fix(classifier): majority-vote key, partial-load guard, sparse label test

scripts/classifier_adapters.py

@@ -399,7 +399,7 @@ class EmbeddingKNNAdapter(ClassifierAdapter):
     classify(subject, body):
       Embeds the input email, computes cosine similarity against all stored exemplar
       vectors, and majority-votes the top-k labels (default k=3). Tie-break: label
-      with the highest mean similarity score among tied vote counts wins.
+      with the highest total similarity score among tied vote counts wins.
 
     unload():
       Releases the cf-orch allocation (DELETE .../allocations/{id}) and clears state.
@@ -489,10 +489,14 @@ class EmbeddingKNNAdapter(ClassifierAdapter):
             orch_url_used = ""
         self._node_url = node_url
         self._orch_url_used = orch_url_used
-        embeddings: dict[str, list[list[float]]] = {}
-        for label, texts in self._exemplar_texts.items():
-            embeddings[label] = self._embed(node_url, texts)
-        self._exemplar_embeddings = embeddings
+        try:
+            embeddings: dict[str, list[list[float]]] = {}
+            for label, texts in self._exemplar_texts.items():
+                embeddings[label] = self._embed(node_url, texts)
+            self._exemplar_embeddings = embeddings
+        except Exception:
+            self.unload()
+            raise
 
     def unload(self) -> None:
         if self._allocation_id and self._orch_url_used:
@@ -523,4 +527,7 @@ class EmbeddingKNNAdapter(ClassifierAdapter):
         votes: dict[str, list[float]] = {}
         for score, label in top_k:
             votes.setdefault(label, []).append(score)
-        return max(votes, key=lambda lbl: sum(votes[lbl]))
+        return max(
+            votes,
+            key=lambda lbl: (len(votes[lbl]), sum(votes[lbl])),
+        )

tests/test_classifier_adapters.py

@@ -592,13 +592,13 @@ def test_classify_tiebreak_by_mean_score():
 
 def test_classify_sparse_label_can_win():
     from unittest.mock import patch
-    # "hired" has only 1 exemplar; query vector is closest to it
+    # "hired" has only 1 exemplar; with k=1, the single closest match wins
     adapter = _adapter_with_embeddings({
         "rejected": [[0.0, 0.0, 1.0], [0.0, 0.1, 0.9]],
         "hired":    [[1.0, 0.0, 0.0]],
-    }, k=3)
+    }, k=1)
 
-    # Query [1,0,0] → hired exemplar scores 1.0; rejected exemplars score ~0
+    # Query [1,0,0] → hired exemplar scores 1.0; closest single match wins
     with patch("httpx.post", return_value=_embed_resp([1.0, 0.0, 0.0])):
         result = adapter.classify("Welcome aboard", "Your first day details")
 
@@ -637,3 +637,4 @@ def test_classify_lazy_loads_when_not_loaded():
     assert result == "rejected"
     assert any("/allocate" in u for u in post_urls), "lazy load must call allocate"
     assert adapter._exemplar_embeddings != {}
+    assert adapter._node_url == "http://navi:11434"