track_names: List[str], **kwargs) -> None:
-        """Fine-tune Sei on new tracks."""
-        # TODO: for now we decided not to implement this functionality
-        raise NotImplementedError("Sei fine-tuning not yet implemented")
+        """Fine-tuning is not supported for Sei.
+
+        Sei's 21,907-class classification head is tied to its training
+        vocabulary; fine-tuning on user-supplied tracks would require
+        re-engineering the head. Use AlphaGenome or Borzoi for
+        workflows that need on-the-fly track adaptation.
+        """
+        raise NotImplementedError(
+            "Sei fine-tuning is not supported. Use AlphaGenome or "
+            "Borzoi for workflows that need on-the-fly track adaptation."
+        )
 
     def _get_context_size(self) -> int:
         """Return the required context size for the model."""