Pin llama.cpp fa226f9: ZAYA chat templates tokenize; HRX all-NaN router row is loud (engine#123)

[bong-water-water-bong]

Moves third_party/llama.cpp from 895d63f to fa226f9, which brings in two fork PRs.

llama.cpp #27: ZAYA chat templates tokenize.

• Every ZAYA GGUF we converted stored <|im_start|> (105) and <eos> (1) as ordinary tokens. llama.cpp only matches control and user-defined tokens in prompt text, so each chat turn's <|im_start|> reached the model as seven text tokens, and ZAYA1-8B chat answers went off-template ("12 + 30" gave "22").
• The converter now marks every token tokenizer.json calls special as a control token. The ZAYA GGUFs on our Hugging Face org were fixed in place (2 bytes of metadata each; tensors unchanged): 8B, VL-8B, base, reasoning-base, with the 74B finishing now.
• Checked: <|im_start|>user now tokenizes as [2, 105, 2364], as transformers does. With thinking off, our F16 matches Zyphra's BF16 model on the reference prompts: "7 + 8" gives "25", which is the model's own answer, not ours.

llama.cpp #26: an all-NaN MoE router row is loud instead of silently decoding a wrong expert (engine#123 follow-up).

• Checked: perplexity on HRX0 is identical to before: Qwen3-Coder-30B 17.2526, ZAYA1-8B 25.2474 (8 chunks). Chat answers are right on HRX0 and Vulkan0 with no false aborts.

ZAYA1-8B end to end through 1bit serve at this pin:

• serve_e2e passes on vulkan, hrx and auto.
• Streaming, 4 concurrent requests and a 5,400-token prompt all work on Vulkan and HRX.

Also here: the registry is regenerated for the pin (counts unchanged). docs/hrx.md covers #26, and docs/vulkan.md warns that ZAYA GGUFs converted before today need re-downloading.

🤖 Generated with Claude Code

[context7]

Docs7 for 1bit-monster/engine

Result	Status	Action
Deployment	➖ Not used	—
Content review	➖ Did not run. This site has no agent runs available this month. Wait for the monthly reset or check your Docs7 plan.	—

Commit 7c37201

[github-actions]

PR Reviewer Guide 🔍

Here are some key observations to aid the review process:

⏱️ Estimated effort to review: 3 🔵🔵🔵⚪⚪

🧪 No relevant tests

🔒 No security concerns identified

⚡ No major issues detected