Pin TRL <0.16 for unsloth 2025.11.1 GRPO API compatibility

requirements-training.txt

@@ -1,6 +1,6 @@
 # Training deps (torch, unsloth installed separately in Dockerfile)
 transformers>=4.51.3,<5.0
-trl>=0.12.0,<1.0
+trl>=0.12.0,<0.16
 peft>=0.13.0
 accelerate>=1.0.0
 datasets>=3.0.0