The prompt-cache layout

Measured on Coral: the provider orders instructions, then tools, then input items; any change to instructions or tools is a full miss, and so is an insertion before a long history; prompt_cache_key partitions strictly, so one is always sent. The layout: instructions holds only the static host system prompt; tools are sorted, a stable superset (under U2); the first input item is the frozen workspace snapshot; then the trim summary (one accepted miss when it changes); then the append-only history; the last item is the single volatile tail. Coral’s cache semantics are recorded in its provider profile; the core layout stays provider-agnostic. Reuse on glm-5.3 is marked unverified.