fix(inference): METADATA-35 fit Call 1/2 images to 1024 long edge

Downscale processed PNGs only on the inference send path (keep alpha, never
upscale). Stored inventory/Woo blobs unchanged. Adds Pillow + unit/e2e-style
tests; cite tests on METADATA-35.

backend/inference.py

@@ -5,6 +5,7 @@ Cites METADATA-35..115, SETTINGS-30/45/90.
 
 import base64
 import hashlib
+import io
 import json
 import time
 import urllib.error
@@ -13,12 +14,53 @@ from typing import Any
 
 from fastapi import APIRouter
 from fastapi.responses import JSONResponse
+from PIL import Image
 
 from .db import get_connection
 from .crypto import decrypt
 
 router = APIRouter(prefix="/api/v1")
 
+# METADATA-35: Call 1 (and Call 2) send clean processed photos at 1024 px long edge or less.
+_INFER_MAX_LONG_EDGE = 1024
+
+
+def fit_long_edge_png(png_bytes: bytes, max_long_edge: int = _INFER_MAX_LONG_EDGE) -> bytes:
+    """Downscale a PNG so max(width, height) <= max_long_edge. Never upscale. Keep alpha.
+
+    Used only for the inference request payload; stored processed_blob / Woo media stay full size.
+    Unreadable or non-image bytes are returned unchanged so a bad blob cannot crash Call 1.
+    """
+    if not png_bytes or max_long_edge < 1:
+        return png_bytes
+    try:
+        im = Image.open(io.BytesIO(png_bytes))
+        im.load()
+    except Exception:
+        return png_bytes
+    w, h = im.size
+    long_edge = max(w, h)
+    if long_edge <= max_long_edge:
+        # Already within budget; re-encode only if we need a stable PNG (caller always base64s).
+        return png_bytes
+    scale = max_long_edge / float(long_edge)
+    nw = max(1, int(round(w * scale)))
+    nh = max(1, int(round(h * scale)))
+    # Preserve alpha: convert palette/etc. to RGBA when needed, LANCZOS downscale.
+    if im.mode not in ("RGB", "RGBA", "L", "LA"):
+        im = im.convert("RGBA")
+    out = im.resize((nw, nh), Image.Resampling.LANCZOS)
+    buf = io.BytesIO()
+    out.save(buf, format="PNG", optimize=True)
+    return buf.getvalue()
+
+
+def _image_part_from_processed(png_bytes: bytes) -> dict:
+    """Build one OpenAI-style image_url part from a processed PNG (METADATA-35 sized)."""
+    fitted = fit_long_edge_png(png_bytes)
+    b64 = base64.b64encode(fitted).decode()
+    return {"type": "image_url", "image_url": {"url": f"data:image/png;base64,{b64}"}}
+
 CALL1_SEED = (
     "Classify the slab's species, wood categories, edge type, and figure from the supplied "
     "photos, choosing ONLY from the supplied synced taxonomy lists (invent no category, term, "
@@ -203,17 +245,12 @@ async def call1(slab_id: str) -> dict:
     client = InferenceClient(base, key, model)
     act = _active_prompt(conn, "call1")
     prompt_text = act[1] if act else CALL1_SEED
-    # Build vision message with processed photos (1024px) + taxonomy lists.
+    # Build vision message with processed photos (METADATA-35: long edge <= 1024) + taxonomy.
     photos = conn.execute(
         "SELECT processed_blob FROM slab_photos WHERE slab_id=? AND processed_blob IS NOT NULL ORDER BY seq LIMIT 8",
         (slab_id,),
     ).fetchall()
-    image_parts = []
-    for p in photos:
-        image_parts.append({
-            "type": "image_url",
-            "image_url": {"url": f"data:image/png;base64,{base64.b64encode(p[0]).decode()}"},
-        })
+    image_parts = [_image_part_from_processed(p[0]) for p in photos]
     tax = _taxonomy_lists(conn)
     text = (f"{prompt_text}\n\nSYNCED TAXONOMY (only legal choices):\n{json.dumps(tax)}\n\n"
             f"SKU={row['sku']} length_in={row['length_in']} thickness_in={row['thickness_in']}")
@@ -359,10 +396,7 @@ async def call2(slab_id: str) -> dict:
         "SELECT processed_blob FROM slab_photos WHERE slab_id=? AND processed_blob IS NOT NULL ORDER BY seq LIMIT 8",
         (slab_id,),
     ).fetchall()
-    image_parts = [
-        {"type": "image_url", "image_url": {"url": f"data:image/png;base64,{base64.b64encode(p[0]).decode()}"}}
-        for p in photos
-    ]
+    image_parts = [_image_part_from_processed(p[0]) for p in photos]
     brand = conn.execute("SELECT value FROM settings WHERE key='brand_voice'").fetchone()
     geo = conn.execute("SELECT value FROM settings WHERE key='geo_context'").fetchone()
     text = (f"{prompt_text}\n\nDIMENSION FACT BLOCK (exact, invent no numbers):\n"