v0.7.0 Ship Campaign 2026-05-28 — SME / Engineer / Architect Briefing

Verdict

SHIP-CLEARED. Six tracks exercised against integrated tip be3347d704dad03bcc210c9eb0a517946dbe555f on release/v0.7.0: build + 1hr dogfood loop, A2A in-host via lan-parity Docker stack, postgres + Apache AGE full regression (cargo test --features sal,sal-postgres --release --no-fail-fast → 8,028 / 9 / 27 / 312), docs + GitHub Pages drift remediation closing #1197 + #1198, scope-disciplined CoALA prior-art citation (PR #1380), and a parallel-dispatched heterogeneous AI-NHI assessment v3 (in-flight at write time). Three codegraph-driven QC audits on the Track-D fix batch (A literal / B structural / C fault-injection) returned ZERO-DEFECTS-CONFIRMED. Substrate untouched by all campaign findings; the one substrate-adjacent fix (PR #1382) lives entirely in tests/**.

Reproducibility contract

Branch:           release/v0.7.0
HEAD:             be3347d704dad03bcc210c9eb0a517946dbe555f
Binary:           target/release/ai-memory  (thin LTO + stripped)
Install symlink:  /opt/homebrew/bin/ai-memory -> (above)
Dogfood daemon:   PID 10338, started 2026-05-27 16:55:25Z,
                  RSS ~18 MB sustained over 16h+
Lan-parity stack: infra/lan-parity-test/docker-compose.yml
                  - ai-memory-lan-parity-alice  (127.0.0.1:19180->19077)
                  - ai-memory-lan-parity-bob    (127.0.0.1:19181->19077)
                  - ai-memory-lan-parity-pg-age (127.0.0.1:15432->5432)
                  PG16 + AGE 1.6.0 + pgvector 0.8.2, schema v51
Env vars:         AI_MEMORY_TEST_POSTGRES_URL=postgresql://aimemory:****@127.0.0.1:15432/aimemory
                  AI_MEMORY_TEST_AGE_URL=postgresql://aimemory:****@127.0.0.1:15432/aimemory
Schema:           v51 (sqlite ladder + postgres ladder both at migrate_v51();
                  v50 #1156 K8 quota PK extension; v51 #1255 / PR #1296
                  federation_nonces durable peer-replay nonces)
Test invocation (Track C):
  cargo test --features sal,sal-postgres --release --no-fail-fast
Authoring agent:  Claude (Opus 4.7, 1M context)

Methodology

The campaign exercised six tracks under one matrix:

1. Track A — Build + 1hr dogfood loop. Binary built from release/v0.7.0 HEAD; install symlink updated; live MCP daemon PID 10338 observed across a 1h00m window (2026-05-28T11:44:53Z → 12:44:53Z) and extended to 16h+ sustained. RSS profile flat at ~18 MB at every probe — no leak.
2. Track B — A2A in-host via lan-parity Docker. Rebuilt 3- container stack against integrated HEAD; alice + bob + pg-age all UP healthy on loopback; federation paths exercised via cargo test --features sal,sal-postgres against the shared pg-age URL.
3. Track C — Postgres + AGE full regression. Full sweep at --no-fail-fast against the lan-parity pg-age; 8,028 / 9 / 27 /
   1. Triage: 5 of 9 = cargo-target races (environmental); 4 of 9 = real test-isolation defects → #1381 → fixed in PR #1382 via new tests/common/postgres_env.rs per-test schema isolation helper.
4. Track D — Docs + Pages drift remediation. Codegraph-driven detection (41 sites / 23 files initial) → 3 commits on PR #1379 (cumulative ~54 sites / ~24–27 files) → 3 codegraph-driven QC audits (Audit A literal, Audit B structural, Audit C fault- injection). Closes #1197 + #1198.
5. Track E — CoALA prior-art citation. PR #1380 scope- disciplined: 3 deliverables (ROADMAP citation, positioning section, new docs/strategy/coala-mapping.md), 3 rejected proposals (§2.8 subsection, §11.4.D/§22 reframing, capabilities-v3 coala block). Substrate untouched.
6. Track F — AI-NHI assessment v3. Parallel-dispatched fresh Opus 4.7 agent against post-fix HEAD. Report expected at docs/v0.7.0/heterogeneous-ai-nhi-assessment/report-claude-opus-4-7-v3.md.

Why three QC audits on Track D

The Track-D drift remediation is the highest-volume mechanical correction the campaign did (~54 sites across ~24–27 files). One detection pass is insufficient because:

• Literal enumeration (Audit A) catches stale numeric literals at scale but misses sites where the number appears as part of a larger expression or in a less-obvious context (e.g., the feature-matrix.html:249-250 pill row that contradicted the banner 4 lines above — a literal grep finds both occurrences, but the consistency check requires structural awareness).
• Structural call-graph (Audit B) traverses from each canonical-source symbol to every doc location that should reference it; catches the cases where a doc page should mention a canonical count but doesn’t (or mentions it in a non- literally-greppable form).
• Fault-injection (Audit C) verifies the enforcing contracts: inject a deliberate drift into the source-side pin tests, confirm each one fails loudly. This is the load-bearing-vs-decorative test: a pin test that doesn’t fail under deliberate drift is worse than no pin test, because it gives false confidence.

The 3-audit discipline surfaced 13 sites that the initial pass missed (Audit A: 6; Audit B: 7) — that’s a 32% miss rate on the single-pass detection. The 3-audit closure is what enables the ZERO-DEFECTS-CONFIRMED verdict.

Per-issue root-cause table

The 4 #1381 sub-defects

1. embedding_dim_migration::auto_migrate_converts_384_schema_to_768_on_daemon_bootstrap
2. issue_1213_atttypmod_age_schema_scope::issue_1213_unscoped_probe_demonstrates_root_cause
3. issue_1213_atttypmod_age_schema_scope::issue_1213_atttypmod_probe_scopes_to_public_schema
4. migrate_links_roundtrip::migrate_links_sqlite_to_postgres_to_sqlite_roundtrip

All four share one root cause: tests assume an empty / known schema state on the postgres container but other tests in the same cargo test --no-fail-fast invocation, AND the long-lived ic_alice / ic_bob daemon schemas from the lan-parity stack, accumulate state. CI’s Postgres-feature gate runs --test-threads=1 against a one-shot container per job so the issue doesn’t manifest there; only the LAN-parity shared-container path surfaces it.

Why no substrate code touched

The PR #1382 fix is entirely on the test side: a new tests/common/postgres_env.rs (421 LOC) ships a PostgresTestEnv helper that, per test invocation, runs CREATE SCHEMA test_<name>_<uuid8> against the base postgres URL and returns a per-test connection URL with ?options=-c%20search_path=<schema>,public set. Drop impl cleans up. Three substrate sites where n.nspname = 'public' is hardcoded (src/store/postgres.rs:770, :2788, :3217) are documented in the helper’s module doc but NOT modified — the hardcoded scope is intentional for the substrate’s own dim probe; the test-side fix mirrors it where needed.

This is the right shape: the bug is at the test-discipline layer (tests not isolating their own state), not at the production-code layer. Production daemons running against a real customer postgres do not accumulate test state because they are not run alongside other tests; the failure mode is specific to the test-runner

• shared-container combination.

The meta-pattern: test-isolation discipline at the realistic-deployment substrate

The 2026-05-22 campaign’s meta-pattern was “substrate evolution outruns test-fixture pin discipline” (20 of 22 issues were fixture-drift, fixed by re-pinning). The 2026-05-28 campaign’s meta-pattern is the next layer in: test-isolation discipline at the realistic-deployment substrate.

The 4 #1381 sub-defects are not pin-drift — the pins are correct. They are tests that work fine in CI (one-shot container per job) but fail under the realistic-deployment substrate (a long-lived shared container with multiple daemons connected). The fix is not a pin update — it is a new helper that ensures each test gets its own schema namespace.

This is meaningful because the lan-parity stack IS the realistic deployment substrate: a single managed-postgres instance backing multiple agents is the v0.7.0 “managed-postgres” story. Test fixtures that only work against CI’s isolated single-binary container would have shipped a v0.7.0 with a known test-discipline debt; the campaign caught it and closed it.

Future-bug prevention — proposed gates

Two CI gates would close the recurrence vector for the 2026-05-28 defect class:

1. Add the lan-parity stack to the CI test matrix. Today, CI runs one-shot containers; promote a “lan-parity-shared-container” job to the matrix that runs the full test sweep against a long-lived dual-daemon backend. Cost: ~10 min CI wall-clock per job; gated by feature flags. This would have caught the 4 #1381 sub-defects at PR time, not at ship time.

2. doc-drift-detection workflow (v0.8). Per the Track-D Audit C latent-drift-risk finding: the doc sites themselves have no enforcing CI pin test. A v0.8 workflow scopes a CI step that greps docs/**.md + docs/**.html for canonical-count literals and validates against codegraph_search of the canonical-source symbols. Adds a consistency check across pages (so the 73 routes vs. 87 registrations framing across two pages cannot drift relative to each other).

Both gates are v0.7.x / v0.8 follow-ups, not v0.7.0 SHIP blockers.

Gate-matrix evidence at HEAD be3347d70

QC trail

The 2026-05-28 campaign’s QC discipline shifted from orchestrator-safeguard C1–C8 enumeration (2026-05-22 style — apt for the 22-issue-batch shape) to 3 codegraph-driven audits (apt for the Track-D drift-sweep shape):

• QC Audit A (literal enumeration). Independent codegraph- driven literal enumeration of every drift-vulnerable surface. Found 6 sites missed in the initial pass; verdict was REMAINING-VIOLATIONS-FOUND on round 1. Cleared on round 2.
• QC Audit B (structural call-graph). Codegraph traversal from each canonical-source symbol to every doc location. Found 7 additional sites concentrated in the compliance/inventory bundle.
• QC Audit C (regression-invariance fault-injection). Deliberate-drift injection into the source-side pin tests; confirmed each fails loudly. ZERO-DEFECTS-CONFIRMED on load-bearing pin tests. Flagged latent-drift-risk for v0.8 doc-drift-detection workflow.

All three audits APPROVED at final state. No HARD-BLOCK fails on the docs+Pages remediation. The Track E (CoALA citation) audit returned ZERO-DEFECTS-CONFIRMED on the first pass (the scope discipline made multi-pass auditing unnecessary).

Postgres + AGE cross-store parity at HEAD be3347d70

Lessons learned, in priority order

1. Test isolation at realistic-deployment substrates is a distinct discipline from test-fixture pin maintenance. The 2026-05-22 campaign closed pin-drift; the 2026-05-28 campaign closed isolation-discipline at the lan-parity shared-container substrate. v0.7.x / v0.8 follow-up: promote the lan-parity stack to the CI matrix so this discipline is mechanically enforced.

2. Doc pins need their own enforcing CI gate. The 3-audit ZERO-DEFECTS-CONFIRMED on Track D is fragile against the next substrate count bump — a future PR that bumps the source count
   • updates the source pin test + forgets the docs would slip. v0.8 doc-drift-detection workflow scoped.

3. Three codegraph-driven audits with different lenses are meaningfully better than one. The 32% miss rate on the initial Track-D detection pass would have been customer-visible docs drift had the campaign closed after one pass. The structural lens (Audit B) caught a class of defects that the literal lens (Audit A) cannot see; the fault-injection lens (Audit C) catches enforcement gaps that no detection pass can.

4. Scope discipline on adjacent-temptation PRs is a CI-saving discipline. The Track-E (CoALA citation) PR rejected 3 plausible-sounding adjacent enhancements (§2.8 subsection, §11.4.D/§22 reframing, capabilities-v3 block). Each rejected enhancement was a substantive design change that should travel under its own work order. The 3-deliverable / 3-rejection contract is the right shape for citation-discipline PRs.

5. 16h+ sustained dogfood with flat RSS is a load-bearing signal for substrate stability. The 1hr window was the operator-set minimum; the extended 16h+ sample is the meaningful one. The PRAGMA-tuned SQLite footprint + HNSW async-rebuild double-buffer
   • Once-gated config-load WARN combine to give a flat memory curve under live load. Any future regression in those three subsystems would show as RSS growth in this window.
6. Substrate-untouched fix PRs are the right shape when the bug is test-discipline. PR #1382 closes 4 deterministic test failures via a new helper module; zero src/** change. The discipline: when a defect manifests at the test substrate, first ask “is this a test-isolation defect or a substrate defect?” The 4 #1381 sub-defects were test-isolation; the fix correctly lived at the test layer.

Recommendation

SHIP. The release-gate Tier-1, Tier-2 (sqlite + A2A in-host via lan-parity + postgres+AGE post-#1382), Tier-3, Tier-5, and Tier-6 (cargo audit + four gates + release-notes) are all GREEN at HEAD be3347d70 (with PR #1379 + PR #1380 + PR #1382 pending operator merge). The Tier-2 Track D cross-node + Tier-6 24h dogfood loop remain operator-gated per #836; they are not engineering- completable.

The 5-issue campaign closure (#1197 + #1198 closed via PR #1379; 4 #1381 sub-defects closed via PR #1382; #1378 open + non-blocking) validates the testing-loop discipline (pm-v3.3) — every defect surfaced was filed at discovery, traced through fix → audit → re-fix → audit → close, with substrate kept untouched where the defect was test-side.

The Track-F (AI-NHI assessment v3) report, when it lands, will provide an independent fresh-eyes structural review of the post-fix HEAD as an additional confidence-building data point; the SHIP-CLEARED verdict on the 5 in-host tracks does not depend on it.

Audit trail

Apache-2.0, © 2026 AlphaOne LLC. Authored autonomously by Claude (Opus 4.7, 1M context).