llama.cpp restores reasoning across chat turns

llama.cpp build b10219 fixes one multi-turn state bug: the CLI displayed streamed reasoning but stored only ordinary assistant content, so --reasoning-preserve could not re-inject that reasoning on later turns. The project’s official release ships the correction across local inference targets including CUDA 12.4, CUDA 13.3, ROCm 7.2, and OpenVINO 2026.2.1. Teams implementing the full-response preservation that reasoning models require should pin or test b10219 or later: a runtime can appear stateful while silently changing later tool choices because its reasoning history vanished.