fix(openai-api): report turn token usage

[fancyboi999]

Summary

• Supersedes fix(openai-api): report real token usage in chat completions #8 with a clean, one-commit branch from current main; the old PR remains open.
• Carries each AgentRunner turn's accumulated usage through AgentLoop outbound metadata to non-streaming /v1/chat/completions responses.
• Normalizes the three OpenAI usage counters and accumulates them if an empty first completion is retried.

Root cause

The API response constructed its usage object as three hard-coded zeroes even though the agent runtime had already accumulated provider usage for the completed turn.

Compatibility and scope

• Missing usage still produces the existing all-zero usage object.
• Streaming behavior is intentionally unchanged; this PR fixes the three hard-coded fields in normal chat-completion responses only.
• The branch changes four files and does not include the unrelated upstream/rebase history present in fix(openai-api): report real token usage in chat completions #8.

Validation

• Targeted OpenAI API and AgentLoop integration tests: 36 passed.
• Full agent suite: 4,628 passed, 3 skipped.
• Root lint, typecheck, build, and test suites passed.
• Built API HTTP check: /health returned OK and /v1/chat/completions returned usage 42 / 17 / 59 from a local provider stub.

Supersedes #8.