Handle n > 1 consistently in output measurement and server token usage #565

Problem

Follow-up to #552, reviewed at c805b32655fe675520bc90a4fe182053816c1e39.

OpenAI chat and completions requests accept n > 1 from dataset rows. The benchmark measures the first choice, while server usage can sum tokens across all choices. PR #552 rejects these rows when server token counts are primary; the default client-primary mode warns and permits them.

For streams, omitting stream_options.include_usage does not guarantee that a server omits usage. Ordinary content frames are reduced to their first choice, losing choice multiplicity. A separate usage-only trailer has no choices, so the accumulator can accept summed counts as first-choice counts. Secondary server OSL/TPS/TPOT and paired comparisons then become misleading. The completions adapter shares this issue.

Reproduction

Pass these frames through OpenAIMsgspecAdapter.decode_sse_message and OpenAISSEAccumulator:

{"choices":[{"index":0,"delta":{"content":"first"}},{"index":1,"delta":{"content":"second"}}]}
{"choices":[],"usage":{"prompt_tokens":10,"completion_tokens":42}}

The result contains only first but publishes completion_tokens: 42. This corresponds to an accepted client-primary n: 2 request when an endpoint sends usage despite the request not asking for it.

Expected behavior

Define and consistently enforce the policy for n > 1. While measurement remains first-choice-only, summed usage must not be presented as valid first-choice server metrics. Carry request multiplicity into response accumulation or another reliable request-level guard. Do not rely solely on choice count in a single streaming frame: choices can arrive separately.

If full multi-choice benchmarking is supported instead, define its output population, per-choice accounting, and throughput semantics explicitly.

Validation

• Cover chat and completions, streaming and non-streaming.
• Cover multiple choices in one frame and choices in separate frames, followed by an empty-choices usage trailer.
• Verify excluded summed counts are absent from events and appear as unavailable/partial in secondary server metrics and paired comparisons.
• Preserve the server-primary setup rejection until multi-choice server measurement is defined.
• Preserve n = 1 behavior.

This is follow-up work; #552 still has the described limitation for permitted n > 1 streams.