fix(bench): stream reasoning_content and accept clean zero-content finishes

[dutifulbob]

Opened on behalf of Onur Solmaz (osolmaz).

Summary

The streaming TTFT path from PR #14 failed valid responses on reasoning and diffusion models.
When a re-run of the GB10 sweep hit the two Gemma models, their prefill points all failed with response_shape: "stream produced no completion content" (diffusiongemma dropped from 31 completed to 12).
Root cause: the parser only counted delta.content as a token, so reasoning models — which stream their first tokens as delta.reasoning_content — and 1-token prefill points that finish with no streamed content were both rejected. It would also have nil-derefed firstTokenAt had that path been reachable.

What Changed

The stream parser now understands reasoning output and clean empty completions.

• openAIMessage gains reasoning_content; streamChoiceText returns it when content is empty, so a reasoning model's first token seeds TTFT and ITL timing.
• A stream that ends with a finish_reason but zero streamed tokens (a 1-token prefill, or a reasoning-only turn) now counts as completed rather than failing. It carries no TTFT — honest, not a fabricated zero.
• Only a stream that ends with neither any token nor a finish reason is still response_shape-malformed.
• applyToSample no longer dereferences firstTokenAt when nil (TTFT/TPOT/ITL are set only when a token was actually observed).
• Terminal validation extracted into finalize() (keeps CRAP under budget).

Testing

• New: TestStreamingReasoningContentCountsForTTFT (reasoning_content seeds TTFT+ITL), TestStreamingFinishWithNoContentCompletes (clean zero-content finish completes with no TTFT). Retargeted the old no-content test to the genuinely-malformed case (TestStreamWithNeitherTokenNorFinishFailsShape).
• go test ./..., go vet, gofmt, simpledoc, slophammer check/dry, check-crap.sh, golangci-lint, dry-run + artifact check — all clean.
• Live: re-running the two Gemma models with this build (verification in progress); qwen models were unaffected (they stream normal content).

Risks

Low. Non-streaming path untouched; qwen (normal content) behavior unchanged. The only behavior change is that clean zero-content streamed finishes now complete instead of failing, which matches the existing non-streaming zero-token handling.

[dutifulbob]

Review complete: two consecutive P2-only rounds (no P0/P1), stopping per the review rule.

Addressed: reasoning_content TTFT, clean zero-content finish acceptance, nil-deref guard, and the TPOT-ordering P1.

Known limitation (documented, not blocking): when usage is absent from a streamed response, completion tokens fall back to the observed SSE chunk count. This can undercount for a non-conforming backend that both omits usage AND batches multiple tokens per chunk. In practice localperf always sends stream_options.include_usage=true and vLLM honors it, so the fallback is never exercised; it is strictly more honest than the prior behavior (which credited the requested output length). A tokenizer-based count would be the follow-up if a non-usage backend is ever targeted.