feat: switch to MiniCPM finetune prompt + fix name==code filter

- Update LIST_PROMPT for MiniCPM finetune output format
- Fix extract_entries: treat name==code as unnamed (finetune echoes code when no colour name known)
- Remove streaming from call_llama (heartbeat proxy handles CF timeout; MiniCPM ~64s < 100s limit)
- Set LLAMA_SERVER default to inference.llmops.pl tunnel

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

app.py

@@ -12,7 +12,7 @@ from html import escape
 from PIL import Image, ImageOps
 import inventory
 
-LLAMA_SERVER = os.environ.get("LLAMA_SERVER", "")
+LLAMA_SERVER = os.environ.get("LLAMA_SERVER", "https://inference.llmops.pl")
 TIMEOUT = 300  # seconds; CPU inference is slow
 
 # ---------- Humbrol → Tamiya lookup ----------
@@ -37,8 +37,9 @@ HUMBROL_TO_TAMIYA = _load_humbrol_tamiya()
 # `- CODE: NAME` list regardless of source layout, so a single tolerant regex parses
 # all of them — no per-format coupling to break when the model varies its output.
 LIST_PROMPT = (
-    "List every paint code visible in this image."
-    # "Output one per line, exactly as `- CODE: NAME`, and nothing else. "
+    # "List every paint code visible in this image."
+    "List all paint codes shown. 'code' = the numeric identifier (e.g. '33'). 'name' = colour name (e.g. 'Matt Black')"
+    # "Output one per line, exactly as `- CODE: NAME`, and nothing else."
     # "For a paint given as a mix ratio, join the codes with + (for example 118+34)."
 )
 
@@ -149,33 +150,13 @@ def call_llama(img_b64: str, prompt: str) -> str:
         }
     elif OUTPUT_MODE == "line":
         payload["grammar"] = LINE_GRAMMAR
-    # stream=True keeps the Cloudflare connection alive via SSE tokens;
-    # without it a 100s+ inference triggers Cloudflare 524 before the response arrives.
-    payload["stream"] = True
     r = requests.post(
         f"{LLAMA_SERVER}/v1/chat/completions",
         json=payload,
         timeout=TIMEOUT,
-        stream=True,
     )
     r.raise_for_status()
-    content = []
-    for line in r.iter_lines():
-        if not line:
-            continue
-        line = line.decode("utf-8") if isinstance(line, bytes) else line
-        if line.startswith("data: "):
-            data = line[6:]
-            if data.strip() == "[DONE]":
-                break
-            try:
-                chunk = json.loads(data)
-                delta = chunk["choices"][0]["delta"].get("content", "")
-                if delta:
-                    content.append(delta)
-            except (json.JSONDecodeError, KeyError, IndexError):
-                pass
-    return "".join(content)
+    return r.json()["choices"][0]["message"]["content"]
 
 # ---------- Parsing ----------
 
@@ -255,7 +236,15 @@ def extract_entries(raw: str, pattern: re.Pattern) -> list[dict]:
         entries = parse_json_entries(raw)
     else:
         entries = parse_entries(strip_parens(raw), pattern)
-    return [e for e in entries if is_paint_name(e.get("name"))]
+    result = []
+    for e in entries:
+        name = e.get("name")
+        # finetune model echoes the code as name when it has no colour name — treat as unnamed
+        if isinstance(name, str) and name.strip() == e.get("code", "").strip():
+            e = {**e, "name": None}
+        if is_paint_name(e.get("name")):
+            result.append(e)
+    return result
 
 
 def expand_codes(entries: list[dict]) -> list[dict]:
@@ -336,9 +325,9 @@ def _run_analysis(image: Image.Image, format_id: str) -> tuple[str, str]:
             codes_str = "<p>(none parsed)</p>"
         return raw, codes_str
     except requests.Timeout:
-        return "Timeout — model took too long.", "<p>(timeout)</p>"
+        return "Timeout — model took too long.", ""
     except requests.RequestException as e:
-        return f"Server error: {e}", f"<p>(error)</p>"
+        return f"Server error: {e}", ""
 
 
 def analyze(image: Image.Image, format_id: str):