Tighten tagged ref routing and anchor presets

dumas_h3_longvideos.py

@@ -4405,15 +4405,20 @@ def _named_refs_for_text(text, ref_slots, kinds=None):
     haystack = str(text or "")
     matched = []
     wanted = {str(k).strip().lower() for k in (kinds or ()) if str(k).strip()}
+    by_name = {}
     for slot_number, ref in enumerate(ref_slots or [], 1):
         if ref is None or _reference_image(ref) is None:
             continue
         if wanted and str(ref.get("kind") or "").strip().lower() not in wanted:
             continue
         for name in _reference_name_keys(ref):
-            if re.search(r"\b" + re.escape(name) + r"\b", haystack, re.I):
-                matched.append((slot_number, ref))
-                break
+            by_name.setdefault(name.lower(), []).append((slot_number, ref, name))
+    for entries in by_name.values():
+        if len(entries) != 1:
+            continue
+        slot_number, ref, name = entries[0]
+        if re.search(r"\b" + re.escape(name) + r"\b", haystack, re.I):
+            matched.append((slot_number, ref))
     return matched
 
 
@@ -4533,7 +4538,7 @@ def resolve_tagged_refs(text, ref_list):
     return out.strip(), [normalized_refs[n - 1] for n in live], dropped
 
 
-def resolve_prompt_refs(text, ref_list):
+def resolve_prompt_refs(text, ref_list, include_named=True):
     """(rewritten text, refs, dropped) for the refs a shot actually carries.
 
     Explicit <Picture N> tags still decide which slot numbers the prompt points at,
@@ -4545,12 +4550,13 @@ def resolve_prompt_refs(text, ref_list):
     rewritten, tagged_refs, dropped = resolve_tagged_refs(text, normalized_refs)
     refs = list(tagged_refs)
     seen = {id(ref) for ref in refs}
-    for _slot_number, ref in _named_refs_for_text(rewritten, normalized_refs, kinds=("character", "location")):
-        marker = id(ref)
-        if marker in seen:
-            continue
-        seen.add(marker)
-        refs.append(ref)
+    if include_named:
+        for _slot_number, ref in _named_refs_for_text(rewritten, normalized_refs, kinds=("character", "location")):
+            marker = id(ref)
+            if marker in seen:
+                continue
+            seen.add(marker)
+            refs.append(ref)
     return rewritten, refs, dropped
 
 
@@ -5999,8 +6005,9 @@ class H3LongVideos:
                                 "for long chains where identity drift matters more than strict per-shot "
                                 "routing. 'where tagged' keeps the old strict behavior, including the "
                                 "first-shot fallback when no tags are found. Tags are renumbered per shot, "
-                                "so <Picture 2> alone still resolves. Character names in the beat can also "
-                                "pull their matching character refs into the real image-conditioning list. "
+                                "so <Picture 2> alone still resolves. In 'auto ref2v', character and "
+                                "location names in the beat can also pull their matching refs into the "
+                                "real image-conditioning list; 'where tagged' does NOT do that. "
                                 "'first shot' / 'every shot' / "
                                 "'every shot + handoff ref' go purely by position. Ignored when no "
                                 "ref_image is connected."}),
@@ -6702,7 +6709,7 @@ class H3LongVideos:
                 for shot_index, gen in enumerate(gens):
                     shot_mode = beat_ref_mode_directive(beats[shot_index] if shot_index < len(beats) else "") or ref_mode
                     if shot_mode in ("where tagged", "auto ref2v") and any_tags_anywhere:
-                        if resolve_prompt_refs(gen, ref_slots)[1]:
+                        if resolve_prompt_refs(gen, ref_slots, include_named=(shot_mode == "auto ref2v"))[1]:
                             on.append(shot_index + 1)
                             tagged_used = True
                     else:
@@ -6796,7 +6803,11 @@ class H3LongVideos:
                 # The prompt itself says where each reference belongs: the shot whose
                 # text names <Picture N> gets image N, renumbered to match what that
                 # shot actually carries. Every untagged shot keeps its handoff.
-                gen_prompt, shot_refs, dropped = resolve_prompt_refs(gen_prompt, ref_slots)
+                gen_prompt, shot_refs, dropped = resolve_prompt_refs(
+                    gen_prompt,
+                    ref_slots,
+                    include_named=(shot_mode == "auto ref2v"),
+                )
                 for n in dropped:
                     if n not in ref_missing:
                         ref_missing.append(n)