fix(pricing): key the fast lane per model in a slash-combined row

[krayong]

Fixes the bug behind #31, which is pure regression: its whole diff was the deletion of fast_mult, with no rate change at all.

The bug

The published fast-mode table now bills one shared lane for several models by naming them in a single cell:

["Claude Opus 5 / Claude Opus 4.8", "$10 / MTok", "$50 / MTok"]

buildFastInput keyed the map on the raw cell text, so the lookups for "Claude Opus 5" and "Claude Opus 4.8" never matched. fast_mult was therefore never derived, and mergeRows replaced each row wholesale and dropped the field.

Downstream, fast_mult defaults to 1 (pricing.ts:56) and gates the speed multiplier (pricing.ts:188), so fast-mode messages on Opus 5 and 4.8 priced at ×1 instead of ×2 — reporting half their true cost.

Changes

• buildFastInput splits the cell and keys each model name separately.
• mergeRows carries an established fast_mult forward when a model has dropped off the fast table. Opus 4.7 has, and transcripts recorded while that lane was offered still need ×6 to price correctly. A model still listed is re-derived normally, and deleting the field by hand still sticks.
• The refresh workflow now writes a minor changeset next to pricing.json. That file ships inside the published ccsidekick package, so CI's changeset gate failed on every automated pricing PR. A fixed filename keeps re-runs idempotent, since the branch is recreated from main each run.

Why the tests were green while production drifted

The fixture pinned the old single-name table shape, so it tested the parser against a page that no longer exists. It now mirrors the live slash-combined row.

Verification

• refresh-pricing.test.ts: 8 pass. Both new tests were confirmed failing first (Expected: 2, Received: undefined).
• bun run test: 1076 pass, 0 fail. The existing ×6 assertion for Opus 4.7 still passes, because the carry-forward leaves pricing.json byte-identical.
• refresh-pricing.ts --check against the live page: in sync, with pricing.json unchanged — confirming chore(pricing): sync pricing.json with published rates #31's diff was entirely spurious.
• The changeset gate was simulated in a scratch worktree by committing a pricing.json edit plus the generated changeset: changeset status --since=origin/main exits 0.
• typecheck, lint, format:check clean.

Follow-up

#31 should be closed rather than merged.

🦇 Generated by Penny-One at Batcave