Aireceipts

Patch: Claude Code receipts were over-reporting cost by ~2.5–3× and are now correct.
Claude Code writes one assistant transcript record per content block of a response, and
every record of that response repeats the same message.id and a byte-identical usage
snapshot. The adapter had counted each record as its own billed turn, multiplying a
response's tokens by its block count. Turns are now deduped by message.id — the first
record of a response books its usage once, later records only add their tool calls — so
turnCount and every $ reflect real billed responses (PR #196). After upgrading, every
Claude Code receipt you render shows lower dollars and fewer turns than 0.7.0 showed for
the same session: a real local session dropped from $102.97 to $36.69 (2.81×), and one PR
slice from $5.17 to $1.61. This was a silent over-report — the first known arithmetic
case of a receipt exceeding the true floor rather than under-reporting it. Codex, opencode,
and Cursor are unaffected (Codex reconciles against the rollout's own cumulative total and
was re-verified correct to the cent). Synthetic-fixture output is byte-identical (goldens
unchanged); the summary-cache version was bumped (2 → 3) so stale inflated totals recompute
on the next run. Receipts posted on this repo's own past PRs carry per-session correction
blocks with exact recomputed figures. Full audit and evidence: PR #196,
docs/cost-model.md, and the 2026-07-08 addendum in
docs/internal/cost-attribution-evidence.md.