convert: ZAYA special tokens are CONTROL, so chat templates tokenize

[bong-water-water-bong]

Bug: in every ZAYA GGUF our converter wrote, <|im_start|> (id 105) and <eos> (id 1) had token type NORMAL, although tokenizer.json marks both special: true (LlamaHfVocab left them normal; both are also in the base vocab). llama.cpp matches special-token text in a prompt only for CONTROL and USER_DEFINED tokens. So every chat turn's <|im_start|> reached the model as seven text tokens instead of token 105:

llama (before): [2, 236820, 236909, 548, 236779, 3041, 111038, 2364]   <|im_start|>user
transformers:   [2, 105, 2364]

Chat answers were off-template: "What is 12 + 30?" gave "22". The teacher-forced checks never saw this, because they feed token IDs directly.

Fix: every token tokenizer.json lists as special becomes CONTROL, for ZAYA and ZAYA1-VL.

Checked:

• Converting ZAYA1-8B with this commit gives the published GGUF's 1,283 tensors unchanged. The token types match the published file with tokens 1 and 105 set to CONTROL. The only other differences are general.* model-card fields (base model, licence, name).
• llama-server now tokenizes <|im_start|>user as [2, 105, 2364].
• With thinking off, on questions where Q4_K_M and BF16 both land on the same token, our F16 answers match Zyphra's model in transformers (BF16): "7 + 8" → "25" (the model's own answer), "100 + 23" → "123", "17 × 23" → "391".

The published GGUFs on Hugging Face get the same two-byte metadata fix in place, so they needn't be reconverted.

🤖 Generated with Claude Code