Audit malformed Ego4D voice segments

docs/ego4d_av_llm_annotation_review.md

@@ -55,6 +55,18 @@ handoffs are added as independently reviewable hard-negative proposals.
 Timing rules only select and describe candidates; they never assign semantic
 positive labels.
 
+Official zero-duration voice records cannot define an activity interval. They
+are never expanded to an invented duration: the generator skips them, reports
+the count on stdout, and stores their complete source records under
+`clip_audits` and each candidate's `temporal_evidence`. A negative-duration
+record remains a hard error.
+
+Ego4D v1 also contains nine train clips whose official voice endpoint extends
+past `clip_end_sec` by 0.42--1.03 seconds. The default audited ceiling is 1.1
+seconds. These endpoints are preserved and flagged rather than silently
+clamped; larger overruns remain errors. A human should normally exclude a
+semantic boundary beyond observable media as `DATA_RECORDING_BOUNDARY`.
+
 ## 2. Inspect what will be uploaded
 
 The LLM client refuses to upload transcript excerpts without an explicit