llama.cpp context shift stalls on RAG

This looks like budget math, not model intuition. Context shift is a KV-cache reuse trick, so if the prompt plus RAG payload leave no safe runway for the completion, the runtime has to rebuild state or clamp generation.

• –n_predict / max output affects whether the server can reserve space for a shifted cache.
• –llama.cpp docs call this "rotating context management" and expose --cache-reuse, so the behavior is an engine policy, not a magical model feature.
• –Big retrieved chunks can destroy the suffix overlap needed for cache reuse.
• –Sliding-window-attention models and backends can be brittle here or outright disable shifting.
• –If latency matters, shrink retrieved chunks or raise ctx-size instead of expecting the model to absorb everything.