fix(billing): apply model pricing policies to account stats cost

[qiujun8023]

Problem

For DeepSeek traffic served during the official peak window (01:00–04:00 and 06:00–10:00 UTC, weekdays only), usage_logs.account_stats_cost is recorded at off-peak rates and understates the real cost by up to 2x.

User-facing billing is unaffected — it goes through calculateTokenCost, where the peak multiplier is applied. Only account cost statistics and the dashboard's account_cost are wrong.

Root cause

Priority 3 of resolveAccountStatsCost (model pricing file) keeps its own simplified unit price × tokens implementation in tryModelFilePricing instead of reusing the billing pipeline:

cost := float64(tokens.InputTokens)*pricing.InputPricePerToken +
    float64(tokens.OutputTokens)*pricing.OutputPricePerToken + ...

Every pricing feature added on the billing side therefore has to be mirrored here by hand, and this has already been patched twice for exactly that reason — 25e6688a for long-context pricing, 9261dd77 for service tiers. The DeepSeek peak/off-peak rates introduced in b5827cfd live inside calculateTokenCost, and this raw-multiplication path never reaches that function, so they have never taken effect for account cost.

Fix

Route the branch through CalculateCostWithServiceTier and drop the duplicated implementation.

channelPricing stays nil, so priority-3 semantics are unchanged: model-file pricing only, no channel overrides. Off-peak results are identical to the previous arithmetic — with no interval pricing and no extra policies, calculateTokenCost reduces to the same multiplication. Returning nil for missing pricing or zero cost is preserved through the err and TotalCost <= 0 checks.

Verification

• go build ./... passes.
• go vet ./internal/service/ reports no issues.
• go test ./internal/service/ passes for the whole package (124s), including the existing deepseek_pricing_test.go and account_stats_pricing_test.go cases.
• gofmt reports no diff.

I did not run the backend integration tests, frontend suite, vulnerability scan or deployment checks — this touches a single backend file and no frontend or deployment configuration.

Tests

I did not add a regression test, and I want to be upfront about why. tryModelFilePricing takes no billing timestamp, so the peak multiplier falls back to time.Now(). Off-peak runs produce identical results before and after this change, which means any test I write would pass against the broken code for most of the day. Making it deterministic requires threading a pricing timestamp through this path, which grows the diff well beyond this one file. Happy to do that if you would rather the fix land with a test.

Searched open PRs before submitting and found nothing else fixing account cost resolution or DeepSeek peak pricing. #6353 (long-context tier pricing moving to a data-driven catalogue) is in the same billing area but changes billing_service.go, so there is no file overlap.

🤖 Generated with Claude Code