Complete procedures with a way back.

Build a repeatable capacity-and-quality matrix for local GGUF models instead of choosing from a model leaderboard or a vendor logo. Inventory the real machine, calculate a safe memory envelope, pin Qwen3 4B, 8B, and 14B candidates, run llama.cpp b9637 with the appropriate CPU, Metal, CUDA, ROCm, or Vulkan backend, and select the smallest candidate that passes the workload at the required concurrency.

Correlate kernel I/O symptoms, SMART or NVMe health data, self-test history, stable device identity, storage-stack membership, and application evidence before replacing a Linux data drive through the layer that actually owns redundancy.