Tokens saved

Two hard facts: (1) prompt-cache savings across every real Claude Code session on this machine -- cache_read_input_tokens is tokens that were served from cache instead of freshly reprocessed, a directly measured number, no guessing; (2) the real token cost of what every rote play-run workspace's own steps processed internally (from rote's own archive, read via the duckdb Python package against the parquet files directly -- no DuckDB CLI required), broken down per play with call/run counts so it is never mistaken for a single-run cost. On top of those two measured facts it shows a live compression example: how many raw tokens THIS play's own steps just processed, versus how few tokens its own final human-readable report actually costs to read -- a genuine, reproducible, measured ratio computed fresh every run, not a historical claim. An AI narrates what the numbers mean in plain words, but is explicitly told never to invent or esti