repo-indexlisted

# One graph, two tiers Governs: shared/scripts/index/ Retrieval fails in repositories for a specific reason: the useful unit is not a paragraph, it is a symbol and its neighbourhood. A chunk-and-embed pipeline returns files that talk *about* authentication; what the task needs is the four functions that would break. So: **one graph, built from source, queried two ways.** Not a stack of retrievers — an extra retriever whose results have to be merged is a second ranking problem on top of the one you had. ## The graph Nodes are `file:<path>`, `sym:<path>:<name>`, and `doc:<path>`. Edges: | Edge | From | Weight | |---|---|---| | defines | file → symbol | tree-sitter tags | | references | file → symbol | tree-sitter tags, split across definers | | imports | file → file | resolved import statements | | calls | symbol → symbol | call sites inside a definition's span | | governs | doc → path | a `Governs: src/billing/` line in the doc's head | | supersedes | doc → doc | `Supersedes: 0004` | **A symbol node is keyed by the file that defines it.** Keying by bare name merges every definition sharing it, and the merged node then dominates the graph: in this plugin's own repository `sym:main` — sixteen unrelated `def main()` — had a higher degree than any real file, and seeding on one guard ranked an *empty template* second because five guards define `check`. Bare names are still how a reference resolves and how `--seed` matches; they are just no longer the node. A reference to an