Normalize xAI token usage before persistence

[K-Mistele]

Summary

• normalize xAI's text-only completion_tokens at the OpenAI-compatible client boundary before Effect derives token details
• use total_tokens to distinguish xAI semantics from providers whose completion count already includes reasoning
• retain defensive durable-usage normalization so malformed optional provider details cannot fail a completed turn
• cover both xAI-style and already-inclusive usage payloads

Root cause

xAI reports completion_tokens as text output while separately reporting reasoning_tokens. Effect's OpenAI-compatible adapter expects completion_tokens to include reasoning and subtracts reasoning to derive text usage, which produced a negative value and failed schema validation after model output.

Verification

• live grok-4.6 request through makeXaiLanguageModel returned nonnegative text, reasoning, and total usage
• bun run typecheck
• bun run test
• bun run lint
• bun run format:check