Report per-shot ref matches in H3 script

dumas_h3_longvideos.py

@@ -3273,6 +3273,45 @@ def annotate_script_guards(gens, anatomy_shots, anatomy_mode):
     return "\n".join(lines + ["", body]) if body else "\n".join(lines)
 
 
+def annotate_script_debug(gens, anatomy_shots, anatomy_mode, ref_slots):
+    parts = [annotate_script_guards(gens, anatomy_shots, anatomy_mode)]
+    ref_report = annotate_script_refs(gens, ref_slots)
+    if ref_report:
+        parts.insert(1, ref_report)
+    return "\n".join(part for part in parts if part)
+
+
+def annotate_script_refs(gens, ref_slots):
+    """Per-shot reference routing summary for the script socket."""
+    lines = []
+    for shot_index, gen in enumerate(gens or [], 1):
+        tagged = _slot_refs_for_text(gen, ref_slots)
+        named = _named_character_refs_for_text(gen, ref_slots)
+        seen = set()
+        merged = []
+        for slot_number, ref in tagged:
+            key = (slot_number, "tag")
+            if key in seen:
+                continue
+            seen.add(key)
+            merged.append((slot_number, ref, "tag"))
+        for slot_number, ref in named:
+            key = (slot_number, "name")
+            if key in seen or any(slot_number == existing[0] for existing in merged):
+                continue
+            seen.add(key)
+            merged.append((slot_number, ref, "name"))
+        if not merged:
+            lines.append(f"# shot {shot_index} refs: none")
+            continue
+        labels = []
+        for slot_number, ref, source in merged:
+            label = _reference_text(ref.get("name")) or _reference_text(ref.get("id")) or f"reference {slot_number}"
+            labels.append(f"Picture {slot_number} {label} (by {source})")
+        lines.append(f"# shot {shot_index} refs: " + "; ".join(labels))
+    return "\n".join(lines)
+
+
 def speech_flags(beats):
     """Per-beat: does it contain scripted (quoted) dialogue? Same rule the prompt
     builder uses to decide silencing, exposed so the renderer can also MUTE the
@@ -6455,7 +6494,7 @@ class H3LongVideos:
             # correctly-SHAPED empty one rather than None: a downstream LATENT input
             # would choke on None, and this keeps the preview wireable exactly like
             # a real run.
-            return (ph_img, ph_audio, plan, annotate_script_guards(gens, anatomy_shots, anatomy_mode), max(plan_lens),
+            return (ph_img, ph_audio, plan, annotate_script_debug(gens, anatomy_shots, anatomy_mode, ref_slots), max(plan_lens),
                     sum(plan_lens), shots, total, float(fps), int(fps),
                     _empty_av_latent(w, h, 5, fps)[0], global_soundscape, [], [])
 
@@ -6762,7 +6801,7 @@ class H3LongVideos:
             watermark_opacity, watermark_margin, intro_text, intro_seconds,
             intro_fade, intro_size, intro_position, overlay_font, overlay_stroke)
 
-        script = annotate_script_guards(gens, anatomy_shots, anatomy_mode)
+        script = annotate_script_debug(gens, anatomy_shots, anatomy_mode, ref_slots)
         actual = all_frames.shape[0] / fps
         uniform_len = len(set(shot_lens)) == 1
         shape_str = (f"{len(gens)} shot(s) x {shot_lens[0]}f (~{shot_lens[0] / fps:.1f}s each) "