Add formatter metadata smoke fixtures

docs/prompt-architecture-improvement-plan.md

@@ -257,8 +257,8 @@ Improve later:
 Near-term:
 
 - Add final row hygiene already done through `prompt_hygiene.py`.
-- Add a metadata smoke checker for representative rows through
+- Add a metadata smoke checker for representative generated rows and static
-  `tools/prompt_smoke.py`.
+  formatter fixtures through `tools/prompt_smoke.py`.
 - Normalize every row with one function before JSON serialization.
 
 Medium-term:
@@ -359,5 +359,5 @@ Medium-term:
    checks.
 2. Extract hardcore role graph generation from `prompt_builder.py` into a
    dedicated `hardcore_role_graphs.py` module.
-3. Add route-level smoke fixtures for more representative Krea/SDXL/caption metadata
+3. Add more route-level smoke fixtures for generated edge cases that are not
-   rows.
+   covered by the current static Krea/SDXL/caption metadata fixtures.

docs/prompt-pool-routing-map.md

@@ -722,6 +722,9 @@ pair metadata through the core Python APIs, then verifies:
   normal camera text, and preserve composition punctuation before the style
   suffix;
 - expression-disabled rows do not fall back to generated expression text.
+- static formatter metadata fixtures keep source-provided action families
+  stable across Krea2 prose, SDXL tags, and natural captions even when raw item
+  text contains distracting wording.
 
 ## Editing Cheatsheet
 

tools/prompt_smoke.py

@@ -273,6 +273,64 @@ def _prompt_row(
     return row
 
 
+def _fixture_hardcore_row(**overrides: Any) -> dict[str, Any]:
+    row: dict[str, Any] = {
+        "source": "json_category",
+        "prompt": "Fixture explicit adult prompt for metadata route.",
+        "caption": "fixture caption",
+        "negative_prompt": "low quality, bad anatomy",
+        "main_category": "Hardcore sexual poses",
+        "subcategory": "Penetrative sex",
+        "category_slug": "hardcore_sexual_poses",
+        "subcategory_slug": "penetrative_sex",
+        "subject_type": "configured_cast",
+        "subject_phrase": "1 adult woman and 1 adult man",
+        "cast_summary": "1 woman, 1 man",
+        "cast_descriptor_text": (
+            "Woman A: 25-year-old adult woman, slim figure, fair skin, blonde hair, blue eyes; "
+            "Man A: 40-year-old adult man, average figure, tan skin, dark hair"
+        ),
+        "cast_descriptors": [
+            "Woman A: 25-year-old adult woman, slim figure, fair skin, blonde hair, blue eyes",
+            "Man A: 40-year-old adult man, average figure, tan skin, dark hair",
+        ],
+        "women_count": 1,
+        "men_count": 1,
+        "person_count": 2,
+        "item": (
+            "missionary position while full-body penetrative sex, hands gripping the ass, "
+            "mouth close to the ear, and explicit genital contact visible"
+        ),
+        "custom_item": "Penetrative sex",
+        "item_label": "Sexual pose",
+        "item_axis_values": {
+            "position": "missionary position",
+            "penetration_act": "full-body penetrative sex",
+            "mouth_detail": "mouth close to the ear",
+        },
+        "scene_text": "private studio room with warm light",
+        "scene_kind": "explicit adult sex scene",
+        "pose": "configured explicit pose",
+        "composition": "front-facing full-body frame",
+        "source_composition": "front-facing full-body frame",
+        "role_graph": (
+            "Woman A lies on her back with legs open around Man A's hips while Man A is above her between her thighs; "
+            "Man A's hips press close and Man A's penis thrusts into her pussy."
+        ),
+        "source_role_graph": (
+            "Woman A lies on her back with legs open around Man A's hips while Man A is above her between her thighs; "
+            "Man A's hips press close and Man A's penis thrusts into her pussy."
+        ),
+        "expression": "focused adult expression",
+        "action_family": "penetration",
+        "position_family": "penetrative",
+        "position_key": "missionary",
+        "position_keys": ["missionary"],
+    }
+    row.update(overrides)
+    return row
+
+
 def smoke_builtin_single() -> None:
     row = _prompt_row(name="builtin_single_woman", category="woman", subcategory="random", seed=1001, men_count=0)
     _expect(row.get("source") == "built_in_generator", "builtin row should come from built-in generator")
@@ -681,6 +739,99 @@ def smoke_no_expression_fallback() -> None:
     _expect_formatter_outputs(row, "hardcore_expression_disabled", target="single")
 
 
+def smoke_formatter_metadata_fixtures() -> None:
+    cases = [
+        {
+            "name": "fixture_penetration_text_noise",
+            "row": _fixture_hardcore_row(),
+            "krea_terms": ("penis thrusts",),
+            "sdxl_terms": ("penetrative sex", "missionary"),
+            "caption_terms": ("penetrative action",),
+        },
+        {
+            "name": "fixture_manual_source_family",
+            "row": _fixture_hardcore_row(
+                subcategory="Manual stimulation",
+                subcategory_slug="manual_stimulation",
+                item="wet fingers moving between the thighs, one hand braced on the hip, wet shine on fingers and inner thighs",
+                custom_item="Manual stimulation",
+                item_axis_values={
+                    "position": "kneeling hand-between-thighs position",
+                    "manual_act": "wet fingers moving between the thighs",
+                },
+                composition="close crop on hands and face",
+                source_composition="close crop on hands and face",
+                role_graph=(
+                    "Woman A reclines with thighs open while Man A's hand is between her legs, "
+                    "fingers visibly stimulating her pussy."
+                ),
+                source_role_graph=(
+                    "Woman A reclines with thighs open while Man A's hand is between her legs, "
+                    "fingers visibly stimulating her pussy."
+                ),
+                action_family="foreplay",
+                position_family="manual",
+                position_key="fingering",
+                position_keys=["fingering", "open_thighs"],
+            ),
+            "krea_terms": ("fingers visibly stimulating",),
+            "sdxl_terms": ("manual stimulation", "fingering"),
+            "caption_terms": ("manual action",),
+        },
+        {
+            "name": "fixture_climax_family",
+            "row": _fixture_hardcore_row(
+                subcategory="Cumshot and climax",
+                subcategory_slug="cumshot_climax",
+                item="external cumshot with cum on lower back and ass, explicit semen aftermath visible",
+                custom_item="Cumshot and climax",
+                item_axis_values={"position": "on all fours with hips raised"},
+                composition="low-angle post-orgasm frame",
+                source_composition="low-angle post-orgasm frame",
+                role_graph=(
+                    "Woman A is on all fours with hips raised while Man A is positioned behind her "
+                    "and ejaculates semen across her ass, thighs, and lower back."
+                ),
+                source_role_graph=(
+                    "Woman A is on all fours with hips raised while Man A is positioned behind her "
+                    "and ejaculates semen across her ass, thighs, and lower back."
+                ),
+                action_family="climax",
+                position_family="climax",
+                position_key="doggy",
+                position_keys=["doggy"],
+            ),
+            "krea_terms": ("ejaculates semen",),
+            "sdxl_terms": ("climax", "semen"),
+            "caption_terms": ("climax action",),
+        },
+    ]
+    for case in cases:
+        name = case["name"]
+        row = case["row"]
+        _expect_custom_row(row, name)
+        metadata = _json(row)
+
+        krea = krea_formatter.format_krea2_prompt("", metadata_json=metadata, target="single")
+        krea_prompt = _expect_text(f"{name}.krea_prompt", krea.get("krea_prompt"), 40).lower()
+        _expect("metadata" in krea.get("method", ""), f"{name}.krea did not use metadata")
+        _expect("role graph:" not in krea_prompt and "sexual pose:" not in krea_prompt, f"{name}.krea leaked raw labels")
+        for term in case["krea_terms"]:
+            _expect(term in krea_prompt, f"{name}.krea missing {term!r}")
+
+        sdxl = sdxl_formatter.format_sdxl_prompt("", metadata_json=metadata, target="single", trigger=SdxlTrigger, prepend_trigger=True)
+        sdxl_prompt = _expect_text(f"{name}.sdxl_prompt", sdxl.get("sdxl_prompt"), 40).lower()
+        _expect("metadata" in sdxl.get("method", ""), f"{name}.sdxl did not use metadata")
+        for term in case["sdxl_terms"]:
+            _expect(term in sdxl_prompt, f"{name}.sdxl missing {term!r}")
+
+        caption, method = caption_naturalizer.naturalize_caption("", metadata_json=metadata, trigger=Trigger, include_trigger=True)
+        caption_text = _expect_text(f"{name}.caption", caption, 40).lower()
+        _expect("metadata" in method, f"{name}.caption did not use metadata")
+        for term in case["caption_terms"]:
+            _expect(term in caption_text, f"{name}.caption missing {term!r}")
+
+
 SMOKE_CASES: list[tuple[str, Callable[[], None]]] = [
     ("builtin_single_woman", smoke_builtin_single),
     ("camera_scene_single", smoke_camera_scene_single),
@@ -694,6 +845,7 @@ SMOKE_CASES: list[tuple[str, Callable[[], None]]] = [
     ("pov_camera_scene", smoke_pov_camera_scene),
     ("krea_pov_penetration_route", smoke_krea_pov_penetration_route),
     ("expression_disabled", smoke_no_expression_fallback),
+    ("formatter_metadata_fixtures", smoke_formatter_metadata_fixtures),
 ]