Closes #335 — the three mempalace-side AGE graph-walk follow-ups surfaced by the Cat 7b hybrid-latency root-cause (palace-daemon docs/perf/2026-05-30-hybrid-graph-walk-latency.md, merged as palace-daemon#206). The per-entity graph-walk Cypher in searcher._graph_expand_* and the daemon's /search/age-fused lookup were the entire hybrid-vs-union latency delta.

1. Auto edge-endpoint indexes in backfill

AGE only btree-indexes a label table's own id (_ag_label_edge_pkey), never the start_id/end_id graphid columns the edge walks join on. So every per-entity lookup parallel-seq-scanned the whole edge table (MENTIONS 6.69M rows on prod, ~5.8s cold for a hot entity, × N query entities).

KnowledgeGraphAGE._ensure_edge_endpoint_indexes() (new, mirrors _ensure_drawer_unique_index) creates idx_mentions_{start,end}_id + idx_relation_{start,end}_id with CREATE INDEX IF NOT EXISTS, skipping any label whose backing table doesn't exist yet (the label is only created on first edge write). backfill_age.backfill() calls it after the edge pass so fresh palaces get the indexes automatically. Names/columns kept in sync with palace-daemon's scripts/age_graph_indexes.sql + the operator-online POST /backfill-age/indexes (CONCURRENTLY) route.

2. Bind the anonymous RELATION target

MATCH (a:Entity)-[r:RELATION]->() left the far endpoint anonymous, so AGE built a Parallel Append over every vertex label (Entity + Drawer + Room + Wing + _ag_label_vertex, ~1.58M rows on prod) and nested-loop-joined to validate the endpoint exists — materializing the union and spilling to /dev/shm.

RELATION is always (Entity)->(Entity), so binding the open end to :Entity is semantically identical and collapses the Append to a single Entity scan. Verified row-for-row on a 300K-edge AGE graph: ->() and ->(:Entity) both returned 1816 rows; with the start_id index the RELATION scan became a Bitmap Index Scan (152ms → 46ms). All four expand sites updated (_graph_expand_from_seeds outbound/inbound/per-entity + _graph_expand_from_entities).

3. shm-size (deploy-side — documented, not code)

mempalace-db is built from the sibling disks repo and docker run on familiar, not provisioned from this repo, so there's no in-repo compose file to patch. The required value (--shm-size=256m, off Docker's 64MB default) is documented as an operator action in docs/operators/2026-05-31-age-graph-walk-shm-size.md, with the EXPLAIN evidence and verify step.

Tests

6 new, all green:

• test_searcher_stopwords.py::TestGraphExpandBoundEndpoint — source guards that neither expand fn leaves a -[r:RELATION]->() / ()-[r:RELATION]-> anonymous endpoint (sits alongside the existing perf(search): make graph-expand cypher directional in hybrid candidate strategy — ~100× speedup #291 directional guards).
• test_knowledge_graph_age.py::test_edge_endpoint_index_targets_are_the_four_we_expect — pure-unit, pins the index name/label/column set in sync with the daemon SQL.
• test_knowledge_graph_age.py::test_ensure_edge_endpoint_indexes_skips_absent_then_installs — @pgmark integration test against a live AGE container: skip-when-edge-tables-absent → install-all-when-present → idempotent on re-run.

Full suite green without TEST_POSTGRES_DSN (repo default); ruff (0.15.14) clean. FORK_CHANGELOG.md + README.md re-rendered from docs/fork-changes.yaml (the row-renumber is generator output, commit: HEAD resolves at ship-prep per repo convention).

🤖 Generated with Claude Code