Prioritize character refs over location refs

dumas_h3_longvideos.py

@@ -4486,10 +4486,20 @@ def _matched_reference_slots(text, ref_slots, normalized_refs=None):
     matched = []
     seen = set()
     normalized_slots = normalized_refs if normalized_refs is not None else _normalized_ref_slots(ref_slots)
-    for slot_number, ref in (
-        _slot_refs_for_text(text, normalized_slots)
-        + _named_refs_for_text(text, normalized_slots, kinds=("character", "location"))
-    ):
+    for slot_number, ref in _slot_refs_for_text(text, normalized_slots):
+        if slot_number in seen:
+            continue
+        seen.add(slot_number)
+        matched.append((slot_number, ref))
+    # Character names should win before location names when a beat mentions both.
+    # That keeps the face conditioning in front of the scene conditioning instead
+    # of letting an early location slot dominate the named-ref list.
+    for slot_number, ref in _named_refs_for_text(text, normalized_slots, kinds=("character",)):
+        if slot_number in seen:
+            continue
+        seen.add(slot_number)
+        matched.append((slot_number, ref))
+    for slot_number, ref in _named_refs_for_text(text, normalized_slots, kinds=("location",)):
         if slot_number in seen:
             continue
         seen.add(slot_number)
@@ -4624,7 +4634,15 @@ def resolve_prompt_refs(text, ref_list, include_named=True, normalized_refs=None
     refs = list(tagged_refs)
     seen = {id(ref) for ref in refs}
     if include_named:
-        for _slot_number, ref in _named_refs_for_text(rewritten, normalized_refs, kinds=("character", "location")):
+        # Keep named character refs ahead of location refs so the identity image
+        # is the first named reference the model sees on untagged beats.
+        for _slot_number, ref in _named_refs_for_text(rewritten, normalized_refs, kinds=("character",)):
+            marker = id(ref)
+            if marker in seen:
+                continue
+            seen.add(marker)
+            refs.append(ref)
+        for _slot_number, ref in _named_refs_for_text(rewritten, normalized_refs, kinds=("location",)):
             marker = id(ref)
             if marker in seen:
                 continue

tests/test_dumas_h3_longvideos.py

@@ -479,6 +479,21 @@ class DumasH3LongVideosHelperTests(unittest.TestCase):
         self.assertFalse(shot_tag_driven)
         self.assertEqual(mode_eff, "auto ref2v")
 
+    def test_resolve_prompt_refs_prioritizes_characters_before_locations(self):
+        refs = [
+            {"kind": "location", "image": "img1", "name": "Hangar"},
+            {"kind": "character", "image": "img2", "name": "Mara"},
+        ]
+
+        text, references, dropped = self.module.resolve_prompt_refs(
+            "[Generation 1] Mara waits in the hangar.",
+            refs,
+        )
+
+        self.assertEqual(text, "[Generation 1] Mara waits in the hangar.")
+        self.assertEqual([self.module._reference_image(ref) for ref in references], ["img2", "img1"])
+        self.assertEqual(dropped, [])
+
     def test_shot_references_uses_all_connected_sparse_slots(self):
         refs = [
             None,