Added#
The day series reaches the markdown. The spend-per-day table the peak
sentence summarises — day, exact dollars, calls, and the biggest label of the
day — for the CI summary and PR comment readers who want to see the week.
Capped at the most recent 14 days with the earlier ones counted out loud
(silent truncation reads as "covered everything"), absent for a single day
because one row is the total again, and the full series still rides --json
as spendByDay.
The price table's age, said out loud when it matters. Every dollar a
profile prints uses the bundled price table, and the one fact that silently
invalidates all of them is a table the provider has re-priced since. Past 45
days — the threshold is in the sentence — the terminal, the markdown, the MCP
and the web all say so loudly: the review date, the age, and that the report
is wrong by exactly whatever changed, with --pricing-live as the fix.
--json carries pricing.lastReviewed/pricing.ageDays always, as
provenance. The tests pin the rule rather than the calendar: whether the line
appears is derived from the table's own date at run time, so a freshly
reviewed table asserts the opposite behaviour and passes the same suite.
--max-cache-loss-usd, the third money gate — and it reads the worst case
on purpose. Exit 1 when caching added more than the limit to the bill. When
the log did not record which write TTL was paid, the settled figure and the
1-hour worst case can straddle the limit, and a gate reading the flattering
half would pass exactly the bills it exists to catch — so the gate reads the
ceiling and its failure message says which claim fired: a settled loss
(exact, the same tokens at the published input rate), or a ceiling only the
missing cache_creation field can settle. Fires under --json too; wired
into the Action as max-cache-loss-usd and self-tested in CI on a +$1.25
loss ($6.25 of 5-minute writes nothing reads back, against $5.00 as plain
input). The read-the-worst-case guard is mutation-tested.