fix(usage): repair cost analytics and display precision
- Automatic OpenRouter pricing sync with cost backfill for traces/snapshots/events
- Live usage data merging for current-hour dashboard accuracy
- 2-decimal API cost formatting across usage/overview pages
- Comprehensive test coverage (PG, SQLite, HTTP, UI)
Merged by github-maintain automation.
CalculateCost previously ignored usage.ThinkingTokens entirely, causing
cost underestimation for every reasoning-capable model (OpenAI o3/o4-mini,
Codex/GPT-5, Anthropic extended thinking).
Thinking tokens are reported by providers as a sub-count of completion
tokens — OpenAI completion_tokens already includes reasoning_tokens,
Codex output_tokens likewise, and Anthropic output_tokens counts thinking
at the output rate. Default behaviour (ReasoningPerMillion=0) now matches
provider billing semantics: full CompletionTokens priced at OutputPerMillion,
no double-counting.
When ReasoningPerMillion is explicitly configured, the completion bucket is
split into visible output + thinking and each priced independently. Defensive
clamp to zero when Anthropic's thinkingChars/4 estimate exceeds reported
OutputTokens.
Adds ReasoningPerMillion field to ModelPricing and 7 regression tests
covering the double-count gate, distinct-rate split, Anthropic estimate
overrun, and backward compatibility for non-reasoning models.
Duc Nguyen