wire-knowledge-graph

Run tools/check_graph_coverage.py for every registered profile; it compares each seed against its graph and exits 1 on a stated relation that reaches no edge. Then grep the model's builder script for each relation-bearing field, since the tool proves an edge exists but not which field produced it.

A seed field that no builder pass reads is silently inert: validation only checks that names resolve, not that anything consumes them, and counts never fall for an edge that was never created. The fix is to write a builder pass that walks that field, mirroring a sibling model's builder that already has it.

EcoSIM supports scripts/build_ecosim_rag.py --graph-only, while FATES requires scripts/build_rag_index.py --rebuild --graph-only --profile <name>; --graph-only alone silently loads and re-saves the existing graph. PFLOTRAN has no graph-only mode, so use the rebuild-rag workflow instead.

wire-knowledge-graph handles relations stated in the seed that no builder pass reads, which fail silently. rebuild-rag handles reindexing, the vector store, and edges whose endpoints do not resolve, which the FATES builder reports loudly as skipped edges.

Minting nodes for unresolved names creates phantom entries that answer queries about variables the model does not have. An unresolved parameter name is a seed defect for a human to fix; the builder should report it and skip, while outputs may legitimately be created since they can precede the parsed registry.

Save the graph JSON before rebuilding, then diff node by node and edge by edge: a correct rebuild adds edges with zero removals, zero node changes, and zero pre-existing edge attribute changes. A count line alone cannot detect replaced edge attributes, and git diff on metadata files catches side effects like rewritten paths.