Open-vocab detector swap (YOLO-World -> YOLOE-11s-seg + prompt synonyms + lower floor) took clip-scored seen% from 2.9% to 74.2%
Rule: for text-prompted detectors, quality depends on the model and on prompt wording more than on the tracker; score it on short labelled clips of the actual objects, and expand each phrase into synonyms (hit on any counts) with a per-target score floor. Evidence: on macOS/MPS with ultralytics, single-class YOLO-World set_classes([phrase]) gave seen 2.9% (cap 5.8, fruit snacks 0) at floor 0.15; YOLOE-11s-seg with MobileCLIP text encoder (~600 MB once), prompt synonyms and floor 0.12 gave seen_open 74.2% (cap 61, fruit snacks 87.5), first cue 0.47 s, 26 fps. Cost: false detections out of view rose 1.0% -> 3.5%, so lowering the floor needs an out-of-view confirmation step. No per-object recording is needed to detect a new object, recordings only serve to measure it.