sembench 0.2.3

0.2.3

• The system turn a manifest row asks for is now sent. Rows that carry
  metadata.system_prompt are replayed as [system, user], the message list
  their expectations were measured against. Earlier releases sent the user
  turn alone.
• --staged-prefill replays a row's metadata.staged_prefill plan: each
  cut is sent as a token-exact /v1/completions prefix (max_tokens=1,
  vllm_xargs.semblend_stage=1), then the full prompt as token ids. TTFT
  counts from the first stage.
• --capture-hint-role marks rows of one manifest role with
  vllm_xargs.semblend_capture=1. The harness knows the roles, so this
  measures the ceiling of hinted capture, not what a router could do.
• Single-population alignment rate published beside the existing metrics
  (alignment_given_opportunity_in_class).
• Test fix: the throughput check allows for wall_seconds rounding.

SemBench 0.2.0

• run-load: donor->recipient streams with a fixed number in flight; requests per second, output tokens per second, TTFT percentiles for donors and recipients.
• run-live-gateway --post-donor-delay-ms for engines that index donors asynchronously.
• ROI cost model recalibrated to 8K to 24K measurements on an A10G.
• Long-context manifest generator and the exact arm recipe in examples/overhead-benefit/.