llama.cpp Metal bug stalls long contexts

This is less a clean product update than a useful stress signal for local AI infrastructure: long-context, reasoning-heavy workloads are still finding hard edges in consumer Apple GPU scheduling.

• –The log points at llama.cpp's Metal backend during near-full-context prompt processing and context checkpoint churn, not an app-layer opencode issue.
• –A 35B-class quantized MoE model plus roughly 122K prompt tokens on 32GB unified memory is an aggressive setup, so memory pressure and GPU interactivity limits are plausible failure factors.
• –The practical mitigation path is likely smaller context, lower batch/ubatch settings, fewer GPU layers, a newer llama.cpp build, or filing a minimal reproducible issue upstream.
• –For developers betting on local coding agents, this is a reminder that "fits in RAM" is not the same as stable under long-running agentic workloads.