test(daemon): run the store suites on real PostgreSQL

[zfy0701]

Summary

Closes #1073.

The daemon pool opens LocalStore over PostgresSyncDatabase, but every store test ran
against SQLite only, so a SQLite-only construct in local-store.ts reached a cluster
before anything noticed — #1068 caught a two-argument MAX(x, y) in review, not in CI, and
postgres-pool-store.int.test.ts was gated on DATA_PLANE_TEST_DATABASE_URL and never ran.

This adds the missing coverage in two layers: a Vitest project that re-runs the store
suites against a real postgres:16-alpine, and a text check over the SQL itself for the
statements no suite reaches.

How the PostgreSQL project works

vitest.postgres.config.ts is a separate config file, not a second project inside
vitest.config.ts, so vitest run and every targeted vitest run <file> already in CI
stay Docker-free.

• test/store-postgres/global-setup.ts boots one postgres:16-alpine via
  Testcontainers and creates one database per Vitest worker — the same shape as the
  control-plane integration project.
• test/store-postgres/setup.ts opens one PostgresSyncDatabase per worker on that
  worker's database, lets the first LocalStore materialize the schema through the real
  SQLite→PostgreSQL rewrite, releases the schema advisory lock, and truncates every store
  table (plus rewinds the revision sequence) before each test.
• test/store-postgres/backend.ts is the seam. Unarmed — the ordinary SQLite run — it
  reports usingPostgresStore() === false and the suites build their node:sqlite stores
  exactly as before. Armed, openPostgresLocalStore() hands them a LocalStore over the
  worker's pool store, with a borrowed handle so a suite's close() cannot take the
  worker-wide connection down.

Suites covered: local-store, memory-capture-outbox, postgres-pool-store.int,
postgres-transcript-org.int — 99 tests, 10 skipped. The skips are the suites that are
about SQLite internals and nothing else: schema versioning through PRAGMA user_version,
the v10 transcript-org migration replay, and two cases that reach for DatabaseSync
directly to stage a race.

The gated .int suites now run for real: the setup file points
DATA_PLANE_TEST_DATABASE_URL at the worker's database.

ci: gate the daemon store suites on real PostgreSQL adds a Daemon Store (PostgreSQL)
job on its own Docker-bearing runner, off the critical path, like the control-plane
integration shards.

What it caught

The activation rendezvous (send-message-routing-rework.md §3.2/§8.6) joined its composite
key and its transcript coordinates with NUL, and PostgreSQL TEXT rejects 0x00
outright: invalid byte sequence for encoding "UTF8": 0x00. On the pool every
claimActivationObservation / attachActivationEnvelope write threw, so paired activation
— the agent-to-agent and relay-forwarded hand-off — could not settle at all. Eleven tests
failed on the first PostgreSQL run for that one reason.

The fix is portable, in the key builder rather than the worker: the separator is now the
unit separator (U+001F). Those rows are TTL-bounded pairing records, so no stored key
outlives the change.

The cheap half

test/local-store-sql-portability.test.ts parses local-store.ts, pulls out every
SQL-shaped string and template literal, and fails on constructs the pool worker does not
rewrite: two-argument MAX/MIN, IFNULL, IIF, datetime/strftime/julianday/
unixepoch, INSERT OR REPLACE|ABORT|FAIL, GROUP_CONCAT, printf, TYPEOF, ||
concatenation and comma LIMIT. A statement may opt out with a -- pg-portable-exempt:
marker and a reason. It runs in the ordinary unit job, so a statement no suite covers still
fails fast, and it carries its own self-check so the patterns cannot rot into matching
nothing.

What the worker already translates — INSERT OR IGNORE, BEGIN IMMEDIATE,
INTEGER PRIMARY KEY AUTOINCREMENT, PRAGMA user_version, sqlite_master,
LIMIT -1 OFFSET, length(CAST(x AS BLOB)) — stays legal, and the check's header says the
list grows by making SQL portable, never by teaching the worker another function name.

Test plan

• pnpm --filter @agentconnect.md/daemon test:store:postgres — 99 passed, 10 skipped,
  ~23s after the container boots.
• pnpm --filter @agentconnect.md/daemon exec vitest run test/local-store.test.ts test/memory-capture-outbox.test.ts test/local-store-permissions.test.ts test/local-store-sql-portability.test.ts test/daemon-agent-mention-routing.test.ts test/postgres-*.test.ts — the SQLite run, unchanged: 151 passed.
• pnpm --filter @agentconnect.md/daemon test — the pre-existing local failures
  (acp-matrix live runtimes, the shim-* macOS sandbox suites,
  workspace-git-runner-seam) reproduce on a clean checkout of main and are untouched
  by this change.
• pnpm --filter @agentconnect.md/daemon typecheck, pnpm lint, pnpm format:check.

[agentconnect-md-test]

No blocking findings on a0b250710a49e9a050a612f60481d91a6632fdf5.

The PostgreSQL harness keeps each Vitest worker on a separate database, releases the schema advisory lock before the gated pool-store suites open their own connections, and prevents suite-level LocalStore.close() calls from closing the worker-wide backend. The activation rendezvous separator change is applied consistently to the production key/coordinate builders and their fixtures, and remains an opaque durable value as required by the rendezvous design.

Validation on this exact revision is green: Build, Check, Unit Test, Daemon Store (PostgreSQL), both integration shards, and Sandbox (Linux). The new PostgreSQL job reports 4 files passed with 99 tests passed and 10 SQLite-specific skips.

_{sent by review-bot (Codex · gpt-5.6-sol) · open in session}