What changed

Signed prerelease b12199 sanitizes invalid UTF-8 before PEG-parsed tool-call arguments reach downstream agents. Later b12202 aligns BF16 and F32 accumulation choices with reference implementations for GLM4, MiMoVL, Qwen3-TTS, DeepSeek V4 and MiniMax-M3 paths and adds BF16 flash-attention tests.

Why it matters

Malformed tool-call arguments can break structured parsers or terminate an agent loop, while unintended accumulation precision can change model behavior or hardware fallback. The builds create exact artifacts for testing both boundaries.

What to do

Canary b12202 against stable 0.9 with captured malformed Unicode tool calls and fixed prompts for every affected model family; compare parsed arguments, outputs, latency and backend selection before promoting any rolling build.

Caveat

These are rolling prereleases, not a new stable llama.cpp version. The repository supplies signed artifacts and tests but no independent cross-backend reproduction, and later same-day builds continued refactoring the multimodal code path.