[coco-alen]

Closes #646

/v1/chat/completions passed a developer message (sent by Codex and other current OpenAI clients) straight to the chat template. Templates that don't name the role handle it badly, and two of the three failures are silent:

template	developer on main
nvidia/Qwen3.8-Flash-Next-NVFP4	TemplateError: Unexpected message role. (400)
Qwen/Qwen3-0.6B	dropped: the instruction never reaches the model
Qwen/Qwen3-Coder-30B-A3B-Instruct	rendered as `<

TokenizeManager._render now renders developer as system when the chat template does not mention developer. Templates that name the role (gpt-oss) get it unchanged. The DSV4 encoder path returns before this and keeps its own handling. The mapping copies the message, so the request is not mutated. /v1/responses already folds developer into the system message, so this only changes the chat completions path.

Not covered here: a request with both a system and a developer message now behaves like one with two system messages. Templates that require a single leading system message (Qwen3.8-Flash-Next) still reject it, now with System message must be at the beginning. Merging the two would be a separate change.

Testing

RTX 6000 Ada (48 GB), driver 595.99.02, torch 2.14.1+cu130, transformers 5.18, Python 3.12. Tokenizer-only check, no model weights: TokenizeManager.render_prompt with [{"role": "developer", "content": "Answer in French."}, {"role": "user", "content": "hi"}] against each checkpoint's real tokenizer and chat template:

tokenizer	main	this branch
nvidia/Qwen3.8-Flash-Next-NVFP4	TemplateError: Unexpected message role.	same prompt as the system spelling
Qwen/Qwen3-0.6B	developer text missing from the prompt	same prompt as the system spelling
Qwen/Qwen3-Coder-30B-A3B-Instruct	`<	im_start
openai/gpt-oss-20b	same prompt as system	unchanged

• New tests/tokenizer/test_tokenize.py::test_developer_role_renders_as_system_unless_the_template_names_it. With the fix reverted, the case for a template without developer fails ({'role': 'developer', ...} != {'role': 'system', ...}). With the fix, both cases pass.
• pytest tests/tokenizer tests/server: 755 passed, 24 skipped (needs_weights).