fix(pricing): normalize Bedrock dot-form model names and add opus-4-7

[estelledc]

Summary

Fixes ~56% cost under-counting on Bedrock-proxy users (and a smaller hit on direct-API Opus 4.7 users). Independent of #16 — that one fixes token counts, this one fixes the per-token rates.

Two bugs converging

PRICING_TABLE uses dash-form patterns (opus-4-6, haiku-4-5). Two cases never match anything in the table:

1. Bedrock proxy emits dot form. aws.claude-opus-4.7 does not contain opus-4-6 or opus-4-7 (or anything else dash-separated in the table). Every Bedrock call falls to FALLBACK.
2. opus-4-7 is missing from the table. Even direct-API users on the newest Opus fall to FALLBACK.

FALLBACK_PRICING was hard-coded Sonnet ($3/$15/$0.30). So:

Model	Old behavior	Should be	Effect
aws.claude-opus-4.7	Sonnet $3/$15	Opus $5/$25	under by ~5/3
aws.claude-opus-4.6	Sonnet	Opus	under by ~5/3
aws.claude-haiku-4.5	Sonnet	Haiku $1/$5	over by 3× (smaller volume)
aws.claude-sonnet-4.5	Sonnet	Sonnet	accidentally correct
claude-opus-4-7 (direct API)	Sonnet	Opus	under by ~5/3

Net on a real Bedrock-proxy archive (28k events, ~12.7k unique messages, 8 active days):

	total cost
Pre-fix CCMeter	$1,718
Post-fix CCMeter (this PR, simulated)	$3,007
Reference (CCCostMonitor / LiteLLM)	$3,007

(Token totals are unaffected; this PR only changes per-token rates. PR #16 separately fixes token counts on the same proxy environment.)

Fix

model_pricing now does a two-stage lookup:

1. Normalize the input — lowercase, replace . with - — then substring-match against PRICING_TABLE. A single dash-form table now covers both name styles.
2. Family fallback on miss — if the normalized string contains opus, haiku, or sonnet, use the latest known pricing for that family. Truly unknown strings still default to Sonnet, preserving prior behavior for non-family inputs.

Adds opus-4-7 to the explicit table.

The family-default change is the load-bearing piece: it means a future Opus release we don't yet know about silently bills at Opus rates instead of Sonnet rates. Under-billing Opus traffic was the larger risk; over-billing a hypothetical future Sonnet (which would still match sonnet substring anyway) is not.

Cache schema bump

Bumped CURRENT_SCHEMA_VERSION 2 → 3. Existing v2 caches on user disks were populated by the buggy pricing and would otherwise freeze the under-counted cost in place. The existing "Cache rebuilt" banner handles the one-time recompute UX with no manual intervention.

Note for the maintainer: PR #16 also bumps schema to v3 for an independent reason (token-count fix). If both PRs land, the second-merged one needs a trivial bump to v4 (or we collapse both bumps into one v3 with a combined comment). The bump itself is just a one-line change in src/data/cache.rs.

Tests

12 new tests in src/data/models.rs::tests covering:

• Direct-API dash form (Opus 4-6, 4-7, Sonnet, Haiku 4-5, legacy Opus 4-1, 3-5-haiku)
• Bedrock dot form (Opus 4-7, 4-6, Haiku, Sonnet)
• Family fallback for unknown versions (claude-opus-9-9 → Opus, not Sonnet)
• Case-insensitive lookup (uppercase ARN-style ids)
• Truly-unknown strings still default to Sonnet (preserves prior contract)

All 64 prior tests still pass. The 2 pre-existing rate_limits failures on main are unrelated to this change.

Test plan

• cargo test models:: → 15 pass (3 baseline + 12 new)
• cargo test → 64 pass (same 2 pre-existing rate_limits failures as on main)
• Real-data sanity: post-fix simulated cost on a 28k-event Bedrock archive matches the LiteLLM-pricing reference exactly
• Reviewer to confirm direct-API users on Sonnet/Haiku see no regression (covered by prices_direct_sonnet_dash and prices_direct_haiku_4_5_dash)

🤖 Generated with Claude Code