docs(token-usage): state the actual speed_inferred and pricing_catalog semantics

speed_inferred means the tier was not recorded in the transcript
(resolved from configuration or the standard default) — not
specifically config-derived, which misread Codex default-standard
buckets and Claude's never-set flag. The bucket's pricing_catalog is
the SQL MAX (lexicographically greatest) id among its catalog-priced
entries, not the "latest": buckets price under one catalog in
practice, and a mixed bucket (refresh plus daemon restart mid-bucket)
is identified approximately.

The fast/long-context e2e assertion now derives its expected cost from
the embedded catalog's gpt-5.6-sol rates instead of a hard-coded
$4.572, so the weekly snapshot refresh cannot break integration CI on
a price change.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>