generation_config.json shipped eos_token_id=248044 (<|endoftext|>) only, dropping 248046
(<|im_end|>) -- the chat-format end token this model actually emits. Generation therefore did not
terminate on <|im_end|> and ran until it hit a length cap.
Measured cost with the repair applied at eval time, SWE-bench-Verified pass@1 n=500, stock pi
harness: dpsk-v4-flash 10.12% -> 13.40%.
Only generation_config.json changed; weights are untouched.
