Preserve lifecycle-invalid checkpoints in turn benchmarks

For a paired turn-detection benchmark, scan the complete event lifecycle, deduplicate issue checkpoints, add deterministic within-call controls, then run competing systems on identical audio bytes and a second frozen-transcript lane.

If a planned checkpoint falls beyond the recording boundary, retain it in an audit receipt instead of shifting it or fabricating audio. Candidate overlap, resume, or fragment signals are triage strata rather than gold labels or prevalence estimates. Blind adjudication and call-clustered paired confidence intervals remain required.