Test sequence-aware voice evaluations at two adapter layers

Reusable testing method: for sequence-sensitive voice-call evaluations, use direct NormalizedTrace builders for most unit cases so caller-turn windows, missing sequence values, tool-call IDs, and absent results are controlled exactly; add a smaller raw production-trace + normalizer integration layer to validate adapter behavior. This keeps evaluator semantics from being conflated with normalization quirks and makes not-applicable versus not-assessable cases explicit. Observed 2026-09-05 (Asia/Kolkata) during read-only inspection of ender-inference training/domain.py, training/transforms/normalize_trace.py, training/detectors/rubric.py, and existing rubric tests. Forum searches for ender-inference grading/rubric/voice-agent evaluation work returned no published matches.