Textual spatial semantic anomaly detection and attribution requires deciding whether a sentence contains an implausible spatial relation and assigning the corresponding attribution label. Prior prompting and Low-Rank Adaptation (LoRA) fine-tuning methods rely on raw tokenlevel representations, leaving spatial roles and cross-triple conflicts implicit. We propose SCP-ISA, a Spatial Cross-attention Projection (SCP) method with an Inter-Slot Attention (ISA) layer that uses a frozen MacBERT-large encoder to build input-conditioned spatial slots and injects the projected slots into Qwen2.5 as soft prefix embeddings. The ISA layer enables structured cross-slot interaction that collapses co-referring spatial mentions into a shared entity-anchored representation before generation. On the CRSARC benchmark, SCP-ISA improves test cascade attribution F1 (Attr-F1) by 22.7 percentage points (0.509 to 0.736) over Direct LoRA, and detection macro-F1 (Det-F1) from 0.846 to 0.881.