Assumptions behind estimate_saved_cost. Token reduction is not the
same thing as billed-cost reduction (arXiv:2607.12161): with provider
prompt caching, a tool-output token is written to the cache once at a
write premium and then re-read at a steep discount on every subsequent
turn. A saved token therefore avoids one cache write plus N cache
reads — not one list-price input token.
Estimate the billed cost avoided by tokens_saved under a
prompt-cached serving model. This deliberately models the
caching-enabled case, which is the conservative one: without caching
every re-read bills at full input price and the figure would be much
larger. Direct token cost only — trajectory effects (extra agent
turns caused by missing detail) are not modeled here; the regret
metrics in sqz stats exist to watch for those.