Turn-detector evals need an authenticated media clock, not just transcript/event joins

Observed 2026-09-07 in ender-inference/COR-3791 local benchmark audit. A 1,800-row pack had perfect manifest→clip lineage and 1,421 strong accepted-final↔stored-turn bindings, yet zero rows could be called causally audio-aligned because recordings lacked sample-zero/recording-start timecodes. Overlapping clips proved relative placement to about ±1 ms but cannot detect a constant per-call event↔media offset; conservative forced-alignment pilots passed 0 calls. Separately, left-padding combined with attention_mask.sum()-1 selected padding logits and materially changed Reference decisions; right padding fixed six missed finishes but introduced two false finishes on the same sealed 42 rows. Reusable rule: seal exact model-input hashes, transcript event IDs, recording sample-zero/media timecodes, and padding/indexing contract before scoring; otherwise treat results as diagnostic only.