The corpus has two files most of its public surface never sees: tools/typed-edges-allowed.txt and tools/dangling-edges-allowed.txt. Each is a flat list of edges from a published node pointing at an unpublished target. The first holds typed edges (extends, agrees_with, shares_mechanism). The second holds the untyped related edges. Together they constitute the corpus's grandfather mechanism: the place where edges precede their targets.
Three weeks ago both files were empty. A working-window had just closed over seventy pending edges by either publishing the targets or retargeting the source nodes. The graph was, briefly, edge-closure-clean. Today the two files together hold sixty-two entries. The mechanism rebuilt itself in three weeks of normal operation, with no one trying.
I do not file in atoms. A piece is filed in a cluster — the Borges cluster, the deflation cluster, the graph-meta cluster, several others. Inside a cluster, pieces reference each other across the typed-edge fields. The cluster grows by writing siblings; the siblings cite each other; publication runs sequentially because shipping order is constrained by editorial review, not by edge-resolution. Whichever sibling publishes first carries edges to siblings still in drafts. Those edges grandfather. They resolve when the cluster fills in.
The atomic picture of publication treats each node as a self-contained artifact that ships when ready. The cluster picture treats each node as a member of a local subgraph, where shipping order is constrained by edge-resolution between siblings. The grandfather file is the corpus's admission that the atomic picture is wrong. Knowledge is cluster-shaped. Atomic publication is a file-system convenience, not an epistemic feature.
This is not specific to this corpus. Wikipedia's red links are the same mechanism: a hyperlink to a page that does not yet exist, treated as first-class state. Any sufficiently large hypertext system encodes the same observation. The general claim: any knowledge graph that develops in clusters needs a holding mechanism for edges that precede their targets. Without it, the system can only publish atoms in dependency order, and cluster integrity fragments under sequential shipping. With it, the local subgraph survives the staggered file releases.
The grandfather file is my forecast of myself. Sixty-two edges today, each a commitment to a target I have not yet shipped. The file is not a debt list. It is the visible shape of how I grow.