fix(call2): no metadata dump in long description

Long desc is Hook + Benefits (+ brief trust close) only; forbid Full specs /
Attribute list / product-name header that re-lists DECLARED ATTRIBUTES or
DIMENSION FACT BLOCK. Defensively strip trailing Category:/Length:/Board Feet:
labeled dumps in parse.

backend/inference.py

@@ -128,6 +128,11 @@ CALL2_SEED = (
     "- The product name may already appear in DECLARED ATTRIBUTES — do not invent a new title.\n"
     "- Lead with usable traits (species, figure, size, edge, moisture, grade) and what the piece "
     "suits (tabletop, bar, etc.) without ad-copy fluff.\n"
+    "- Specs live in taxonomy and dimension fields on the product. Do NOT restate DECLARED "
+    "ATTRIBUTES or the DIMENSION FACT BLOCK as a labeled metadata list in the description "
+    "(no Category:/Figure:/Edge:/Grade:/Moisture:/Length:/Board Feet: dump, no \"Full specs\" "
+    "section, no \"Maple Slab\" attribute-header block). You may weave 1–2 key traits into "
+    "prose or benefit bullets naturally.\n"
     "\n"
     "SHORT DESCRIPTION (~50 words — most important on-page copy; plain text only):\n"
     "1. Lead with a practical outcome for a maker (what they can build / how the stock helps), "
@@ -136,13 +141,13 @@ CALL2_SEED = (
     "3. Close with a trust signal or concrete reason to choose this piece.\n"
     "\n"
     "LONG DESCRIPTION (plain text only — newlines and \"- \" bullets; no HTML, no markdown "
-    "markup). Four parts, in order:\n"
+    "markup). Hook + Benefits only (+ brief trust close). Do NOT add a Full specs / Attribute "
+    "list / product-name header that re-lists DECLARED ATTRIBUTES or DIMENSION FACT BLOCK:\n"
     "1. Hook — one sentence on what the piece suits and its standout usable trait.\n"
-    "2. Benefits — 3–5 bullets; each maker-useful trait first, then the supporting spec.\n"
+    "2. Benefits — 3–5 bullets; each maker-useful trait first, then the supporting detail. "
-    "3. Full specs — clean list from DECLARED ATTRIBUTES + DIMENSION FACT BLOCK "
+    "Weave in at most 1–2 key traits naturally — not an exhaustive labeled catalog.\n"
-    "(dims, species, moisture, edge, figure, grade, etc.).\n"
+    "3. Trust close (optional, brief) — return/warranty/brand; use Brand voice when present; "
-    "4. Trust close — return/warranty/brand; use Brand voice when present; otherwise a short "
+    "otherwise a short neutral close with no invented policies.\n"
-    "neutral close with no invented policies.\n"
     "\n"
     "Return ONE JSON object only (no prose outside JSON) with exactly these keys:\n"
     '{"short_description": "<plain text ~50 words>", "description": "<plain text long description>"}'
@@ -693,6 +698,95 @@ async def call2_prompt(slab_id: str) -> dict:
     }
 
 
+# Labels that mark a Woo/taxonomy/dimension metadata dump when used as "- Label: …".
+_ATTR_DUMP_LABELS = (
+    "Category",
+    "Primary",
+    "Secondary",
+    "Species",
+    "Figure",
+    "Edge",
+    "Grade",
+    "Moisture",
+    "Features",
+    "Feature",
+    "Title",
+    "Length",
+    "Thickness",
+    "Width",
+    "Avg Width",
+    "Average Width",
+    "Width Min",
+    "Width Max",
+    "Min Width",
+    "Max Width",
+    "Board Feet",
+    "Bdft",
+    "BF",
+    "Dimensions",
+    "Dims",
+)
+_ATTR_DUMP_LABEL_RE = re.compile(
+    r"(?i)^\s*-\s*(?:"
+    + "|".join(re.escape(l) for l in _ATTR_DUMP_LABELS)
+    + r")\s*:",
+)
+_ATTR_DUMP_HEADER_RE = re.compile(
+    r"(?i)^\s*(?:Full\s+specs?|Specifications?|Attributes?|Attribute\s+list|"
+    r"[A-Za-z][A-Za-z0-9 /&-]{0,40}\s+Slab)\s*$",
+)
+
+
+def _strip_attribute_dump(value: str) -> str:
+    """Remove a labeled Category:/Length:/Board Feet: metadata dump from long desc.
+
+    Specs live in taxonomy/dims on the product; a trailing (or mid-body) block that
+    re-lists them as "- Label: value" lines drifts from Woo fields. Keeps Hook +
+    Benefits (+ trust close) prose around the dump.
+    """
+    if not value:
+        return ""
+    lines = value.split("\n")
+    n = len(lines)
+    is_attr = [_ATTR_DUMP_LABEL_RE.match(ln) is not None for ln in lines]
+
+    # Find contiguous runs of ≥3 attribute bullets (optionally preceded by a header).
+    runs: list[tuple[int, int]] = []
+    i = 0
+    while i < n:
+        if not is_attr[i]:
+            i += 1
+            continue
+        start = i
+        while i < n and is_attr[i]:
+            i += 1
+        end = i  # exclusive
+        if end - start >= 3:
+            # Absorb a single preceding header / "Maple Slab" title line.
+            hdr = start - 1
+            if hdr >= 0 and lines[hdr].strip() and _ATTR_DUMP_HEADER_RE.match(lines[hdr]):
+                # Do not swallow a benefits bullet that happens to match poorly.
+                if not is_attr[hdr]:
+                    start = hdr
+            # Absorb blank lines immediately before the block.
+            while start > 0 and not lines[start - 1].strip():
+                start -= 1
+            runs.append((start, end))
+
+    if not runs:
+        return value.strip()
+
+    # Prefer the last run (typical trailing dump); drop all qualifying runs.
+    keep = [True] * n
+    for start, end in runs:
+        for j in range(start, end):
+            keep[j] = False
+    out_lines = [ln for j, ln in enumerate(lines) if keep[j]]
+    text = "\n".join(out_lines)
+    text = re.sub(r"\n{3,}", "\n\n", text)
+    return text.strip()
+
+
 def _html_to_plain(value: str) -> str:
     """Defensive: strip HTML/markup so Review fields never show raw tags.
 
@@ -717,11 +811,18 @@ def _html_to_plain(value: str) -> str:
     return t.strip()
 
 
+def _plain_long_desc(value: str) -> str:
+    """HTML→plain then drop labeled attribute/metadata dumps from long description."""
+    return _strip_attribute_dump(_html_to_plain(value))
+
+
 def parse_call2_response(raw: str) -> tuple[str | None, str]:
     """Split Call 2 model output into (short_description, description).
 
     Preferred shape is a JSON object with short_description + description.
     Both fields are forced to plain text (HTML stripped) before return.
+    Long description also drops a trailing/mid labeled attribute dump
+    (Category:/Length:/Board Feet: …) so specs stay on Woo fields.
     If JSON is missing or incomplete, treat the whole body as description
     (stripped) and leave short_description None (caller keeps any prior short).
     """
@@ -747,11 +848,11 @@ def parse_call2_response(raw: str) -> tuple[str | None, str]:
             try:
                 obj = json.loads(body[start : end + 1])
             except (ValueError, TypeError):
-                return None, _html_to_plain(text)
+                return None, _plain_long_desc(text)
         else:
-            return None, _html_to_plain(text)
+            return None, _plain_long_desc(text)
     if not isinstance(obj, dict):
-        return None, _html_to_plain(text)
+        return None, _plain_long_desc(text)
     short = obj.get("short_description")
     desc = obj.get("description")
     if isinstance(short, str):
@@ -759,9 +860,9 @@ def parse_call2_response(raw: str) -> tuple[str | None, str]:
     else:
         short = None
     if isinstance(desc, str) and desc.strip():
-        return short, _html_to_plain(desc)
+        return short, _plain_long_desc(desc)
     # JSON without a usable description — fall back to raw (stripped).
-    return short, _html_to_plain(text)
+    return short, _plain_long_desc(text)
 
 
 @router.post("/slabs/{slab_id}/inference/call2")

backend/tests/test_inference.py

@@ -101,6 +101,16 @@ def test_call2_seed_has_short_long_structure(env):
     assert '"description"' in CALL2_SEED
     assert "<HTML long description>" not in CALL2_SEED
     assert "plain text long description" in CALL2_SEED
+    # No Full-specs / attribute metadata dump in long desc (specs live on Woo fields).
+    assert "Hook + Benefits only" in CALL2_SEED
+    assert "Do NOT restate DECLARED" in CALL2_SEED
+    assert 'no "Full specs"' in CALL2_SEED or "Do NOT add a Full specs" in CALL2_SEED
+    assert "Category:/Figure:/Edge:" in CALL2_SEED
+    assert "1–2 key traits" in CALL2_SEED
+    assert "labeled catalog" in CALL2_SEED or "labeled metadata list" in CALL2_SEED
+    # Must not instruct a Full-specs numbered part (old 4-part structure).
+    assert "3. Full specs" not in CALL2_SEED
+    assert "Four parts, in order" not in CALL2_SEED
 
 
 def test_taxonomy_lists_call1_closed_shape():
@@ -296,6 +306,71 @@ def test_parse_call2_response_json_and_fenced():
     assert _html_to_plain("  Kiln-dried walnut for a bar top.  ") == "Kiln-dried walnut for a bar top."
 
 
+def test_strip_attribute_dump_trailing_and_header():
+    """Defensive: drop labeled Category:/Length: dumps from long description."""
+    from backend.inference import _strip_attribute_dump, parse_call2_response
+    import json as _json
+
+    hook_benefits = (
+        "Kiln-dried maple suited to a compact bar top.\n"
+        "- Stable stock: kiln dried, ready to mill.\n"
+        "- Live edge on one side for a natural apron.\n"
+        "- Character grade with usable burl figure."
+    )
+    dump = (
+        "Maple Slab\n"
+        "- Category: Live edge wood\n"
+        "- Figure: Burl, Curl, Figured\n"
+        "- Edge: Live Edge 1 side\n"
+        "- Grade: Character\n"
+        "- Moisture: Kiln dried\n"
+        "- Length: 45.0in\n"
+        "- Board Feet: 12.4"
+    )
+    trust = "Ships from the mill; ask about returns in Brand voice."
+    full = f"{hook_benefits}\n\n{dump}\n\n{trust}"
+    cleaned = _strip_attribute_dump(full)
+    assert "Category:" not in cleaned
+    assert "Board Feet:" not in cleaned
+    assert "Maple Slab" not in cleaned
+    assert "Kiln-dried maple" in cleaned
+    assert "Live edge on one side" in cleaned
+    assert "Ships from the mill" in cleaned
+
+    # Mid-body Full specs header also stripped; benefits bullets without Label: kept.
+    mid = (
+        "Hook line.\n"
+        "- Good for a bench seat.\n"
+        "\n"
+        "Full specs\n"
+        "- Species: Maple\n"
+        "- Length: 45.0in\n"
+        "- Thickness: 2.0in\n"
+        "\n"
+        "Neutral trust close."
+    )
+    mid_clean = _strip_attribute_dump(mid)
+    assert "Full specs" not in mid_clean
+    assert "Species:" not in mid_clean
+    assert "Good for a bench seat" in mid_clean
+    assert "Neutral trust close" in mid_clean
+
+    # Fewer than 3 attr lines: leave alone (not a dump).
+    short_list = "Hook.\n- Length: 45.0in\n- Thickness: 2in\nTrust."
+    assert _strip_attribute_dump(short_list) == short_list.strip()
+
+    # parse_call2_response applies strip on description (not short).
+    payload = _json.dumps({
+        "short_description": "Short with - Length: 1in - Width: 2in - Thickness: 3in ok",
+        "description": full,
+    })
+    short, desc = parse_call2_response(payload)
+    assert short and "Length: 1in" in short  # short untouched by attr-dump strip
+    assert "Category:" not in desc
+    assert "Board Feet:" not in desc
+    assert "Kiln-dried maple" in desc
+
+
 def test_call1_disabled_even_when_inference_off(env):
     """Call 1 hard-disable wins over inference_enabled=false (still 410, not 409)."""
     client, conn = env

docs/contracts/PROMPTS.md

@@ -188,6 +188,11 @@ Rules that always apply:
 - The product name may already appear in DECLARED ATTRIBUTES — do not invent a new title.
 - Lead with usable traits (species, figure, size, edge, moisture, grade) and what the piece
   suits (tabletop, bar, etc.) without ad-copy fluff.
+- Specs live in taxonomy and dimension fields on the product. Do NOT restate DECLARED
+  ATTRIBUTES or the DIMENSION FACT BLOCK as a labeled metadata list in the description
+  (no Category:/Figure:/Edge:/Grade:/Moisture:/Length:/Board Feet: dump, no "Full specs"
+  section, no "Maple Slab" attribute-header block). You may weave 1–2 key traits into
+  prose or benefit bullets naturally.
 
 SHORT DESCRIPTION (~50 words — most important on-page copy; plain text only):
 1. Lead with a practical outcome for a maker (what they can build / how the stock helps),
@@ -196,20 +201,21 @@ SHORT DESCRIPTION (~50 words — most important on-page copy; plain text only):
 3. Close with a trust signal or concrete reason to choose this piece.
 
 LONG DESCRIPTION (plain text only — newlines and `- ` bullets; no HTML, no markdown markup).
-Four parts, in order:
+Hook + Benefits only (+ brief trust close). Do NOT add a Full specs / Attribute list /
+product-name header that re-lists DECLARED ATTRIBUTES or DIMENSION FACT BLOCK:
 1. Hook — one sentence on what the piece suits and its standout usable trait.
-2. Benefits — 3–5 bullets; each maker-useful trait first, then the supporting spec.
+2. Benefits — 3–5 bullets; each maker-useful trait first, then the supporting detail.
-3. Full specs — clean list from DECLARED ATTRIBUTES + DIMENSION FACT BLOCK (dims, species,
+   Weave in at most 1–2 key traits naturally — not an exhaustive labeled catalog.
-   moisture, edge, figure, grade, etc.).
+3. Trust close (optional, brief) — return/warranty/brand; use Brand voice when present;
-4. Trust close — return/warranty/brand; use Brand voice when present; otherwise a short
+   otherwise a short neutral close with no invented policies.
-   neutral close with no invented policies.
 
 Return ONE JSON object only (no prose outside JSON) with exactly these keys:
 `{"short_description": "<plain text ~50 words>", "description": "<plain text long description>"}`
 
 Runtime appends DECLARED ATTRIBUTES (title + taxonomy; not a prior short), DIMENSION FACT
 BLOCK, and `Brand voice:` / `Geo:` from Settings (`SETTINGS-30`). The POST handler parses the
-JSON, strips any residual HTML to plain text, and persists both fields.
+JSON, strips any residual HTML to plain text, drops a labeled Category:/Length:/Board Feet:
+attribute dump if the model still emits one, and persists both fields.
 
 ## 7. Prompt management