Deduplicate pair caption cast descriptors

caption_metadata_routes.py

@@ -363,12 +363,11 @@ def insta_of_pair_from_row_result(
     same_soft_cast = options.get("softcore_cast") == "same_as_hardcore"
 
     parts = []
-    if cast_descriptor_text and same_soft_cast:
+    if not soft_text and not hard_text:
+        if cast_descriptor_text:
             parts.append(deps.natural_cast_descriptor_text(cast_descriptor_text))
         elif descriptor:
             parts.append(f"A {descriptor}")
-    if cast_descriptor_text and not same_soft_cast:
-        parts.append(deps.natural_cast_descriptor_text(cast_descriptor_text))
     if same_soft_cast and include_soft:
         parts.append(
             deps.softcore_caption_setup_phrase(

docs/prompt-architecture-improvement-plan.md

@@ -33,6 +33,9 @@ The map audit currently sees:
   caption outputs can be debugged by category, action/position family, selected
   pair side, scene profile, position keys, and POV labels instead of only
   proving that a metadata branch was used.
+- Insta/OF side-target training captions no longer prepend shared cast
+  descriptors when the selected side row already emits its own cast prose, and
+  route simulation flags repeated cast descriptors.
 
 ## Architectural Finding
 

docs/prompt-pool-routing-map.md

@@ -1004,6 +1004,7 @@ issues for:
   action/position family, selected pair side, scene profile, position keys, and
   POV labels;
 - raw builder labels leaking into Krea output;
+- repeated cast descriptors in training-caption formatter output;
 - duplicate negative-prompt comma items;
 - softcore prompt noise;
 - POV routes emitting third-person camera text or losing first-person wording;

tools/prompt_map_audit.py

@@ -110,6 +110,10 @@ AUDIT_DOC_SNIPPETS: tuple[tuple[str, str], ...] = (
         "docs/prompt-pool-routing-map.md",
         "formatter route traces exposing selected row metadata",
     ),
+    (
+        "docs/prompt-pool-routing-map.md",
+        "repeated cast descriptors in training-caption formatter output",
+    ),
 )
 
 PROMPT_ROW_READ_SCAN_GLOBS: tuple[str, ...] = (

tools/prompt_route_simulation.py

@@ -505,6 +505,22 @@ def _formatter_expectation_issues(
     return issues
 
 
+def _caption_cast_descriptor_issues(name: str, row: dict[str, Any] | None, caption_text: str) -> list[str]:
+    if not isinstance(row, dict):
+        return []
+    descriptor = row.get("cast_descriptor_text") or row.get("shared_cast_descriptors")
+    if isinstance(descriptor, list):
+        descriptor_text = "; ".join(str(item or "").strip() for item in descriptor if str(item or "").strip())
+    else:
+        descriptor_text = str(descriptor or "").strip()
+    if not descriptor_text:
+        return []
+    natural_descriptor = caption_naturalizer._natural_cast_descriptor_text(descriptor_text)
+    if natural_descriptor and caption_text.count(natural_descriptor) > 1:
+        return [f"{name}.caption: repeated_cast_descriptor"]
+    return []
+
+
 def _trace_dict(formatter_name: str, payload: dict[str, Any]) -> tuple[dict[str, Any], str]:
     trace_text = str(payload.get("route_trace_json") or "")
     if not trace_text:
@@ -637,6 +653,7 @@ def _formatter_issues(
         if "metadata" not in str(method or ""):
             issues.append(f"{name}.{formatter_name}: not_metadata_route:{method}")
     issues.extend(_formatter_trace_issues(name, formats, target=target, row=row))
+    issues.extend(_caption_cast_descriptor_issues(name, row, caption_text))
 
     for label, value in (
         (f"{name}.krea_negative", krea.get("negative_prompt")),

tools/prompt_smoke.py

@@ -4341,6 +4341,17 @@ def smoke_caption_metadata_routes() -> None:
     _expect("Softcore side:" not in hard_route.prose, "Caption hardcore target should not include soft label")
     _expect("Hardcore side:" not in hard_route.prose, "Caption hardcore target should not keep combined pair labels")
     _expect(soft_route.prose != hard_route.prose, "Caption pair soft/hard targets should produce distinct prose")
+    shared_cast = pair.get("shared_cast_descriptors")
+    if isinstance(shared_cast, list):
+        shared_cast_text = "; ".join(str(item or "").strip() for item in shared_cast if str(item or "").strip())
+    else:
+        shared_cast_text = str(shared_cast or "").strip()
+    shared_cast_caption = caption_naturalizer._natural_cast_descriptor_text(shared_cast_text)
+    if shared_cast_caption:
+        _expect(
+            hard_route.prose.count(shared_cast_caption) <= 1,
+            "Caption hardcore target repeated shared cast descriptors",
+        )
     public_hard, public_hard_method = caption_naturalizer.naturalize_caption(
         "",
         metadata_json=_json(pair),