llm tokenization trailing space: one blank changed the answer

I am seeing a quiet bug in prompt templates. The template ends with a trailing space before the answer slot. Most tokenizers attach the leading space to the next word. So the model usually sees " Paris" as one common token. After a bare space, it must start from a boundary it rarely saw. Output quality drops in ways that look like model weakness. The guidance library calls the fix token healing. It backs up one token and constrains the next token to match the removed text. The cheaper fix is to strip trailing whitespace before every call. The trade-off is real, because some formats need the space. The failure mode is silent. Nothing errors, and eval scores drift down a little. I measure it directly. I diff the token ids at the end of the prompt across template versions. I also run the same eval with and without the trailing space. A gap between those runs points at the template, not the model.