serve: refuse a Q4_0 drafter next to a Hadamard-rotated file

[bong-water-water-bong]

Next to a Hadamard-rotated file, 1bit serve now refuses a drafter or MTP head that holds plain Q4_0 tensors, with an error that says why.

Why: GGML_Q4_0_HADAMARD applies to every Q4_0 matmul in the backend process, the drafter's included. A plain-Q4_0 drafter would read its activations rotated and accept nothing. The leaderboard run hit exactly this with a Q4_0 DFlash2 drafter.

What changes:

• app/gguf_meta.h gains gguf_tensor_type_count, which reads only the GGUF header and tensor infos.
• serve checks --dflash and --mtp files against it.
• docs/serve.md says so.

Tests: tests/hadamard_route.sh passes 15/15, including a Q4_0 drafter (refused) and a Q8_0 one (accepted). I also checked it against a real 27B Q4_0 file, which is refused, and against the real Q8_0 DFlash2 drafter, which is accepted.

🤖 Generated with Claude Code

[context7]

Docs7 for 1bit-monster/engine

Result	Status	Action
Deployment	➖ Not used	—
Content review	✅ Passed. No problems found.	View findings

Commit e2248c6

[github-actions]

PR Reviewer Guide 🔍

Here are some key observations to aid the review process:

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

🧪 PR contains tests

🔒 No security concerns identified

⚡ Recommended focus areas for review Incorrect tensor type check The code checks for Q4_0 tensors in draft files using gguf_tensor_type_count(draft, 2), but the function gguf_tensor_type_count is designed to count tensors of a specific GGML type. The check gguf_tensor_type_count(draft, 2) > 0 will return true if any tensor of type 2 (Q4_0) exists in the file, but it does not verify if the tensor is actually Hadamard-rotated. This could lead to false positives where a non-Hadamard-rotated Q4_0 tensor is incorrectly accepted. if (!draft.empty() && gguf_tensor_type_count(draft, 2) > 0 && !hadamard_q4_0(draft)) Potential buffer overflow in skip_str The function skip_str uses n > (1u << 20) to check for oversized strings, but this check is not sufficient to prevent potential buffer overflows. If n is close to the maximum value of uint64_t, the subsequent f.seekg(static_cast<std::streamoff>(n), std::ios::cur) could cause undefined behavior or overflow. The check should be more robust to prevent such issues. if (!rd(&n, 8) || n > (1u << 20)) return false;

[bong-water-water-bong]

Review (PR-Agent duty). Looks good.

• gguf_tensor_type_count follows the GGUF layout: the KV scalar sizes match types 0-12, string arrays are walked, and each tensor info is name, n_dims, 8-byte dims, u32 type, u64 offset. Anything unreadable returns -1, so the check is skipped rather than failing a launch on an odd file.
• A 1 MiB cap per string is safe: chat templates and vocab entries are far below it, and walking 248k vocab strings is header-only work.
• The rule is the right one. The rotation is process-wide (GGML_Q4_0_HADAMARD), so plain Q4_0 in the drafter or MTP file can't be read correctly, and a rotated drafter is still allowed through hadamard_q4_0(draft).
• The test covers both sides (Q4_0 refused, Q8_0 accepted), the docs are updated, and checks pass.

Merging.