[euforic]

The Orion deployment validation fails for xAI chat responses that report visible completion tokens separately from reasoning tokens: the fork retains normalizeXAIChatUsage, but unary and streaming chat entrypoints only normalize provider cost. Both existing deployment assertions see 5 completion tokens instead of 56 (5 visible + 51 reasoning), so image publication never starts.

Route both entrypoints through the existing idempotent usage normalizer, retaining provider-reported cost. Add a pure streaming regression and two live-harness cases for unary and terminal SSE token totals.

Validation: the new pure regression failed before the fix with 5 versus 56 completion tokens and passes afterward. Existing pure normalization/cost regressions pass. Gateway/provider builds pass. Harness augmentation and xAI filtering retain exactly two cases. The existing Orion workflow will validate the full provider-account implementation before building/promoting an immutable image. Paid live provider inference remains an operator check.

make dev PORT=8081 APP_DIR=$(pwd)/tests/integrations/python
make run-provider-harness-test PROVIDER=xai FEATURE="reasoning usage normalization" BASE_URL=http://localhost:8081

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}