fix(celery+lmstudio): PYTHONPATH for forked workers + qwen3.5-9b thinking model support

Three root causes fixed:

1. docker-compose.yml — celery-worker and celery-beat missing PYTHONPATH=/app
   ForkPoolWorker processes inherit sys.path with '' (CWD), but fork can change
   CWD so lazy imports like `from db.session import ...` raised ModuleNotFoundError.
   Adding PYTHONPATH=/app ensures /app is always explicit in the path.

2. provider_config.py — lmstudio SUPPORTS_JSON_MODE set to False
   LM Studio only accepts response_format type 'json_schema' or 'text', not
   'json_object'. GenericOpenAIProvider now omits response_format for lmstudio
   and relies on parse_classification() fallback (same path as Gemini).

3. generic_openai_provider.py — max_tokens raised 1024→4096 (classify) / 256→1024 (suggest)
   Qwen3 thinking models (like qwen/qwen3.5-9b) consume reasoning tokens from the
   same budget. With max_tokens=1024 the model exhausts the budget in the thinking
   trace, leaving content=''. 4096 gives room for both thinking and the JSON output.

Also updates lmstudio PROVIDER_DEFAULTS model to qwen/qwen3.5-9b (confirmed loaded
in LM Studio) and context_chars to 32_000 (Qwen3 context window).

Tested: classify(invoice_text, ['Finance','Legal','HR']) → topics=['Finance'] ✓
Backend tests: 376 passed ✓

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

backend/ai/generic_openai_provider.py

@@ -60,7 +60,7 @@ class GenericOpenAIProvider(OpenAIProvider):
         )
         create_kwargs = dict(
             model=self._model,
-            max_tokens=1024,
+            max_tokens=4096,
             messages=[
                 {"role": "system", "content": system_prompt},
                 {"role": "user", "content": user_msg},
@@ -87,7 +87,7 @@ class GenericOpenAIProvider(OpenAIProvider):
         )
         create_kwargs = dict(
             model=self._model,
-            max_tokens=256,
+            max_tokens=1024,
             messages=[
                 {"role": "system", "content": system_prompt},
                 {"role": "user", "content": user_msg},

backend/ai/provider_config.py

@@ -83,16 +83,17 @@ PROVIDER_DEFAULTS: dict[str, dict] = {
     },
     "lmstudio": {
         "base_url": "http://host.docker.internal:1234/v1",
-        "model": "gemma-4-e4b-it",
+        "model": "qwen/qwen3.5-9b",
-        "context_chars": 8_000,
+        "context_chars": 32_000,
     },
 }
 
 # Whether the provider honours response_format={"type": "json_object"}.
 # Gemini's OpenAI-compat endpoint does NOT support the string form (D-02/D-03).
-# Ollama and LMStudio accept the parameter but some models ignore it — the
+# LMStudio only accepts "json_schema" or "text" (not "json_object") — set False
-# GenericOpenAIProvider always wraps the raw response with parse_classification()
+# so GenericOpenAIProvider omits response_format and relies on parse_classification().
-# regardless, so they are left as True (the parameter is still sent).
+# Ollama accepts the parameter but some models ignore it — left True; the
+# GenericOpenAIProvider always wraps the raw response with parse_classification().
 SUPPORTS_JSON_MODE: dict[str, bool] = {
     "openai": True,
     "anthropic": True,
@@ -103,5 +104,5 @@ SUPPORTS_JSON_MODE: dict[str, bool] = {
     "openrouter": True,
     "mistral": True,
     "ollama": True,
-    "lmstudio": True,
+    "lmstudio": False,
 }

docker-compose.yml

@@ -102,6 +102,7 @@ services:
       - CLOUD_CREDS_KEY=${CLOUD_CREDS_KEY}
       - SECRET_KEY=${SECRET_KEY}
       - PYTHONDONTWRITEBYTECODE=1
+      - PYTHONPATH=/app
     labels:
       logging: "promtail"
     volumes:
@@ -134,6 +135,7 @@ services:
       - MINIO_BUCKET=${MINIO_BUCKET}
       - REDIS_URL=${REDIS_URL}
       - PYTHONDONTWRITEBYTECODE=1
+      - PYTHONPATH=/app
     extra_hosts:
       - "host.docker.internal:host-gateway"
     command: celery -A celery_app beat --loglevel=info --schedule /tmp/celerybeat-schedule