Per-render fingerprint of the cacheable prefix — the segments a provider
caches as a stable prefix (system blocks + history turns). Excludes
state_turn (the volatile uncached tail) and token_count metadata (not on the
wire). This is the metrics-first instrument (P0-A) behind the optimization work:
two renders share a reusable KV / prompt-cache prefix iff their system hashes
match and one’s turn_hashes is a prefix of the other’s. Pure and derived —
never stored in snapshots, session logs, or event logs.
Render with Layer-4 read-time projection driven by handles: tool results whose handle is
Collapsed render as previews (originals untouched), freeing budget for more recent turns.