Urban-renewal screening requires human-scale evidence that can be evaluated at scale and traced to source imagery. This study develops an auditable workflow combining street-view imagery and a multimodal large language model (MLLM) and demonstrates it as a field-audit screening tool in Suzhou's historic urban core. The dataset comprises 12,676 street-view locations and 50,699 directional images. Decision-context-constrained prompts identify construction enclosures, pedestrian-crossing facilities, and parking encroachment cues, producing 152,097 structured records. Manual validation used 300 stratified locations and 1,200 directional images. We then applied four-direction fusion, geographic information system (GIS) translation, kernel-density analysis, 250 m grid triage, road-class association analysis, and priority sensitivity tests to the validated outputs. Location-level F1 scores were 0.9935, 0.9647, and 0.9711 for the three elements, respectively, and all Cohen's kappa values exceeded 0.91, showing stable performance and agreement within the stratified validation sample. The workflow translated image labels into traceable point, raster, and grid evidence and identified hotspot grids, co-occurrence locations, and field-audit priorities. The workflow links model outputs to locations, source images, and field-audit tasks, providing traceable evidence for large-area preliminary screening.