stats() and list(kind=edge) disagree on namespace scope for identity-bearing callers

[ohdearquant]

Observed

For a caller whose request frame carries a visible-namespace set beyond local:

• list(kind="edge") with the keyset cursor walks edges across all frame-visible namespaces (observed: 11,174 unique edges = 11,173 in local + 1 in the caller's own namespace).
• stats() reports edges: 11173 — one namespace only.
• stats().edges_by_relation sums to 11,173, so it shares the scalar's single-namespace scope (count_edges_by_relation in crates/khive-db/src/stores/graph.rs pins WHERE namespace = ?1).

So a full list walk and the stats totals cannot be reconciled by a caller with any non-trivial visibility: the two verbs compute over different namespace scopes.

Analysis

The store-layer functions are all namespace-pinned; the divergence is at the runtime layer: edge listing merges pages across the caller's visible namespaces, while the stats path queries a single namespace regardless of frame visibility.

Per the per-request identity model, read surfaces should honor the frame's visibility consistently. Either:

1. stats() aggregates over the caller's visible namespaces (consistent with list), or
2. both accept an explicit namespace= parameter and default identically.

Option 1 matches the direction the listing surface already took. Whichever is chosen, the scalar totals and edges_by_relation must share one scope, and the contract should be stated in the verb docs.

Acceptance

• stats() totals == count of a full list keyset walk under the same caller identity, for entities, edges, and notes.
• edges_by_relation sums to the edges scalar under all identities.
• Regression test with two namespaces and an identity-bearing frame.