Raise ZeroGPU duration 180->300s: 180 was killing longer completions mid-generation

app.py

@@ -105,9 +105,14 @@ print(
 )
 
 
-# Raised alongside MAX_NEW_TOKENS above -- 2048 tokens of greedy decoding needs
-# more wall-clock headroom than the SQL-only completion this was tuned for.
-@spaces.GPU(duration=180)
+# This is the GPU time the request is allowed, not a hint: ZeroGPU kills the
+# task the moment it runs over, and the caller sees only "GPU task aborted".
+# At 180s that happened intermittently -- short completions finished, longer
+# ones were killed -- which reads as a flaky Space rather than a budget that
+# is simply too small. Generation has to cover a 14B model in 4-bit emitting
+# a JSON plan and then the SQL, so the budget is set to the maximum rather
+# than trimmed to the average.
+@spaces.GPU(duration=300)
 def generate_sql(question: str) -> tuple[str, str]:
     """Generate a plan+SQL completion for a natural-language question.