Track C — Postgres + Apache AGE Full Regression Results (2026-05-22)

Post-campaign update (2026-05-23): #1156 (per-namespace K8 quota dimension extension) bumped both adapters to schema v50 after this campaign closed. The v49 references below reflect the state at the 2026-05-22 gate; the v50 bump is additive (agent_quotas PK (agent_id) → (agent_id, namespace)) and does not invalidate any of the C.1-C.11 results. The current v15 → v50 ladder is exercised by tests/per_namespace_quota.rs + the existing tests/postgres_schema_parity.rs::current_version_matches_canonical assertion (now pinned to 50). See CHANGELOG.md [Unreleased] for the full v50 narrative.

Track C is the postgres + Apache AGE backend regression at the v0.7.0 release-gate tip. Two surfaces are exercised under one matrix:

• 76 store::postgres::tests::live_* rows in src/store/postgres.rs covering the SAL trait methods on the postgres backend.
• 36 postgres-gated integration tests distributed across tests/store_parity_gaps.rs, tests/serve_postgres_*.rs, tests/v49_archive_roundtrip_1025_postgres.rs, tests/postgres_schema_parity.rs, tests/admin_*.rs, tests/kg_timeline_postgres.rs, and the AGE-gated tests/kg_*.rs rows.

The lan-parity container ai-memory-lan-parity-pg-age on 127.0.0.1:15432 hosts PG16 + AGE 1.6.0 + pgvector 0.8.2. Both AI_MEMORY_TEST_POSTGRES_URL and AI_MEMORY_TEST_AGE_URL point at the same DB. The --include-ignored --test-threads=1 invocation flags ensure (a) every previously-#[ignore]-gated live test runs, (b) tests serialize on the shared DB without cross-test interference.

Phase summary

Verdict at a glance: SHIP — postgres + AGE cross-store parity is green at tip fd172f2cf against the lan-parity container.

Phase C.1 — Postgres SAL trait coverage (76 live_* rows)

The src/store/postgres.rs::tests module enumerates every SAL trait method against the live PG backend. Sample (truncated; full set in the run log):

test store::postgres::tests::live_delete_removes_row ... ok
test store::postgres::tests::live_detect_kg_backend_returns_cte_on_missing_extension ... ok
test store::postgres::tests::live_governance_allow_owner_at_leaf ... ok
test store::postgres::tests::live_governance_deny_non_owner_inherited ... ok
test store::postgres::tests::live_governance_inheritance_cap_at_five ... ok
test store::postgres::tests::live_governance_pending_on_approve_level ... ok
test store::postgres::tests::live_kg_backend_resolves_to_age_when_extension_present ... ok
test store::postgres::tests::live_kg_backend_resolves_to_cte_without_age ... ok
test store::postgres::tests::live_kg_invalidate_dispatches_to_cypher_under_age ... ok
[...]

29 of these 76 are visible in the abbreviated log grep (store::postgres::tests::live_); the rest are present at the unabbreviated level and accounted for in the 7,321 total.

The pgvector dimension warning (“memories.embedding column dimension (768) does not match the requested embedder dim (384); run ai-memory schema-init --store-url <url> --embedding-dim 384”) is expected — the test DB was provisioned with the default 768-dim column; the test runner brings up a 384-dim embedder for speed and the warning surfaces. This is an intended diagnostic, not a test failure; the schema-init recipe is documented in src/store/postgres.rs:391 and the test path tolerates the dim mismatch by skipping the vector-similarity branch and asserting on the FTS branch alone.

Phase C.2 — v15 → v49 migration ladder

CURRENT_SCHEMA_VERSION = 49 lives in src/storage/migrations.rs (sqlite) and src/store/postgres.rs:391 (postgres). The postgres ladder runs migrate_v15() through migrate_v49() idempotently on the first daemon open against a fresh DB; the lan-parity container’s on-disk state is at v49 at test time.

Notable per-version contributions:

#1132 pinned POSTGRES_CURRENT_VERSION from 48 to 49 in the postgres_schema_parity test (commit 2b8e704b3). The schema itself had been at v49 since #1025 landed; the test fixture had gone stale.

#1140 added bypass_visibility to the postgres half of the v49 archive→restore roundtrip (commit 07f22e6d3). The production code path is unchanged; the test fixture needed the explicit bypass so the cross-agent archive→restore parity could be verified without tripping the post-#1075 SAL visibility gate.

Phase C.3 — AGE backend dispatch (CTE-vs-Cypher)

The detect_kg_backend() helper resolves whether the connected DB has the age extension installed; if yes, KG traversal dispatches to Apache AGE Cypher (MATCH (m)-[r]-(n) WHERE …); if no, it falls back to recursive-CTE in plain Postgres.

The CTE fallback path is fully implemented and tested — operators running plain Postgres without AGE still get KG traversal, just slower. This matters for the v0.7.0 “managed-postgres” deployment posture where AGE is not always available (RDS, Cloud SQL, etc. ship AGE-less by default).

Phase C.4 — pgvector schema pin (#1120 substrate fix)

Before commit 1cdc67da6, init-age.sql ran CREATE EXTENSION IF NOT EXISTS vector; after CREATE EXTENSION age; had already pushed ag_catalog to the top of the search_path. pgvector consequently landed in ag_catalog, where the post-init SET search_path = public, ... for the daemon could not resolve the vector type. The downstream symptom: 30 live_* tests that touched embeddings (live_recall_*, live_store_with_embedding, live_insert_if_newer_with_embedding, etc.) failed with “type vector does not exist”.

The fix pins the extension creation to the public schema:

-- post-#1120 (commit 1cdc67da6)
CREATE EXTENSION IF NOT EXISTS vector SCHEMA public;

This is a real substrate defect that existed since the AGE extension landed; it surfaced only when the lan-parity container was rebuilt from scratch and the init-age.sql ordering became load-bearing. The retest evidence: all 30 previously-red rows are GREEN at tip fd172f2cf against a freshly-initialized lan-parity DB.

Phase C.5 — Postgres admin-gate allowlist contract (#1133/#1135)

The admin-gate sweep (#946/#957/#1027) tightened the for_admin privacy-bypass requirement to a populated allowlist; the production code refuses admin operations with an empty allowlist and returns HTTP 403. Two test fixtures needed updating to match:

Both fixtures now match the production contract: admin operations require an explicit allowlist, and the test demonstrates that.

Phase C.6 — Schema version pin (#1132)

tests/postgres_schema_parity.rs::current_version_matches_canonical pins the test-side schema-version constant to the canonical POSTGRES_CURRENT_VERSION in src/store/postgres.rs. Bumped from 48 to 49 in commit 2b8e704b3 after #1025’s v48 → v49 archive column expansion landed.

Phase C.7 — column_exists / index_exists schema-qualify (#1131)

tests/postgres_schema_parity.rs previously called column_exists("memories", "version") against information_schema.columns without a table_schema filter, which caused ag_catalog matches to interfere (AGE injects its own memories-shaped placeholders). Commit 202d09cf1 added the table_schema = 'public' filter to both column_exists and index_exists; the test now asserts on the canonical public schema only.

This is a test-instrumentation hardening, not a production code change. Production daemons always query against the public schema directly via the postgres backend’s connection setup; the test was the only path that read information_schema without qualification.

Phase C.8 — bypass_visibility for v49 archive→restore (#1140)

tests/v49_archive_roundtrip_1025_postgres.rs exercises the archive → restore lineage across agents. The post-#1075 SAL visibility gate would normally hide the cross-agent rows; the test explicitly opts into bypass_visibility=true for the parity check. Commit 07f22e6d3 added the bypass; the sqlite half of the same parity test already had it.

Phase C.9 — bypass_visibility for store_parity_gaps (#1138)

tests/store_parity_gaps.rs::trait_update_*_1024 and list_filters_by_agent_id_1030 are the canonical SAL-trait parity tests. They need to assert on data across agents to prove that the SQLite and Postgres backends agree on the contract; without bypass_visibility=true, the post-#1075 gate would hide half the assertion target. Commit 1620aaa45 added the bypass.

Phase C.10 — kg_timeline postgres owner-gate (#1134 substrate)

The SQLite path of kg_timeline had enforced an owner-gate (only the row owner can read the KG timeline for their memory) since the post-#944/#937/#938 governance sweep. The postgres path had been missed — the SAL trait method on PostgresStore returned the timeline unconditionally. Commit 3f911f630 added the owner-gate to the postgres path, restoring cross-store parity.

This is the second of two genuine substrate defects the campaign surfaced (#1120 being the first). The fix is a 3-line SQL WHERE clause addition in src/store/postgres.rs; the regression test is the same kg_timeline_postgres test that caught it. Without #1134, an attacker with PG-direct access through a non-owner agent could read KG timelines for memories they did not own.

Phase C.11 — kg_invalidate dispatches to Cypher under AGE

test store::postgres::tests::live_kg_invalidate_dispatches_to_cypher_under_age ... ok

When AGE is present, kg_invalidate writes a Cypher MATCH (m) WHERE m.id = $1 DETACH DELETE m rather than a plain SQL DELETE — this preserves the graph-edge cleanup that AGE’s DETACH semantics provide. When AGE is absent, the same SAL trait method falls back to a SQL DELETE in the recursive-CTE fallback. Both branches GREEN at tip.

Cross-store parity scorecard

Verdict: SHIP

Postgres + Apache AGE cross-store parity is GREEN at tip fd172f2cf against the lan-parity container. Two substrate defects were surfaced + fixed in-campaign (#1120 pgvector schema pin, #1134 kg_timeline postgres owner-gate); all other Track-C fixes were test-fixture pin-update discipline.

Strengths

• All 76 live_* rows GREEN, including the 30 previously-red embedding-dependent rows that #1120 unblocked.
• AGE backend dispatch resolves correctly in both directions (AGE-present → Cypher, AGE-absent → CTE) — the v0.7.0 managed-postgres deployment story is sound.
• Schema-version contract enforced by postgres_schema_parity (#1132) means future v50+ bumps will surface as test failures, not silent drift.

Audit trail

• Full-suite log: .local-runs/full-suite-final-v18-2026-05-22.log
• Substrate fix commits: 1cdc67da6 (#1120), 3f911f630 (#1134 + #1135)
• Postgres-specific retest log: .local-runs/postgres-lib-retest-with-age-2026-05-22.log (the isolated postgres + AGE re-run that pre-validated the full-suite v18 mint).

Recommendation

SHIP. The cross-store parity invariants are mechanically pinned by the SAL-trait parity test set; #1134 closed the last known substrate parity gap. The lan-parity container is the canonical test target for v0.7.0; the cross-node Track D campaign (192.168.50.100 ↔ 192.168.1.50) remains operator-action-gated on subnet routing per #836 Tier-2.

Drafted by Claude (Opus 4.7, 1M context).