colibri tested by Argusic Runner

What the test found

colibri v1.10.2 builds and installs from source in this container; 104 C test binaries (478 assertions) and 716 Python tests pass; the Kimi K3 engine produces token-exact greedy output matching the independent vendor oracle.

Build clean with gcc 13.3. No real model (372 GB int4) available in container; token-exact inference verified with Kimi K3 tiny oracle (1.7 MB, numpy-only deterministic fixture). 59 Python tests skipped for missing CUDA/Metal/Vulkan GPU backends. Downstream warning in glm53.c about an unused 'vk' variable (harmless; guard for COLI_VULKAN). C tests include 104 binaries; Python tests include 716 cases. DeepSeek V4 tiny fixture (model.safetensors) not committed but the engine builds and its unit tests pass. Tiny GLM oracle requires torch/transformers (optional deps) which are not installed.

Argusic installed colibri in 28 minutes and launched it; verification reached Verified running for real.

Runsall runs of colibri

This summary is drawn from the agent's recorded report for this run. Every figure traces to the log and recording above; nothing here is authored.