[diegosouzapw]

Summary

#11883 (merged) fixed the OpenAI→Claude stream translator dropping usage carried on a
trailing {"choices":[],"usage":{...}} chunk — but only fixed it for state.usage
internally. When the finish_reason chunk arrives before the trailing usage chunk (the
normal order for Fireworks/vLLM/Together and other stream_options.include_usage
upstreams), the terminal message_delta/message_stop events were still emitted
immediately on finish_reason, carrying whatever state.usage held at that moment
(stale/zero) — the trailing chunk's correction arrived too late to matter, since nothing
re-emits after the finish event already went out to the client. Verified directly against
the merged #11883 code: the actual emitted message_delta still carried
{"input_tokens":0,"output_tokens":0} in this exact sequence.

This ports #11915's fix (HouMinXi): the finish emission now defers via
state.pendingClaudeFinishChoice until either the real usage has arrived or a genuine
end-of-stream flush (chunk === null, already an established mechanism in
open-sse/translator/index.ts for formats like openai-responses) forces it through with
whatever was available.

This is a hand-merged combination of #11883's already-merged trackUsageFromChunk()
extraction with #11915's deferral mechanism (the two conflicted on the same lines), not a
straight cherry-pick — full credit for the deferral design and its regression coverage
goes to #11915.

Verification (on release/v3.8.51 tip)

• Direct reproduction before the fix: finish_reason chunk emits message_delta with
  {"input_tokens":0,"output_tokens":0}; after the fix: null (deferred), then the
  trailing chunk emits the correct message_delta with real usage.
• 86/86 tests passing across tests/unit/openai-to-claude-trailing-usage-11817.test.ts
  (fix(translator): keep upstream usage from trailing empty-choices chunks #11883's suite), tests/unit/translator/openai-to-claude-trailing-usage.test.ts
  (fix(translator): preserve usage metrics from trailing choices-empty chunks (#11817) #11915's suite, ported), and 14 other openai-to-claude translator regression files
  (including the fix(translator): prevent doubled tool args in OpenAI-to-Claude #5828 and fix(backend): internal reasoning-replay placeholder leaks into assistant message content #8081/fix: strip internal reasoning placeholder from user-visible content (#8081) #8162 regression guards, and the existing
  "two finish_reason chunks emit finish events exactly once" duplicate-finish guard).
• npm run typecheck:core — clean
• node scripts/check/check-file-size.mjs — pass

Related to #11817, ports #11915 (HouMinXi).