Normalize built-in row action metadata

docs/prompt-pool-routing-map.md

@@ -132,7 +132,7 @@ Core helper ownership:
 | `krea_row_fields.py` | Shared Krea normal-row field extraction for item, scene, pose, expression, composition/source-composition, camera, and style used by normal and configured-cast routes. |
 | `krea_cast.py` | Shared formatter cast descriptor parsing, cast labels, cast prose, natural cast descriptor text, and label replacement used by Krea2 and caption routes. |
 | `prompt_hygiene.py` | Generic prompt, caption, and negative-prompt cleanup, including route-agnostic negative-prompt merge/dedupe. |
-| `row_normalization.py` | Final prompt-row and pair metadata normalization: legacy built-in subject/count/scene metadata enrichment, trigger prepending, extra-positive append, negative merge/dedupe, caption-part joining, embedded soft/hard row output and side-metadata synchronization, and embedded row sanitation. |
+| `row_normalization.py` | Final prompt-row and pair metadata normalization: legacy built-in subject/count/scene/item/pose/expression metadata enrichment, trigger prepending, extra-positive append, negative merge/dedupe, caption-part joining, embedded soft/hard row output and side-metadata synchronization, and embedded row sanitation. |
 | `formatter_detail.py` | Shared formatter detail-level choices, normalization, and concise/balanced/dense gates used by Krea2 and caption routes. |
 | `formatter_input.py` | Shared formatter input parsing: text cleanup, metadata/source JSON detection, trigger-prefix stripping, shared prompt field-label inventory, fallback field-label stripping, `Avoid:` splitting, prompt-field extraction, and metadata row-value fallback. |
 | `formatter_target.py` | Shared formatter target choices and normalization for `auto`, `single`, `softcore`, and `hardcore`, including pair-side selection and combined-caption inclusion policy. |

row_normalization.py

@@ -1,5 +1,6 @@
 from __future__ import annotations
 
+import re
 from typing import Any
 
 try:
@@ -67,6 +68,51 @@ def _legacy_subject_metadata(row: dict[str, Any]) -> tuple[str, str, int | None,
     return "", subject, None, None
 
 
+_LEGACY_PROMPT_FIELD_LABELS = (
+    "Ages",
+    "Body types",
+    "Scene",
+    "Pose",
+    "Facial expressions",
+    "Facial expression",
+    "Clothing",
+    "Prop/detail",
+    "Composition",
+    "Use",
+    "Avoid",
+)
+
+
+def _clean_text(value: Any) -> str:
+    text = "" if value is None else str(value)
+    text = re.sub(r"\s+", " ", text.replace("\n", " ")).strip()
+    return re.sub(r"\s+([,.;:])", r"\1", text)
+
+
+def _legacy_prompt_field(row: dict[str, Any], label: str) -> str:
+    prompt = _clean_text(row.get("prompt"))
+    if not prompt:
+        return ""
+    labels = "|".join(re.escape(name) for name in _LEGACY_PROMPT_FIELD_LABELS)
+    pattern = rf"{re.escape(label)}:\s*(.*?)(?=\. (?:{labels}):|\. Use\b|\. Avoid\b|$)"
+    match = re.search(pattern, prompt)
+    if not match:
+        return ""
+    return _clean_text(match.group(1)).rstrip(".")
+
+
+def _clean_legacy_pose(value: Any) -> str:
+    text = _clean_text(value)
+    text = text.replace(", affectionate and flirtatious but non-explicit", "")
+    return text
+
+
+def _clean_legacy_clothing(value: Any) -> str:
+    text = _clean_text(value)
+    text = re.sub(r",?\s*(?:fashion editorial|resort) styling$", "", text, flags=re.IGNORECASE)
+    return text.strip(" ,")
+
+
 def enrich_legacy_row_metadata(row: dict[str, Any]) -> dict[str, Any]:
     if row.get("source") != "built_in_generator":
         return row
@@ -87,6 +133,15 @@ def enrich_legacy_row_metadata(row: dict[str, Any]) -> dict[str, Any]:
         if scene_text:
             row["scene_text"] = scene_text
             row.setdefault("scene_entry", {"slug": scene_slug, "prompt": scene_text})
+    pose = _clean_legacy_pose(_legacy_prompt_field(row, "Pose"))
+    _setdefault_nonempty(row, "pose", pose)
+    expression = _legacy_prompt_field(row, "Facial expression") or _legacy_prompt_field(row, "Facial expressions")
+    _setdefault_nonempty(row, "expression", expression)
+    clothing = _clean_legacy_clothing(_legacy_prompt_field(row, "Clothing"))
+    _setdefault_nonempty(row, "clothing", clothing)
+    _setdefault_nonempty(row, "item", clothing)
+    if clothing:
+        _setdefault_nonempty(row, "item_label", "Clothing")
     return row
 
 

tools/prompt_smoke.py

@@ -578,6 +578,11 @@ def smoke_builtin_single() -> None:
     _expect_text("builtin_single_woman.scene_text", row.get("scene_text"), 12)
     _expect(row.get("scene_slug") == row.get("scene"), "builtin single row lost legacy scene slug metadata")
     _expect(row.get("scene_entry", {}).get("slug") == row.get("scene"), "builtin single row lost scene_entry slug")
+    item = _expect_text("builtin_single_woman.item", row.get("item"), 8)
+    pose = _expect_text("builtin_single_woman.pose", row.get("pose"), 8)
+    _expect(row.get("item_label") == "Clothing", "builtin single row lost item label")
+    _expect(row.get("clothing") == item, "builtin single row did not mirror clothing into item metadata")
+    _expect("fashion editorial styling" not in item.lower(), "builtin single item kept generic styling suffix")
     _expect("cast_summary" not in row, "builtin single row should not masquerade as configured cast")
     _expect_trigger_once("builtin_single_woman.prompt", row.get("prompt"), Trigger)
     krea = krea_formatter.format_krea2_prompt("", metadata_json=_json(row), target="single")
@@ -593,6 +598,28 @@ def smoke_builtin_single() -> None:
     _expect(caption_method.endswith("metadata(single)"), "builtin single caption route did not use single metadata branch")
     _expect("woman" in caption.lower(), "builtin single caption route lost normalized subject")
     _expect(row.get("scene_text") in caption, "builtin single caption route used scene slug instead of scene text")
+
+    metadata_only = dict(row)
+    metadata_only["prompt"] = ""
+    metadata_only["caption"] = ""
+    krea_metadata = krea_formatter.format_krea2_prompt("", metadata_json=_json(metadata_only), target="single")
+    sdxl_metadata = sdxl_formatter.format_sdxl_prompt("", metadata_json=_json(metadata_only), target="single", trigger=SdxlTrigger, prepend_trigger=True)
+    caption_metadata, caption_metadata_method = caption_naturalizer.naturalize_caption(
+        "",
+        metadata_json=_json(metadata_only),
+        target="single",
+        trigger=Trigger,
+        include_trigger=True,
+    )
+    _expect(item in str(krea_metadata.get("krea_prompt", "")), "Krea metadata-only built-in route lost explicit item")
+    _expect(pose in str(krea_metadata.get("krea_prompt", "")), "Krea metadata-only built-in route lost explicit pose")
+    item_anchor = " ".join(re.findall(r"[a-z0-9]+", item.lower())[:3])
+    pose_anchor = " ".join(re.findall(r"[a-z0-9]+", pose.lower())[:4])
+    sdxl_metadata_prompt = str(sdxl_metadata.get("sdxl_prompt", "")).lower()
+    _expect(item_anchor in sdxl_metadata_prompt, "SDXL metadata-only built-in route lost explicit item")
+    _expect(pose_anchor in sdxl_metadata_prompt, "SDXL metadata-only built-in route lost explicit pose")
+    _expect(caption_metadata_method.endswith("metadata(single)"), "Caption metadata-only built-in route did not use single metadata branch")
+    _expect(item in caption_metadata and pose in caption_metadata, "Caption metadata-only built-in route lost explicit item or pose")
     _expect_formatter_outputs(row, "builtin_single_woman", target="single")
 
 
@@ -2826,7 +2853,12 @@ def smoke_row_normalization_policy() -> None:
             "source": "built_in_generator",
             "primary_subject": "two women",
             "scene": "office",
-            "prompt": "Two adults in a clean legacy prompt.",
+            "prompt": (
+                "Two adults in a clean legacy prompt. Scene: old room. "
+                "Pose: standing close, affectionate and flirtatious but non-explicit. "
+                "Facial expressions: one with a calm smile, the other with a side glance. "
+                "Clothing: coordinated satin outfits, resort styling. Composition: vertical old frame."
+            ),
             "caption": "legacy couple caption",
             "negative_prompt": "bad anatomy",
         },
@@ -2839,6 +2871,10 @@ def smoke_row_normalization_policy() -> None:
     _expect(legacy_couple.get("scene_slug") == "office", "Legacy couple row lost scene slug")
     _expect("cozy office desk" in str(legacy_couple.get("scene_text", "")), "Legacy couple row lost readable scene text")
     _expect(legacy_couple.get("scene_entry", {}).get("slug") == "office", "Legacy couple row lost scene entry")
+    _expect(legacy_couple.get("pose") == "standing close", "Legacy couple row did not clean pose suffix")
+    _expect(legacy_couple.get("item") == "coordinated satin outfits", "Legacy couple row did not clean clothing suffix")
+    _expect(legacy_couple.get("item_label") == "Clothing", "Legacy couple row lost item label")
+    _expect("calm smile" in str(legacy_couple.get("expression", "")), "Legacy couple row lost expression metadata")
     _expect("cast_summary" not in legacy_couple, "Legacy couple row should not gain configured-cast summary")
 
     legacy_group = row_normalization.normalize_prompt_row(