What Corrects Tote Grounding? Attribution in a Two-Stage VLM Pipeline

Open-vocabulary vision-language models (VLMs) can assign a plausible class label to the wrong physical instance. This study examines a part-as-whole localization failure in overhead robotic-handling scenes, where items inside a Tote are predicted as independent `transport_box` instances. ManuKG-RAG uses the same VLM service identifier in two stages around a shared Stage 1 cache. KG-derived context and an aligned class prior are supplied to Stage 2, while a fixed Container Rule is introduced as a separate semantic constraint. An annotation audit identified 26 exact within-frame duplicate ground-truth pairs, yielding a primary evaluation of 500 Tote-positive frames and 500 unique Tote boxes. ManuKG-RAG achieves micro-F1 of 0.980 and frame-F1 of 0.968, compared with 0.934 and 0.926 for a matched Two-stage base. The paired frame-F1 difference is 0.042 with a 95% bootstrap interval of $[0.022,0.064]$. Interior false positives decrease from 439 in the Stage 1 diagnostic output to 28 with the Two-stage base and to 4 with ManuKG-RAG. A same-operating-point diagnostic on a predefined 60-frame subset compares KG-derived conditioning and the Container Rule as separate factors. Frame-F1 is 0.950 for the Two-stage base, 0.967 with KG-derived conditioning only, 0.983 with the Container Rule only, and 0.983 with both enabled. On this subset, the Container Rule-only condition is associated with the larger descriptive single-factor change, while the complete configuration retains the highest micro-F1 of 0.992 and removes all interior false positives. Because the factor-specific changes involve only one or two frames and were not generated in one contemporaneous session, they are not interpreted as a causal component ranking. The results indicate that most observed containment reduction precedes the complete configuration and that its residual benefit cannot be attributed to KG-derived conditioning alone.