PC-PLF: Path-Conditioned Per-Layer LoRA Fusion for Open-Vocabulary ROADWork Segmentation

Construction work zones are a difficult case for open-vocabulary semantic segmentation. Their layouts are temporary, safety-relevant objects that are often small and long-tailed, and generic models readily confuse them with background. We address these failures inside an LoRA adapter space rather than retraining the backbone. Using only ROADWork training data, we audit a CAT-Seg RoadWork LoRA for false-positive- and recall-dominated cases and pair them with anchor images to train a residual adapter. Path-conditioned per-layer LoRA fusion (PC-PLF) then distributes a global correction budget across adapted layers using each layer’s first-order tangent magnitude along the stored factor path. Under group-disjoint out-of-fold evaluation on ROADWork, the complete method raises the mIoU from 61.72 for the RoadWork LoRA baseline to 62.29, with a shared-budget allocation gain of 0.33 mIoU over uniform fusion. Most of the total improvement appears before per-layer allocation. Uniform residual fusion contributes 0.69 points over the baseline, confirming that failure-driven residual training supplies the larger share; PC-PLF contributes a smaller allocation effect when tested on the same trained base-residual pair. The allocation effect is reproducible across four curation rules but near zero under SAN architecture transfer and MUSES second-target-domain evaluation. Three-group and text-weighted controls do not recover the full gain. Improvements concentrate in several long-tail safety classes. Cross-architecture, cross-dataset, and calibration audits define the operating regime rather than universal advantage. Residual curation supplies the larger share of the improvement; layer-wise allocation contributes a smaller, pair-specific gain.