litellm: add local_qwen2.5-14b, the structured-output tier

paperless-ai needs a model that holds strict JSON without being asked,
because it does not send response_format. llama3.1:8b breaks the shape
and qwen2.5:72b cannot answer inside its client timeout. 14B clears both.

Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>

vars/main.yml

@@ -94,6 +94,16 @@ litellm_models:
     backend: ollama_chat/llama3.1:8b
     api_base: "{{ anvil_ollama_base }}"
     api_key: ollama-no-auth
+  # The structured-output tier, added 2026-09-12. llama3.1:8b answers fast but
+  # will not hold strict JSON: paperless-ai got "Expected ',' or '}' after
+  # property value in JSON at position 263" on a 1,921-token reply. The Kestra
+  # flows escape this by setting response_format={"type":"json_object"};
+  # paperless-ai does not, so it needs a model that keeps the shape unprompted.
+  # 14B is the smallest that does, and at ~9 GB it co-exists with anything here.
+  - name: local_qwen2.5-14b
+    backend: ollama_chat/qwen2.5:14b
+    api_base: "{{ anvil_ollama_base }}"
+    api_key: ollama-no-auth
   # Shadow alias SOLELY for paperless-ai's hardcoded gpt-4o-mini validation
   # probe (see the "Narrow exception" note above). Routed to free local
   # llama3.1:8b so the probe costs nothing. Do not point real traffic here.