fix(gateway): keep hard budgets conservative when stream usage is incomplete

[islahuddinn]

Summary

When an Anthropic (or other streaming) provider connection drops after partial usage telemetry arrives, the gateway was reconciling hard-budget reservations down to that partial measured cost. That releases budget headroom the provider may still charge against.

This change tracks usage completeness separately from observed usage:

• UsageTee.usageComplete is set only on terminal provider usage frames
• billableCostCents() keeps hard-budget reconciliation conservative when a provider-charged error ends with incomplete usage
• Observed token counts are still stored on the llm_requests row; only budget billing uses the preflight estimate when final usage is unknown

Fixes #248

Test plan

• pnpm --filter @facility/gateway test (33 passed)
• Unit tests for billableCostCents and usageComplete
• Integration test: incomplete client abort keeps full reservation; complete abort bills measured cost

Made with Cursor