fix: seq2seq decode returns primary choice (strip alternates)

src/interscript/ml/model.py

@@ -10,6 +10,13 @@
 from interscript.ml.loader import load_manifest, verify_and_read
 
 
+def first_alternates(text: str) -> str:
+    """Seq2seq output contract: models trained on multi-reference
+    corpora emit '/'-separated alternates per token; the runtime
+    returns the primary choice only."""
+    return " ".join(part.split("/", 1)[0] for part in text.split())
+
+
 class Model:
     """A loaded, checksum-verified IMF v1 model.
 
@@ -61,7 +68,7 @@ def id(self) -> str:
 
     def translate(self, text: str, max_len: int = 256) -> str:
         token_ids = self.generate(text, max_len=max_len)
-        return decode(token_ids)
+        return first_alternates(decode(token_ids))
 
     def generate(self, text: str, max_len: int = 256) -> list[int]:
         ids = np.array([encode(text)], dtype=np.int64)

tests/test_first_alternates.py

@@ -0,0 +1,15 @@
+from interscript.ml.model import first_alternates
+
+
+def test_first_alternates_keeps_primary_choice_per_token():
+    assert first_alternates("وِكَالَةُ/وِكَالَةُ اَلْأَرْصَادِ") == "وِكَالَةُ اَلْأَرْصَادِ"
+
+
+def test_first_alternates_plain_text_untouched():
+    assert first_alternates("مرحبا بالعالم") == "مرحبا بالعالم"
+
+
+def test_first_alternates_edges():
+    assert first_alternates("") == ""
+    assert first_alternates("a/b c/d") == "a c"
+    assert first_alternates("مَارْغِرِيتْ/مَارْغِرِيتْ/مَارْغِرِيتُ اَلثَّانِيَةُ") == "مَارْغِرِيتْ اَلثَّانِيَةُ"