Germany's Heat Planning Act (Wärmeplanungsgesetz, WPG) requires municipalities to divide their territory into prospective heat-supply areas including district-heating areas (Wärmenetzgebiete), areas to be examined (Prüfgebiete), and areas for decentralized supply by 2026/2028. In doing so, the responsible planning authorities must combine spatially rich but analytically isolated geodata (building footprints, ALKIS, the 2022 Census, LoD2 heights) with extensive and frequently amended legal and methodological texts. Existing energy atlases deliver static maps without legal traceability, while document-based AI tools remain location-agnostic. Neither answers the question that is decisive for any review: why is this hectare classified this way, and on what authority? We present a working geospatial decision-support prototype that treats auditability as a first-class design goal. Per-building heat demand is aggregated to a hectare grid and classified into three-level supply zones. Every classification is linked to both its underlying data and the governing legal clause. A spatially-grounded Retrieval-Augmented Generation engine (GeoRAG) answers free-form questions over the real corpus (the WPG, the federal Leitfaden Wärmeplanung, and a state guideline (KEA-BW)) returning deep-linked citations that explicitly distinguish binding law from advisory convention (e.g., the 600/300 MWh/(ha·a) density threshold is flagged as methodological, not statutory) and expose the underlying legal version (Rechtsstand). In the pilot city of Lippstadt (NRW; ≈1293 GWh/a total heat demand), 11806 hectare cells are classified into 537 district-heating candidates, 1112 Prüfgebiete, and 10157 decentralized cells and the 537 network-candidate cells concentrating 45.2% of total demand. The system fuses map and legal reasoning, surfaces measured-vs-modeled data-confidence, supports live what-if heat-source scenarios, and quantifies the classification's sensitivity to convention-based thresholds. Because the answers come from a swappable set of source documents, the same tool can be adapted to another region simply by exchanging federal guidance for state guidance. The contribution is a reusable design for heat-planning tools that stay spatially precise, always cite their legal source, and clearly mark where the data is uncertain, for use by municipalities, utilities, and planning offices. Keywords: GeoAI, Retrieval-Augmented Generation, municipal heat planning, WPG, spatial decision support, auditability, LLM grounding, Wärmewende