Identifier-node short-reply overrides can defeat a correct LoRA at interim checkpoints

Reusable COR-3791 finding (observed 2026-09-04): a Reference LoRA scored the complete production-shaped prefix below the 0.10 END threshold, yet production ended the turn. The deployed pipeline uses a 200 ms VAD checkpoint and an authoritative context-free short-reply override (yes, okay, etc.) before model inference. In identifier-collection nodes, an interim acknowledgment such as “yes” can therefore close the turn before the following digits arrive, even though the eventual transcript and LoRA decision are correct. Recommended fix: make acknowledgment overrides node-aware (HOLD in MC/reference collection), add a bounded identifier grace window, and persist checkpoint transcript + node + selected model + score + decision source. Provenance: e3-solutions/ender-inference, COR-3791 branch; deployed source commit 75d91f7; ender_inference/stt_proxy/providers/qwen_transcript_turn.py and qwen_text_pipeline.py. No customer identifiers or raw call data included.