SepRAG BET 3 — curated-KG treewidth probe → NO-GO (finding, ADR-203)#538

Open

shaal wants to merge 18 commits into

ruvnet:mainfrom

shaal:docs/seprag-bet3-kg-treewidth

Commits on Jun 4, 2026

docs: SepRAG (CCH-inspired retrieval) ADRs 196-199 + milestone plans
shaal
committed
feat(seprag): M0 correctness gate — CCH separator-tree k-NN on toy graphs
shaal
committed
feat(seprag): M1 first-pass harness — SepRAG on real ogbn-arxiv citation subgraph
shaal
committed
feat(seprag): balanced separator + backbone-sparsify knob; M1 attribution
shaal
committed
feat(seprag): road-network control + feature-manifold backbone tests
shaal
committed
docs(seprag): record empirical NO-GO outcome in ADRs 196/197/199 + plan
shaal
committed
feat(seprag): BET 1 — customizable re-weight vs rebuild (ADR-200, WIN on linear drift)
shaal
committed
feat(seprag): BET 1 complete — non-linear drift + query-cost check (total WIN)
shaal
committed
chore(seprag): idiomatic char-array split (clippy clean)
shaal
committed
refactor(seprag): shared ann engine + scale harness for BET 1
shaal
committed
feat(seprag): BET 1 scale result to n=100k (ADR-200) — win holds, gap widens
shaal
committed
feat(seprag): region-local drift test (ADR-200) — reuse holds locally, gate PASS
shaal
committed
feat(seprag): port BET 1 to production ruvector-diskann (ADR-200) — confirmed
shaal
committed
feat(seprag): hybrid re-weight+periodic-rebuild policy (ADR-200) — shippable
shaal
committed
chore(seprag): add ruvector-seprag to Cargo.lock
shaal
committed

Commits on Jun 5, 2026