Spot the tokenization boundary bug in this manual prompt-building code.

The bug is the missing spaces at the boundaries between the role labels and the content. There are two of them. In a BPE tokenizer, ` word` (with a leading space) and `word` (without) are different tokens with different IDs. When the code writes `'system:' + system_prompt`, the first word of `system_prompt` is glued directly onto the colon. So the tokenizer sees something like `system:Hello`, and the `Hello` part lands on its no leading space variant, which is not how the word normally appears in text. The same thing happens at `'user:' + user_msg`: the first word of `user_msg` gets glued on with no space. That is the second boundary error. The model was trained on naturally spaced text where words come after spaces. Feeding it these smushed boundary tokens gives it IDs it rarely saw, which can subtly throw off its behavior. The clean fix is to add the spaces, but the truly safe fix is to stop hand-building the prompt and call `apply_chat_template()`, which handles all the spacing and role formatting for you.

**The bug is two missing word-boundary spaces** caused by raw string concatenation. In BPE, ` word` and `word` are distinct tokens, so where you glue strings matters. First boundary: `'system:' + system_prompt` produces `system:Hello...`, so the first word of the system prompt resolves to its no leading space variant instead of the spaced form the model expects. Second boundary: `'user:' + user_msg` does the same to the first word of the user message. Two boundaries, two mismatches. Why it matters: the model trained on naturally spaced text, where role markers and content are separated by whitespace. The concatenated form yields token IDs from off the training distribution at exactly the points the model uses to parse turn structure. That can nudge logits and degrade instruction following in ways that are hard to trace. The local fix is to restore the spaces, for example `'system: ' + system_prompt` and `' user:' + user_msg`. The real fix is to never assemble chat prompts by hand. Use `apply_chat_template()`, which applies the model's own Jinja chat template, inserts the correct special tokens, and gets spacing right for that specific model. Hand-rolled formatting also breaks the moment you switch models, since each one expects different markers.