jinja input marking is never consumed by llama-server #26532

#18462 added jinja::string with an is_input flag so special tokens coming from user content can be told apart from ones the template emitted, and listed "implement input marking on llama-server" as a follow-up. That follow-up doesn't seem to have landed.

At 3581ba0 the flag is produced but nothing reads it:

• common/chat-auto-parser.h:73 defaults mark_input = true
• common/chat.cpp:924 calls global_from_json(ctx, inp, inputs.mark_input)
• tools/server/ has zero references to is_input or mark_input
• all four tokenize_input_prompts calls in server-context.cpp pass parse_special = true

So a two-message conversation with a ChatML payload in the user content tokenizes to the same ids as a real system turn (151644, 8948 on qwen2.5). Nothing downstream can distinguish them.

I know special token injection is a known class and this is a documented TODO rather than a surprise. What made me want to bring it up, transformers has no working equivalent. apply_chat_template(tokenize=True) forges the turns, split_special_tokens isn't a parameter it accepts, and on the two-step path that flag strips the template's own delimiters too, leaving no role structure at all. Keeping the template's delimiters while dropping the content's needs origin tracking, and is_input looks like the only implementation of it anywhere.

Is the follow-up still on the roadmap, or was it dropped deliberately? Happy to help if it's just unowned.