From 4caea21098d0d85e1338d516599aca50314b79d6 Mon Sep 17 00:00:00 2001
From: "glm-5.3-flash" <glm-5.3-flash@alk.dev>
Date: Thu, 8 Oct 2026 22:49:00 +0000
Subject: [PATCH] =?UTF-8?q?Wave=204=20decomposed:=20Postgres=20engine=20?=
 =?UTF-8?q?=E2=80=94=2011=20tasks=20(schema,=20open/opts,=20seam,=20forwar?=
 =?UTF-8?q?der,=20mechanisms,=20scheduler/outbox,=20integration,=20review?=
 =?UTF-8?q?=20gate);=20plan=20synced?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 docs/plans/implementation.md        |  90 +++++++++++++++-
 tasks/pg-engine-integration.md      | 101 ++++++++++++++++++
 tasks/pg-engine-locks.md            |  93 +++++++++++++++++
 tasks/pg-engine-notify-listen.md    | 139 +++++++++++++++++++++++++
 tasks/pg-engine-open-opts.md        | 113 ++++++++++++++++++++
 tasks/pg-engine-queues.md           | 141 +++++++++++++++++++++++++
 tasks/pg-engine-scheduler-outbox.md | 154 ++++++++++++++++++++++++++++
 tasks/pg-engine-schema.md           | 110 ++++++++++++++++++++
 tasks/pg-engine-seam-tx.md          | 129 +++++++++++++++++++++++
 tasks/pg-engine-streams.md          | 126 +++++++++++++++++++++++
 tasks/review-wave-4.md              |  94 +++++++++++++++++
 11 files changed, 1287 insertions(+), 3 deletions(-)
 create mode 100644 tasks/pg-engine-integration.md
 create mode 100644 tasks/pg-engine-locks.md
 create mode 100644 tasks/pg-engine-notify-listen.md
 create mode 100644 tasks/pg-engine-open-opts.md
 create mode 100644 tasks/pg-engine-queues.md
 create mode 100644 tasks/pg-engine-scheduler-outbox.md
 create mode 100644 tasks/pg-engine-schema.md
 create mode 100644 tasks/pg-engine-seam-tx.md
 create mode 100644 tasks/pg-engine-streams.md
 create mode 100644 tasks/review-wave-4.md

diff --git a/docs/plans/implementation.md b/docs/plans/implementation.md
index 822763b..3fe0a51 100644
--- a/docs/plans/implementation.md
+++ b/docs/plans/implementation.md
@@ -1,6 +1,6 @@
 ---
 status: draft
-last_updated: 2026-10-08 (wave 3 implemented + reviewed; review-gate fixes: substrate boundary move, writer-slot error-arm release; wave 4/5 decomposition may proceed)
+last_updated: 2026-10-08 (wave 4 decomposed — 11 tasks; waves 1–3 implemented + reviewed)
 ---
 
 # alkstore — Implementation plan
@@ -35,7 +35,7 @@ parallel).
 | [1](#wave-1--foundations) | Workspace scaffold; core crate (errors, value types, full trait surface); contract-suite scaffold | — | implemented + reviewed (2026-10-08) |
 | [2](#wave-2--sqlite-substrate-fork) | honker-core fork into `alkstore-sqlite/src/substrate/`: port, deltas, provenance, test floor | wave 1 (workspace scaffold only) | implemented + reviewed (2026-10-08) |
 | [3](#wave-3--sqlite-engine) | SQLite engine: connection architecture, re-derived queue ops on contract v1, scheduler/outbox, tx seam, SQLite backlog column | waves 1 + 2 | implemented + reviewed (2026-10-08) |
-| 4 | Postgres engine: schema bootstrap, pool/open, listener/forwarder, all mechanisms, tx seam, pg backlog column | wave 1 | not yet decomposed |
+| 4 | [Postgres engine](#wave-4--postgres-engine): schema bootstrap, pool/open, listener/forwarder, all mechanisms, tx seam, pg backlog column | wave 1 | decomposed (11 tasks) |
 | 5 | Contract suite: the cross-engine equivalence properties (core-contract.md §Verification backlog), version-stamped per ADR-017 | waves 3 + 4 | not yet decomposed |
 | 6 | Release readiness: crate docs, deployment matrix final pass, README (written last, honestly), publish prep; mem-engine and fuzzing decisions | wave 5 | not yet decomposed |
 
@@ -101,6 +101,86 @@ shaped this wave's task set:
   panic probe → wave 5's contract-suite rows; N-1's growth-posture
   review → wave 6.
 
+## Wave 4 — Postgres engine
+
+The second engine: the core contract implemented on tokio-postgres +
+deadpool-postgres per engine-postgres.md — schema bootstrap (one
+engine-owned schema, default `alkstore`, ADR-010 §8), the pool/open
+constructor, the hand-rolled LISTEN forwarder (dedicated non-pooled
+listener connection, bounded broadcast fanout, exponential-backoff
+reconnect with re-LISTEN, the synthetic reconnect-wake on
+`__alkstore_listener_reconnected__`), the full `Store`/`TxHandle`/
+mechanism-handle trait impls (natively async — no `spawn_blocking`
+seam; the tx handle holds the pooled object directly, ADR-007's pg
+arm), the re-derived queue machinery (`FOR UPDATE SKIP LOCKED` claims,
+LISTEN-driven wake with re-poll safety net), the scheduler leader loop
+and outbox helper, and the engine's backlog column in the contract
+suite (factory-parameterized; wave 5 runs the same rows against both
+engines).
+
+The wave is deliberately shaped as wave 3's structural twin — same
+task rhythm (constructor → seam → mechanisms in parallel →
+scheduler/outbox → integration → review gate) — because the two
+engines implement the same contract and the wave-3 task set's
+boundaries proved clean. The deltas are the engine's own: no
+substrate (the pg machinery is re-derived greenfield per ADR-004/005,
+with the pg-boss schema family as design reference), a natively-async
+seam (Send+Sync client, POC #2 compile-probe), the forwarder as the
+wake substrate (with its two POC-pinned deadlock pitfalls owned as
+test-pinned failure modes), and the `PayloadTooLarge` client-side
+check (this engine *produces* the variant — the one runtime carriage
+of an engine asymmetry, ADR-016 §5).
+
+Wave-3 outcomes absorbed into the task set:
+
+- **The `#[doc(hidden)]` constructors** (`Job::from_row`,
+  `StreamEvent::from_row`, `Schedule::new`, `Wake::new`) already
+  exist — wave 3's pre-work task built them engine-agnostically; wave
+  4 consumes them as-is (no wave-4 equivalent needed).
+- **The engine-side arithmetic is re-owned, not shared**: the
+  equal-jitter backoff curve, the opts-resolution rules
+  (delay-over-`run_at`, relative-`expires`, the derived stamp sets
+  300/3/5/none and 60/5/5), and the outbox backing-queue derivation
+  are each engine's own implementation (ADR-012 §2's one-owner rule
+  applied per engine); the pg tasks re-implement them from the ADR
+  text, and the contract suite (wave 5) pins the two engines'
+  arithmetic to identical outputs. The SQLite engine's `resolution.rs`
+  is reference, not a dependency.
+- **The `@every` spec grammar has no shared parser on this side** —
+  the SQLite engine reuses the substrate's `parse_every_interval`
+  (one-line delta D-31's neighbor); the pg engine owns its parser
+  (the grammar is small and pinned: `@every <n><unit>`, `s|m|h|d`).
+  Equivalence of *behavior* (what fires, when) is suite-pinned; the
+  parsers themselves are per-engine.
+- **The wave-3 review's deferred notes** (reader-pool close path,
+  commit-error-arm coverage) stay SQLite-scoped — the pg engine's
+  pool is deadpool-managed (no hand-rolled close path to review) and
+  its commit-error arm gets coverage in the pg tasks' own tests.
+- **The suite's factory contract** (isolation + idempotent teardown)
+  is what the pg factory implements — a fresh schema per `open` per
+  the SQLite factory's precedent (fresh temp file there); the
+  POC's shared-server parallel-interference caveat (findings,
+  invocation note) is answered by exactly that isolation, not by
+  sequential-only harnesses.
+
+Test posture: the engine's tests need a live Postgres server. The
+harness convention is the POC's (dockerized `postgres:16-alpine` on
+:15432, the `pglo-poc` container — still running in the dev
+environment); connection settings ride environment/config, never
+hardcoded. Tests that cannot reach a server skip cleanly (the
+`SKIP`-posture), so the workspace gates stay green server-less — but
+the wave's acceptance requires the suite green *against* the server.
+
+One design point the architecture leaves to implementation (pinned in
+the wave-4 tasks, flagged here for visibility): **the queue/stream
+wake-channel naming**. The contract pins LISTEN as the pg wake
+trigger and pins the channels/queues shared namespace, but no doc
+names the channel an enqueue/publish notifies on. The tasks pin the
+natural realization — the mechanism name *is* the wake channel
+(`listen(queue_name)` receives that queue's enqueue wakes; same for
+streams), legal by construction under the shared namespace — with
+the reserved reconnect-wake channel staying the only reserved string.
+
 ## Decided points
 
 - **Contract-suite layout — option (a)**: a small internal
@@ -159,4 +239,8 @@ decomposes. Specific gates:
   `open_writer_connection` boundary move out of `substrate/mod.rs`;
   the writer-slot error-arm stranding in
   `with_writer`/`begin`/`commit`); two minor notes deferred to
-  wave 5/6 (reader-pool close path, commit-error-arm coverage).
\ No newline at end of file
+  wave 5/6 (reader-pool close path, commit-error-arm coverage).
+- **Wave 4 decomposition** (2026-10-08) — shaped as wave 3's
+  structural twin; wave-3 outcomes absorbed (constructors exist;
+  arithmetic re-owned per engine; per-engine `@every` parser;
+  SQLite-scoped deferred notes stay put). See the Wave 4 section.
\ No newline at end of file
diff --git a/tasks/pg-engine-integration.md b/tasks/pg-engine-integration.md
new file mode 100644
index 0000000..35b12d7
--- /dev/null
+++ b/tasks/pg-engine-integration.md
@@ -0,0 +1,101 @@
+---
+id: pg-engine-integration
+name: Postgres engine — integration (full-surface wiring, suite adoption, crate docs)
+status: pending
+depends_on: [pg-engine-notify-listen, pg-engine-streams, pg-engine-queues, pg-engine-locks, pg-engine-scheduler-outbox]
+scope: moderate
+risk: medium
+impact: phase
+level: implementation
+tags: [wave-4, postgres-engine, integration]
+---
+
+## Description
+
+Close the wave's integration gaps once every mechanism is wired — the
+structural twin of wave 3's integration task (no lint-suppression
+removal on this side: the pg engine is greenfield, no `#![allow]`
+posture ever existed — the sweep below is the same discipline without
+that step):
+
+- **Full-surface sweep**: every `Store`/`TxHandle`/mechanism trait
+  method implemented (no `todo!`/`unimplemented!`/error-stub
+  remaining); the engine crate's lib re-exports its public surface
+  (`open`, `PgOpts`, the concrete store type) per the module-per-file
+  convention. Genuinely dead code must be **cut, not suppressed** —
+  clippy `-D warnings` green without allows.
+- **Contract-suite adoption** (the plan's "waves 3 and 4 adopt the
+  harness"): implement `StoreFactory` for the pg engine — **a fresh
+  schema per `open`** (the SQLite factory's fresh-temp-file
+  precedent; the POC's shared-server parallel-interference caveat is
+  answered by exactly this isolation — each property's rows live in
+  its own schema, so parallel property runs never observe one
+  another), idempotent teardown (`DROP SCHEMA … CASCADE`; safe
+  against an abandoned store per the factory contract's teardown
+  posture). Pin the factory's isolation/idempotence contract (the
+  SQLite factory's pin shape).
+  Wire the engine's backlog column — the rows this engine owns now,
+  running against the pg factory:
+  - The ADR-023-stamped rows (extent-clamp, duration-refusal,
+    `encode_payload` round-trip) — already written suite-side,
+    factory-parameterized; they run against pg by this task's
+    factory existing (verify, don't rewrite).
+  - **The pg arm of the engine-scoped rows**: `PayloadTooLarge`
+    produced (the pg rejection arm — oversized `notify`/`notify_tx`
+    rejected client-side with the limit in the variant; the row text
+    the SQLite integration task left for "wave 4's adoption"),
+    drop = rollback no-ghosts, in-tx read-your-own-writes,
+    enqueue-opts resolution, receiver close arms (the pg arm: stays
+    open across reconnects, closes at shutdown — the row's
+    engine-asymmetric legs may need the row split or parameterized
+    per the suite's convention; implementer's choice, documented).
+  - The SQLite-scoped rows (plain-path open §3, poll cadence §4)
+    simply don't run against pg — the SQLite integration task's
+    Notes already pinned that disposition.
+- **Crate docs**: the engine crate's lib-level doc comments state the
+  multi-host posture and the listener budget line (ADR-016,
+  deployment.md) — the identity statements this crate owns (the open
+  task drafted them; this task verifies they reflect the finished
+  engine).
+- **Gates**: full workspace build/test/clippy/fmt (the pg tests skip
+  cleanly server-less but must be green against the harness server);
+  coverage spot-check on the engine crate (no large uncovered regions
+  outside error arms — the wave-3 gate's bar).
+
+## Acceptance Criteria
+
+- [ ] No `todo!`/`unimplemented!`/error-stub in the engine crate;
+      full trait surface implemented; clippy `-D warnings` green
+      without allows
+- [ ] `StoreFactory` implemented for pg (fresh schema per `open`,
+      idempotent CASCADE teardown); the factory's
+      isolation/idempotence contract pinned
+- [ ] The backlog column runs green against the pg factory: the three
+      ADR-023 rows verified (not rewritten), the pg arm of the five
+      engine-scoped rows present + stamped
+- [ ] The `PayloadTooLarge` pg arm row present (the SQLite
+      integration task's deferred adoption discharged)
+- [ ] Crate lib docs carry the multi-host + listener-budget posture
+      statements, reflecting the finished engine
+- [ ] Workspace gates green: `cargo build`, `cargo test` (incl.
+      against the harness server), clippy `-D warnings`, fmt clean
+- [ ] Coverage spot-check: no large uncovered regions outside error
+      arms
+
+## References
+
+- docs/plans/implementation.md (Wave 4 section)
+- docs/architecture/core-contract.md (§Verification backlog)
+- docs/architecture/decisions/022-contract-suite-layout.md
+- docs/architecture/decisions/023-fourth-review-round.md (backlog additions)
+- docs/architecture/decisions/016-deployment-honesty.md §5
+- alkstore-sqlite/tests/contract_suite.rs (the factory precedent)
+- alkstore-contract-suite/src/properties.rs (the rows to run/adopt)
+
+## Notes
+
+> To be filled by implementation agent
+
+## Summary
+
+> To be filled on completion
\ No newline at end of file
diff --git a/tasks/pg-engine-locks.md b/tasks/pg-engine-locks.md
new file mode 100644
index 0000000..1f8bcad
--- /dev/null
+++ b/tasks/pg-engine-locks.md
@@ -0,0 +1,93 @@
+---
+id: pg-engine-locks
+name: Postgres engine — named locks (`try_lock`, `Lock` handle, duration guards)
+status: pending
+depends_on: [pg-engine-seam-tx]
+scope: narrow
+risk: low
+impact: component
+level: implementation
+tags: [wave-4, postgres-engine]
+---
+
+## Description
+
+Implement the named-lock mechanism's pg arm — the smallest mechanism,
+and the one with the strongest POC ground (the pg POC's lock probe
+pinned acquire/release/renew with TTL and expiry re-acquisition
+directly). Over the schema task's locks table:
+
+- **`try_lock(name, owner, ttl)`** — shared-namespace + non-empty
+  owner validation (`InvalidName`); the **duration guard** (`ttl <=
+  0` → opaque `Err(Database)`, detail in the source chain — ADR-023
+  §2's duration kind; an exclusion mechanism must not silently hand
+  back a lease the caller believes it holds); the closed-store check;
+  then the acquire op on a pool connection: opportunistic-expiry-
+  delete + insert-or-reacquire + read-back (the substrate's
+  `lock_acquire` semantics, re-derived — the POC's probe shape).
+  `Ok(None)` = held elsewhere (no-work is a value); `Ok(Some(boxed
+  Lock))` on grant.
+- **`Lock` handle** (`renew`, consuming `release`):
+  - `renew(ttl)` — the same duration guard; the new **full TTL
+    window from now** (not additive); `false` = lost it (expired and
+    re-acquired by another owner — the owner-scoped update's 0-rows).
+  - `release(self: Box<Self>)` — consuming; the owner-scoped delete;
+    `true` = was held (row gone by this owner's hand), `false` =
+    row-less or foreign (the substrate's row-scoped truth-teller
+    posture — no handle-local "held" flag caching).
+- **Guarantee row posture** (ADR-008 §7): mutual exclusion bounded by
+  TTL + renewal; expiry lapses *silently* (no revocation event —
+  expired rows are deleted opportunistically by the next acquire's
+  expiry-delete, no general expiry sweeper — the inherited posture,
+  documented); re-acquirable after expiry (the POC-pinned property —
+  pin it engine-side here; the contract-suite row remains for the
+  suite itself).
+- **Same-owner re-acquire**: pin the inherited disposition (the
+  substrate's shape: the same owner's re-acquire keeps the original
+  row, TTL *not* refreshed, reports granted — the SQLite task's
+  pinned posture; engine-sqlite.md's mapping row calls it inherited
+  deliberately, and the pg arm should match it for cross-engine
+  equivalence — flag in Notes if the pg realization differs).
+- **Post-lapse release arms**: pin both (the stale holder's release
+  deletes its own stale row → `true`; after a foreign re-acquire →
+  `false`) — the SQLite task's pinned shape, matched here.
+- **Closed-store posture**: `try_lock`/`renew`/`release` fail closed
+  with opaque `Database` (the engine-wide shape).
+- **No `lock_tx`** — deliberate absence (core-contract.md's named-
+  locks section); nothing to build, just don't invent one.
+
+## Acceptance Criteria
+
+- [ ] `try_lock`: validation (name + owner), duration guard
+      (`ttl <= 0` → opaque `Err`, detail in the source chain),
+      closed-store check, then acquire — `Ok(None)` on held-elsewhere
+- [ ] Exclusion + loser shape pinned (two owners, one wins); expiry
+      re-acquisition pinned (the POC's property, engine-side)
+- [ ] `renew`: duration guard; full-window reset; `false` on
+      lost/expired-and-reacquired
+- [ ] `release`: consuming, owner-scoped, both boolean arms pinned
+      (stale-row delete = `true`; foreign/row-less = `false`)
+- [ ] Same-owner re-acquire disposition pinned (and flagged if it
+      differs from the SQLite arm's)
+- [ ] Silent lapse posture documented (no revocation event; no
+      general expiry sweeper)
+- [ ] `cargo test -p alkstore-postgres` (harness server), clippy
+      `-D warnings`, fmt clean; gates green server-less
+
+## References
+
+- docs/architecture/core-contract.md (named locks section)
+- docs/architecture/decisions/008-contract-v1-pinning.md §7 (guarantee row)
+- docs/architecture/decisions/023-fourth-review-round.md §2 (duration kind)
+- docs/architecture/engine-postgres.md (Mapping the contract: named locks)
+- docs/research/poc-pg-posture-findings.md (the lock probe row)
+- alkstore-sqlite/src/lock.rs (the structural twin — match its pinned dispositions)
+- alkstore/src/lock.rs (the trait)
+
+## Notes
+
+> To be filled by implementation agent
+
+## Summary
+
+> To be filled on completion
\ No newline at end of file
diff --git a/tasks/pg-engine-notify-listen.md b/tasks/pg-engine-notify-listen.md
new file mode 100644
index 0000000..bb33368
--- /dev/null
+++ b/tasks/pg-engine-notify-listen.md
@@ -0,0 +1,139 @@
+---
+id: pg-engine-notify-listen
+name: Postgres engine — LISTEN forwarder (full behavior) + notify/listen surface
+status: pending
+depends_on: [pg-engine-seam-tx]
+scope: moderate
+risk: high
+impact: component
+level: implementation
+tags: [wave-4, postgres-engine, forwarder]
+---
+
+## Description
+
+Complete the LISTEN forwarder's behavior (the open task stood up the
+skeleton) and implement `Store::notify` and `Store::listen` — the pg
+arm of the wake contract (ADR-006). The forwarder is hand-rolled
+machinery (ADR-004's decision; `postgres-notify` derive-not-adopted)
+— its failure modes are *learned territory*, POC-pinned; this task
+owns them as passing tests.
+
+**Forwarder completion** (on the skeleton's dedicated non-pooled
+connection + poll loop + bounded broadcast):
+
+- **Reconnect**: exponential backoff 50 ms → 2 s cap (the POC's
+  policy); on each successful reconnect, re-LISTEN from the live
+  channel list (the store's channel registry — `listen` adds to it,
+  receiver drop removes it) and broadcast the **synthetic
+  reconnect-wake** on the reserved channel
+  `__alkstore_listener_reconnected__` (ADR-008 §4's one v1-reserved
+  string — the POC's ad-hoc `__listener_reconnected__` renamed into
+  the namespace) to *every* subscriber's receiver (ADR-006's
+  reconnect-recovery semantics — the no-replay hole's honest
+  communication).
+- **Fanout**: `poll_message`'s notifications fan out per channel —
+  each `listen(channel)` subscriber receives wakes for its channel
+  only (plus the reserved reconnect-wake, which every subscriber
+  receives). Bounded broadcast per subscriber or one shared bounded
+  channel with per-subscriber routing — implementer's choice, but
+  **lag is surfaced, not silent** (the POC's `Lagged(n)` posture:
+  log/record and keep going; a lagged subscriber's missed wakes are
+  the contract's best-effort-hint shape, recovery is re-read).
+- **Wake content**: `Wake { channel }` only — never payloads, never
+  ids (ADR-008 §3). Per-notify delivery (no coalescing — LISTEN
+  delivers each committed NOTIFY; the SQLite engine's coalescing is
+  the documented engine asymmetry).
+- **Receiver close semantics** (ADR-021 §5's pg arm — the documented
+  engine asymmetry): the receiver **stays open across forwarder
+  reconnects** (reconnect-wakes arrive on it; no close) and closes
+  only at engine shutdown (`close()`/Drop → all receivers terminal
+  `None`). This is the inverse of SQLite's death-closes-receivers —
+  pin it by test (kill the listener's backend mid-subscription;
+  receivers stay open; the reconnect-wake arrives; post-reconnect
+  deliveries work).
+- **Identifier quoting**: channel names are quoted identifiers
+  (`LISTEN "{name}"` with `"` doubling — the POC's `quote_ident`
+  shape); validation already rejects empty/reserved, but quoting is
+  what makes arbitrary legal names (colons, etc.) work.
+
+**`Store::notify(channel, payload)`**: auto-commit path — pool
+checkout, `SELECT pg_notify($1, $2)` in its own implicit transaction,
+return to pool. The 8000-byte client-side check (typed
+`PayloadTooLarge` before any round trip) — the seam task built it for
+`notify_tx`; this task wires the auto-commit twin through the same
+check (one owner of the limit constant).
+
+**`Store::listen(channel)`**: validate (shared-namespace kinds —
+`InvalidName`/`ReservedName`), register the channel in the store's
+channel registry (issuing `LISTEN` on the forwarder's connection —
+the dynamic channel set), return the boxed `PgWakeReceiver` bridging
+the fanout:
+- `recv()` → `Some(Wake { channel })` per notification; `None` only
+  at engine shutdown (terminal — never reopens; ADR-021 §5's pg arm).
+- `try_recv` / `recv_timeout` per the core trait's pinned arms
+  (`Closed` distinguishable from idle; timeout expiry = `Ok(None)`).
+- Receiver drop unregisters the channel (issuing `UNLISTEN` when the
+  last subscriber for that channel drops — the channel registry's
+  refcount; zero subscribers ⇒ no LISTEN needed).
+- `listen` on a closed store fails closed (`Database`).
+
+**Wake-arrives honesty**: tests pin wake-arrives (cross-connection
+commit → wake), the reconnect-wake path, and close-at-shutdown — not
+wake-exclusivity (a wake may arrive for unrelated activity on the
+shared channel namespace; the contract's best-effort-hint posture).
+
+**The POC's two deadlock pitfalls, pinned as tests** (they are the
+hand-rolled posture's owned failure modes, ADR-004):
+1. The poll loop runs before the first client query on the listener
+   connection (structurally excluded by the skeleton; pin the
+   ordering with a test that issues a query immediately after
+   listener spawn).
+2. The listener's `Client` stays alive for the listener's lifetime
+   (pin: a long-lived listener survives unrelated client churn).
+
+## Acceptance Criteria
+
+- [ ] Forwarder reconnects with 50 ms → 2 s exponential backoff;
+      re-LISTENs from the live channel set; broadcasts the synthetic
+      reconnect-wake on `__alkstore_listener_reconnected__` to every
+      subscriber after each reconnect — tested via backend kill
+- [ ] No-replay honesty pinned: a commit during a connection gap is
+      never re-delivered (`!saw_replay` posture); recovery is the
+      reconnect-wake + consumer re-read
+- [ ] `notify` delivers commit-atomic auto-commit notifies; the
+      8000-byte check is typed, client-side, before any round trip
+      (both `notify` and `notify_tx` — one limit owner)
+- [ ] `listen` returns a working `WakeReceiver`: per-notify wakes,
+      `Wake { channel }` only, channel-scoped fanout, multi-channel
+      on one listener connection, N subscribers per channel
+- [ ] Receiver close semantics: stays open across reconnects
+      (reconnect-wake arrives), closes only at engine shutdown
+      (terminal `None`); `try_recv`/`recv_timeout` arms match the
+      pinned shapes
+- [ ] Receiver drop unregisters (UNLISTEN at last-subscriber drop)
+- [ ] Channel validation on both entry points; closed-store
+      fail-closed on both
+- [ ] The two deadlock pitfalls pinned by tests
+- [ ] `cargo test -p alkstore-postgres` (harness server), clippy
+      `-D warnings`, fmt clean; gates green server-less
+
+## References
+
+- docs/architecture/engine-postgres.md (Connection architecture: forwarder)
+- docs/architecture/decisions/004-postgres-driver.md (forwarder decision, pitfalls)
+- docs/architecture/decisions/006-wake-and-delivery-contract.md (wake contract, reconnect-wake)
+- docs/architecture/decisions/008-contract-v1-pinning.md §3, §4 (wake type, reserved string)
+- docs/architecture/decisions/016-deployment-honesty.md §5 (PayloadTooLarge)
+- docs/architecture/decisions/021-tx-reads-and-value-shape-fixes.md §5 (receiver arms, pg close semantics)
+- docs/research/poc-pg-posture-findings.md (Sub-module L — the measured ground)
+- alkstore-sqlite/src/notify.rs (the structural twin — inverse close semantics)
+- /workspace/alkstore-pg-posture-poc/src/listen.rs (the POC's ~90-line shape)
+
+## Notes
+
+> To be filled by implementation agent
+
+## Summary
+
+> To be filled on completion
\ No newline at end of file
diff --git a/tasks/pg-engine-open-opts.md b/tasks/pg-engine-open-opts.md
new file mode 100644
index 0000000..e211241
--- /dev/null
+++ b/tasks/pg-engine-open-opts.md
@@ -0,0 +1,113 @@
+---
+id: pg-engine-open-opts
+name: Postgres engine — `open` constructor, `PgOpts`, pool + listener wiring
+status: pending
+depends_on: [pg-engine-schema]
+scope: moderate
+risk: medium
+impact: component
+level: implementation
+tags: [wave-4, postgres-engine]
+---
+
+## Description
+
+Stand up the pg engine's public surface: the `open` constructor,
+`PgOpts`, and the connection architecture of engine-postgres.md's
+"Connection architecture" section — the structural twin of wave 3's
+`sqlite-engine-open-opts`, with the pg deltas.
+
+- `PgOpts` (engine-crate type, ADR-008 §6's config split; NOT
+  `#[non_exhaustive]` — opts structs consumers construct, ADR-017 §3's
+  exemption; `Default` + `Clone`): at minimum
+  - connection string / config (the tokio-postgres `Config`-parseable
+    form — the consumer's server address, credentials; the engine
+    never invents defaults for these, `open` fails `Database` if
+    unparseable/unreachable),
+  - `schema: String` (default `"alkstore"` — ADR-010 §8),
+  - `max_size` (the deadpool pool size; default documented, the
+    deployment matrix's budget line — the listener adds +1 outside
+    the pool),
+  - `synchronous_commit: bool` (default `true` — ship config,
+    ADR-004/deployment.md's measured trade; wired as a per-session
+    SET at connect, POC-verified mechanics).
+- `open(config, opts) -> Result<Box<dyn Store>>` (concrete
+  `PgStore` re-exported, mirroring the SQLite engine's
+  boxed-constructor + pub-concrete shape):
+  - Build the deadpool pool (`RecyclingMethod::Fast` — no `DISCARD
+    ALL` recycling, the POC's zero-error posture).
+  - Run the schema task's bootstrap on a pool connection at open
+    (idempotent; open fails `Database` if bootstrap fails).
+  - Spawn the **forwarder** (the listener task) — the wake substrate
+    every later mechanism task rides. The forwarder is this task's
+    hard part; its full behavior (reconnect, re-LISTEN, reconnect-
+    wake) is the notify-listen task's, but the *skeleton* lives here:
+    a dedicated non-pooled connection (deadpool#360 — pooled
+    connections cannot deliver), the `poll_message` loop fanning out
+    into a bounded broadcast channel (lag surfaced, not silent), the
+    two POC-pinned deadlock pitfalls respected (the poll loop running
+    before the first client query on the listener connection; the
+    `Client` kept alive for the listener's lifetime), and the
+    dynamic channel set (LISTENs re-issued from the live channel
+    list — the notify-listen task adds channels to it).
+  - The store struct holds pool + forwarder handle + schema name
+    behind the `Store` trait; `close()`/`Drop` tear down cleanly
+    (drop the listener connection, close the pool).
+- **Trait stubs**: all `Store` methods return
+  `Err(Database("… wiring lands with the … task"))` — the wave-3
+  posture; the mechanism tasks replace them. `with_tx` surfaces the
+  `begin_tx` stub naturally.
+- **Error mapping posture established here** (the seam task's
+  foundation): tokio-postgres/deadpool errors → `Error::Database`
+  with the source chain preserved (ADR-008 §5's opaque fallback) —
+  the two mappings (`pg_error`, `pool_error` or one helper) the
+  mechanism tasks reuse.
+- **No `spawn_blocking` seam on this engine** — natively async
+  (Send+Sync client, POC #2 compile-probe); the engine crate's tokio
+  dep already carries what's needed. State the posture in the crate
+  docs (the SQLite engine's `# Posture` heading shape: multi-host
+  native per ADR-016, the listener budget line per deployment.md).
+
+Tests against the harness server: open/close round-trip, bootstrap
+ran (tables exist), pool bounds, forwarder skeleton up (a LISTEN
+issued through it delivers — the minimal fanout proof, full forwarder
+behavior in the notify-listen task), unreachable-server `open` fails
+`Database`, opts flow (schema name lands in the created schema's
+name; `synchronous_commit` observable via `SHOW`).
+
+## Acceptance Criteria
+
+- [ ] `PgOpts` with connection config, `schema` (default
+      `alkstore`), `max_size`, `synchronous_commit` (default on);
+      documented; not `#[non_exhaustive]`
+- [ ] `open` boots pool + bootstrap + forwarder skeleton; failure at
+      any step is a typed `Database` error (source chain preserved)
+- [ ] Forwarder skeleton: dedicated non-pooled connection, poll loop
+      before first query, bounded broadcast fanout, dynamic channel
+      set; the two POC deadlock pitfalls structurally excluded
+- [ ] `close()`/`Drop` teardown clean (listener connection dropped,
+      pool closed); post-close ops fail closed
+- [ ] Error-mapping helpers exist and are the documented reuse point
+- [ ] Crate docs carry the `# Posture` statements (multi-host,
+      listener budget, no spawn_blocking)
+- [ ] `cargo test -p alkstore-postgres` (harness server), clippy
+      `-D warnings`, fmt clean; gates green server-less (tests skip)
+
+## References
+
+- docs/architecture/engine-postgres.md (Connection architecture)
+- docs/architecture/decisions/004-postgres-driver.md (the forwarder
+  shape, the two pinned pitfalls, RecyclingMethod::Fast)
+- docs/architecture/decisions/008-contract-v1-pinning.md §6
+- docs/architecture/deployment.md (budgets, synchronous_commit)
+- docs/research/poc-pg-posture-findings.md (Sub-modules L/T; the
+  invocation note's isolation caveat)
+- alkstore-sqlite/src/store.rs, opts.rs (the structural twin's shape)
+
+## Notes
+
+> To be filled by implementation agent
+
+## Summary
+
+> To be filled on completion
\ No newline at end of file
diff --git a/tasks/pg-engine-queues.md b/tasks/pg-engine-queues.md
new file mode 100644
index 0000000..db5c739
--- /dev/null
+++ b/tasks/pg-engine-queues.md
@@ -0,0 +1,141 @@
+---
+id: pg-engine-queues
+name: Postgres engine — queues (`Queue`, `FOR UPDATE SKIP LOCKED` claims, job handles, backoff curve)
+status: pending
+depends_on: [pg-engine-seam-tx, pg-engine-notify-listen]
+scope: moderate
+risk: medium
+impact: component
+level: implementation
+tags: [wave-4, postgres-engine]
+---
+
+## Description
+
+Implement the queue mechanism's pg arm — the re-derived queue
+machinery (ADR-004/005's greenfield posture, pg-boss schema family as
+design reference; ADR-010's depth) over the schema task's job/dead
+tables and the forwarder's wake substrate:
+
+- **`Store::queue(name, opts)`** — validated constructor (shared-
+  namespace kinds; construction stays free, queues are names not
+  registered objects). The handle carries its `QueueOpts` as the
+  stamp source.
+- **`enqueue`** (auto-commit): resolves `EnqueueOpts` over the
+  handle's stamps via the seam task's `resolve_enqueue_opts` +
+  `stamps_with_override` (the pg engine's own `resolution.rs` —
+  one clock read per enqueue, delay-over-`run_at`, relative-
+  `expires`, `max_attempts` the one per-job override), `encode_payload`
+  at the seam, `INSERT` on a pool connection. Then **wake**:
+  `pg_notify` on the queue's wake channel (the mechanism-name-is-
+  the-channel realization) — the LISTEN-driven claim path's push
+  half.
+- **`claim_one` / `claim_batch`** — exactly-once handout under
+  concurrency: the `FOR UPDATE SKIP LOCKED` claim (POC-pinned under
+  4×4 concurrency) on a pool connection, in one statement or a
+  CTE-scoped transaction — implementer's choice, but the
+  exactly-once property is what the tests pin (two workers, one row;
+  N×M concurrent drain, zero double-claims). Semantics per ADR-010:
+  ordering `priority DESC, run_at ASC, enqueue order`; `attempts +=
+  1` per claim (a reclaim consumes an attempt); the claim sets the
+  row's deadline from the job's own stamped `visibility_timeout_s`;
+  the pre-claim dead-letter sweep for exhausted reclaimables (the
+  substrate's D-12 shape — a worker dying right after its last
+  allowed claim dead-letters, not strands); `worker_id` stamps the
+  claimant column (non-empty validated, consumer-local — ADR-019 §2).
+  The **extent guard** (`n <= 0` → empty `Vec` at trait-impl entry,
+  ADR-023 §2). Claims return boxed `JobHandle`s (ops + row value).
+- **`JobHandle` impl** (ADR-019 §3, ADR-010 §2's uniform validity
+  predicate — `processing` + unexpired claim deadline; refusal =
+  `Ok(false)`, never error):
+  - one-shot `ack`/`retry`/`fail` as `'static` boxed futures over the
+    consumed handle (mirroring `TxHandle::commit`'s shape);
+    repeatable `heartbeat(&self, extend)` — absolute reset of the
+    claim deadline (not additive), late heartbeat refused.
+  - `ack` deletes the row; `retry(err, None)` computes the queue's
+    **equal-jitter exponential curve** — the pg engine's own
+    implementation of ADR-010 §3's pinned form
+    (`base * 2^(attempt-1)`, jitter over `[half, cap]` integerized
+    inclusive, capped 3600 s; attempt index = the row's post-claim
+    count). **The jitter's randomness source is this task's flagged
+    choice**: the SQLite engine used a std-only `RandomState`-hash
+    draw (no `rand` dep, ADR-005's posture) — reuse that approach
+    unless a pg-native source is materially better, and **flag the
+    choice in Notes**. `retry(err, Some(d))` overrides; `fail(None)`
+    pins `"failed"`; `retry`'s exhaust string pins
+    `"max attempts exceeded"` (the engine-default strings land
+    identically on both engines — the backlog row's wording).
+  - Dead-letter moves (retry-at-budget, fail) write the dead table
+    with `last_error`/`died_at`; dead rows are `get_job`-visible.
+- **`ack_batch(ids)`** — the batch ack applies the uniform predicate
+  per id (lapsed or non-claimed ids silently not counted; count
+  returned; no worker filter — the wave-3 D-31 delta's shape is the
+  contract's, not a substrate artifact).
+- **`cancel(job_id)`** — unconditional delete (pending or processing,
+  not an interrupt); `Ok(false)` when row-less.
+- **`get_job(job_id)`** — dead-visible pure read (pool connection;
+  `Option<Job>` with `last_error`/`died_at` on dead rows; the
+  resolved values — not raw opts — are what the row shows, ADR-020).
+- **`sweep_expired()`** — both-states no-stranded-rows move (any
+  past-expiry row → dead with `last_error='expired'`) + retention
+  enforcement (`dead_letter_retention_s`, `None` = forever) —
+  ADR-010 §5; the queue-scoped handle form.
+- **LISTEN-driven claim with re-poll safety net** (the engine's
+  default consumption posture, ADR-004): the *contract* claim ops are
+  pull-shaped (callers claim in their own loops); the wake substrate
+  makes those loops reactive — a consumer that `listen(queue_name)`s
+  and claims on wake gets the measured 3–6 ms path; the re-poll
+  safety net (interval-tunable, `PgOpts`-carried if wired —
+  implementer's choice, documented either way) covers LISTEN loss
+  (the no-replay hole). The engine's own tests exercise the
+  wake-driven claim loop (commit → notify → wake → claim, the POC's
+  property) to pin the posture; the contract surface itself stays
+  pull-shaped.
+- **Closed-store posture** on all queue ops (the wave-3 shape).
+
+## Acceptance Criteria
+
+- [ ] `enqueue` resolves opts over the handle's stamps (the shared
+      resolution arithmetic; `max_attempts` the one override);
+      payload bytes exact; wake notify fires on the queue's channel
+- [ ] `claim_one`/`claim_batch`: exactly-once under concurrency
+      (two-workers-one-row; 6×40 concurrent drain, zero
+      double-claims); ordering row pinned; reclaim consumes an
+      attempt; deadline from the row's stamp; extent guard
+      (`n <= 0` → empty `Vec`); pre-claim exhausted-reclaimable sweep
+- [ ] `JobHandle`: validity predicate uniform (lapse refuses all ops,
+      `Ok(false)`); heartbeat absolute-reset + late-refused; ack
+      deletes; retry curve matches ADR-010 §3's range/cap
+      (statistical + cap tests, the SQLite task's shape); dead-letter
+      defaults (`"max attempts exceeded"`, `"failed"`, caller strings
+      verbatim); dead rows `get_job`-visible
+- [ ] `ack_batch` per-id predicate (mixed lapsed/valid ids); `cancel`
+      unconditional both states; `sweep_expired` both-states +
+      retention (incl. forever-none)
+- [ ] Wake-driven claim loop pinned end-to-end (enqueue → wake →
+      claim, cross-connection)
+- [ ] Validation (empty/reserved queue names, empty `worker_id`) +
+      closed-store fail-closed on all queue ops
+- [ ] `cargo test -p alkstore-postgres` (harness server), clippy
+      `-D warnings`, fmt clean; gates green server-less
+
+## References
+
+- docs/architecture/core-contract.md (queues section)
+- docs/architecture/decisions/010-queue-semantics-depth.md (§1–§5, the depth)
+- docs/architecture/decisions/019-mechanism-handle-surfaces.md §1–§3
+- docs/architecture/decisions/020-enqueue-opt-semantics-and-bridges.md
+- docs/architecture/decisions/023-fourth-review-round.md §2 (extent guard)
+- docs/architecture/decisions/004-postgres-driver.md (re-derive posture, LISTEN-driven default)
+- docs/architecture/queues.md (lifecycle, consumption posture)
+- docs/research/reference-pgboss-rs-semantics.md (design reference)
+- alkstore-sqlite/src/queue.rs (the structural twin — same semantics, own SQL)
+- alkstore/src/queue.rs, job_handle.rs (the traits)
+
+## Notes
+
+> To be filled by implementation agent
+
+## Summary
+
+> To be filled on completion
\ No newline at end of file
diff --git a/tasks/pg-engine-scheduler-outbox.md b/tasks/pg-engine-scheduler-outbox.md
new file mode 100644
index 0000000..e013bf2
--- /dev/null
+++ b/tasks/pg-engine-scheduler-outbox.md
@@ -0,0 +1,154 @@
+---
+id: pg-engine-scheduler-outbox
+name: Postgres engine — scheduler (`schedule`/`unschedule`/`run_schedules`) and outbox helper
+status: pending
+depends_on: [pg-engine-queues, pg-engine-locks]
+scope: broad
+risk: high
+impact: component
+level: implementation
+tags: [wave-4, postgres-engine]
+---
+
+## Description
+
+Implement the scheduler collapse's pg arm (ADR-009) and the outbox
+helper (ADR-014/019 §5) — the wave's last mechanism pair, riding the
+queues task's claim machinery and the locks task's lock machinery:
+
+**Scheduler** (over the schema task's schedule table):
+
+- **`schedule(name, spec, queue, payload, opts)`** — upsert by name;
+  entry-point validation on all three name-bearing arguments
+  (schedule name like a queue name; the **queue argument** identically
+  per ADR-021 §3 — a schedule cannot target a reserved name, so the
+  outbox backing queue stays reachable only through the outbox
+  surface); payload encoded at the seam (`Codec` propagation); the
+  spec grammar is `@every <n><unit>` only (`s|m|h|d`) — **the pg
+  engine owns its parser** (no shared substrate parser on this side;
+  the grammar is small and pinned; cron strings → `InvalidSpec`).
+  Rejection is pre-storage (`InvalidSpec` before any round trip).
+- **`unschedule(name)`** — the unregister; `Ok(false)` when absent.
+- **`run_schedules(stop: StopToken)`** — the leader loop, the
+  structural twin of the SQLite engine's runner with the pg deltas:
+  - **Leadership via the engine's own lock machinery** on the
+    reserved name `__alkstore_scheduler` (ADR-009 §1) — `try_lock`
+    with a per-runner-instance-unique owner token (the wave-3
+    finding's lesson: a shared constant would make two processes each
+    see "self" holding; the token is pid + subsec-nanos or
+    equivalent). TTL engine-internal (10 s), renewed at the top of
+    every loop iteration **and** on a cadence across the sleep (the
+    wave-3 posture: without in-sleep renewals, a boundary further out
+    than the TTL lapses its own lease by drift). A failed mid-sleep
+    renewal is not acted on at the sleep site — the loop's
+    top-of-iteration renew owns the loss decision.
+  - **Acquire-time refusal is `Err(LeadershipLost)`** — the same
+    matchable loss a mid-run loser returns (the wave-3 posture: no
+    runner-loop exists without the lock; the respawn recipe is
+    identical either way).
+  - **The tick**: fire enqueues + boundary advance commit together —
+    on pg this is one pool-checked-out transaction (`BEGIN` …
+    `COMMIT` on the held object, not `BEGIN IMMEDIATE` — pg's
+    default isolation gives the row-locked fire tx ADR-009 §4
+    names; the schedule row's `SELECT … FOR UPDATE` inside the tick
+    tx is the row lock that serializes rogue second tickers). The
+    tick body: boundary math over the registered specs (the
+    **one-firer-per-boundary** guarantee — at-least-once per elapsed
+    boundary, the **64-boundary catch-up cap with skip-forward**
+    past the cap, honker parity), each fire an ordinary enqueue
+    stamped with the `ScheduleOpts` resolved over the target queue's
+    *derived engine defaults* (no handle is open — ADR-020 §3), then
+    the soonest-boundary read for the sleep deadline.
+  - **Wake-driven tick advance** (the pg delta over the SQLite
+    runner's pure-sleep shape): the runner parks on the sleep
+    deadline, but an enqueue-wake on the scheduler's own wake channel
+    (or a plain bounded-wait interrupt — implementer's choice,
+    documented) may shorten the sleep so a newly-registered schedule
+    is noticed promptly; correctness never rides the wake (the
+    deadline always fires; the wake only shortens idle latency —
+    the SQLite runner's 60 s idle-tick posture is the acceptable
+    floor if the wake path proves fussy; flag the choice in Notes).
+  - **Clean stop** (`StopToken::cancel`) → `Ok(())`, the leadership
+    lock released cleanly; **loss** → `Err(LeadershipLost)` before
+    any tick.
+- **Boundary math is this engine's own implementation** (ADR-012 §2's
+  one-owner rule — the substrate's `scheduler_tick`/`scheduler_soonest`
+  are SQLite-side): the `@every` interval arithmetic, the boundary
+  advance, the 64-cap catch-up, and the skip-forward. The contract
+  suite (wave 5) pins the two engines' boundary behavior to identical
+  outcomes.
+
+**Outbox** (a helper over queues — no separate mechanism):
+
+- **`Store::outbox(name)`** — validated constructor (empty →
+  `InvalidName`, reserved-prefixed → `ReservedName`); carries no opts
+  of its own (the backing queue's derived 60/5/5 set is
+  engine-documented, ADR-010 §3a).
+- **`enqueue(payload, opts)`** — the auto-commit path into the
+  derived backing queue: derive `__alkstore_outbox:{name}` (the
+  reserved prefix makes it unreachable by `enqueue_tx`/`queue()` by
+  design), stamp the 60/5/5 set with the `EnqueueOpts::max_attempts`
+  override, full ADR-020 resolution, ride the queues task's enqueue
+  machinery.
+- **`run_once(worker_id, delivery) -> bool`** — the pull op: claim
+  one job from the backing queue via the ordinary claim machinery
+  (the queues task's), run the delivery callable (core's `Delivery`
+  trait object — ADR-019 §5's pinned shape), `ack` on `Ok`,
+  `retry(err, None)` (the queue's curve; the delivery error's
+  `Display` string as the retry's exhaust string — the wave-3
+  posture) on `Err`, `false` on empty. **No engine-issued heartbeat
+  inside delivery** (honker parity, inherited deliberately — the
+  boxed `JobHandle`'s own `heartbeat` remains available to the
+  closure); the boxed `JobHandle` handed to the caller.
+- **`outbox_enqueue_tx`** landed with the seam task; this task's
+  tests may extend its coverage (the commit-atomicity row is
+  suite-owned in wave 5).
+
+## Acceptance Criteria
+
+- [ ] `schedule`: validation triad (name/spec/queue), `@every`-only
+      grammar (cron → `InvalidSpec` pre-storage), upsert read-back;
+      the pg-owned parser pinned by tests (accepts/rejects the
+      grammar's exact surface)
+- [ ] `run_schedules`: leadership via `__alkstore_scheduler` with a
+      per-instance owner token; acquire-time refusal =
+      `Err(LeadershipLost)`; tick fires + boundary advance in one
+      transaction (row-locked — a rogue second ticker cannot
+      double-fire); loss returns `Err(LeadershipLost)` before any
+      tick; clean stop releases the lock; in-sleep lease renewals
+      hold across boundaries further out than the TTL
+- [ ] One-firer-per-boundary: two runners on the same db, exactly one
+      fire per boundary (the band test); 64-cap catch-up +
+      skip-forward with no in-window refire
+- [ ] Fired jobs stamp the derived defaults resolved over
+      `ScheduleOpts` (`get_job`-visible)
+- [ ] `outbox`: constructor validation; backing queue unreachable by
+      `enqueue_tx`/`queue()`; 60/5/5 stamps + `max_attempts`
+      override; `run_once` ok/err/false dispositions with the curve
+      gap asserted; in-delivery heartbeat renewal available;
+      exhaustion to dead with the delivery error string
+- [ ] `cargo test -p alkstore-postgres` (harness server), clippy
+      `-D warnings`, fmt clean; gates green server-less
+
+## References
+
+- docs/architecture/decisions/009-scheduler-collapse.md (§1–§6)
+- docs/architecture/decisions/014-outbox-tx-enqueue.md
+- docs/architecture/decisions/019-mechanism-handle-surfaces.md §4, §5
+- docs/architecture/decisions/020-enqueue-opt-semantics-and-bridges.md §3
+- docs/architecture/decisions/021-tx-reads-and-value-shape-fixes.md §3
+- docs/architecture/engine-postgres.md (Mapping the contract: scheduler/outbox)
+- docs/architecture/queues.md (scheduler section)
+- alkstore-sqlite/src/scheduler.rs, outbox.rs (the structural twin —
+  same semantics, own boundary math; its Notes carry the flake-fix
+  lessons: drive catch-up tests from far-out horizons so no live
+  boundary races the assertion window)
+- alkstore/src/stop_token.rs, outbox.rs (the traits)
+
+## Notes
+
+> To be filled by implementation agent
+
+## Summary
+
+> To be filled on completion
\ No newline at end of file
diff --git a/tasks/pg-engine-schema.md b/tasks/pg-engine-schema.md
new file mode 100644
index 0000000..c77925d
--- /dev/null
+++ b/tasks/pg-engine-schema.md
@@ -0,0 +1,110 @@
+---
+id: pg-engine-schema
+name: Postgres engine — schema bootstrap (engine-owned schema, table family)
+status: pending
+depends_on: []
+scope: narrow
+risk: low
+impact: component
+level: implementation
+tags: [wave-4, postgres-engine]
+---
+
+## Description
+
+Stand up the pg engine's data model: the engine-owned PostgreSQL schema
+and its table family — the dependency root every later wave-4 task
+rides. Per ADR-010 §8 (one engine-owned schema, default `alkstore`,
+queues as rows — no per-queue tables/schemas) and engine-postgres.md's
+mapping table.
+
+A `schema.rs` module in `alkstore-postgres`:
+
+- **Schema name is engine configuration** (default `alkstore`) — it
+  rides `PgOpts` (the open task wires it; this task's bootstrap takes
+  the schema name as an argument).
+- **Bootstrap is idempotent DDL**, runnable at open on any connection:
+  `CREATE SCHEMA IF NOT EXISTS` + `CREATE TABLE IF NOT EXISTS` +
+  indexes. All engine-owned objects live in the one schema; consumer
+  tables co-tenant the instance untouched (deployment.md's co-tenancy
+  posture).
+- **Table family** (column sets follow the contract's value shapes —
+  the `#[doc(hidden)]` constructors' full field lists in
+  `alkstore/src/{job,stream,schedule}.rs` are the field lists of
+  record; ADR-019 §1/§3, ADR-015 §3, ADR-009 §3, ADR-021 §2):
+  - **job table** (live rows — pending and processing): queue name,
+    payload, enqueue/ready-time resolution columns (`run_at`,
+    `expires_at`), `priority`, the stamped `QueueOpts` columns
+    (`visibility_timeout_s`, `max_attempts`, `backoff_base_s`,
+    `dead_letter_retention_s` — ADR-010 §3a), `attempts`,
+    `worker_id`, `claimed_at`, `claim_expires_at`, `last_error`.
+    Claim-ordering index per ADR-010's ordering row (`priority DESC,
+    run_at ASC, enqueue order`).
+  - **dead table**: the job row's diagnosis fields + `died_at`,
+    `last_error`; retention index for the sweep's retention
+    enforcement.
+  - **stream events table**: `id BIGSERIAL` (the offset — engine-
+    assigned, monotone, immutable, never renumbered), stream name,
+    nullable `key` column (carried metadata, ADR-015 §1), payload,
+    `created_at` (unix-seconds, informational). The `(stream, id)`
+    index serving the `offset ASC` read paths.
+  - **offsets table**: (stream, consumer) → offset, the monotone
+    checkpoint upsert target.
+  - **schedule table** (ADR-009): name (unique), `@every` spec, target
+    queue, payload, the `ScheduleOpts` stamps, and the boundary-state
+    columns the tick math advances (the substrate's
+    `__alkstore_scheduler_tasks` shape is the reference —
+    `reference-honker-machinery.md`).
+  - **locks table**: name (unique), owner, expiry timestamp — the TTL
+    lock machinery's storage (the substrate's lock-table semantics,
+    re-derived; the locks task builds ops on it).
+- **Payload columns store the exact encoded bytes** (`bytea`) — not
+  `JSONB`: JSONB normalization (key reordering, duplicate folding)
+  would break the contract suite's byte-identical-rows row (ADR-020
+  §4). The engine stores exactly what `encode_payload` returned.
+- **No notifications table** — pg notify is native (`pg_notify`); the
+  SQLite notify-table carriage has no pg counterpart. The 8000-byte
+  check is client-side (the seam task), not storage-shaped.
+- **No migration machinery in v1** — tables are created fresh per
+  schema; append-column migrations are the substrate's evolution
+  discipline, not a greenfield v1 need (note the posture in the module
+  docs; wave 6's release pass revisits if the deployment story
+  demands in-place upgrades).
+
+Tests use a raw `tokio_postgres::Connection`/Client against the
+harness server (no engine code yet): bootstrap runs twice cleanly
+(idempotence), every table exists with its pinned columns, the
+bigserial offset behaves (monotone, gap-on-delete), the claim-ordering
+index exists.
+
+## Acceptance Criteria
+
+- [ ] `schema.rs` with an idempotent `bootstrap(conn, schema_name)`
+      creating the full table family + indexes in the one schema
+- [ ] Column sets match the contract's value shapes (the constructors'
+      field lists; stamp columns per ADR-010 §3a; `claimed_at` per
+      ADR-021 §2)
+- [ ] Payload columns are `bytea` (exact-bytes posture, JSONB
+      rejected — documented in the module docs with the reason)
+- [ ] Schema name is a parameter (default `alkstore` decided at the
+      `PgOpts` layer, next task)
+- [ ] Idempotence + table-shape tests green against the harness server
+- [ ] `cargo test -p alkstore-postgres`, clippy `-D warnings`, fmt clean
+
+## References
+
+- docs/architecture/engine-postgres.md (Identity and posture; mapping table)
+- docs/architecture/decisions/010-queue-semantics-depth.md §3a, §8
+- docs/architecture/decisions/015-streams-depth.md §1, §3
+- docs/architecture/decisions/009-scheduler-collapse.md §3
+- docs/architecture/deployment.md (co-tenancy)
+- docs/research/reference-honker-machinery.md (schema reference read)
+- alkstore/src/job.rs, stream.rs, schedule.rs (the field lists of record)
+
+## Notes
+
+> To be filled by implementation agent
+
+## Summary
+
+> To be filled on completion
\ No newline at end of file
diff --git a/tasks/pg-engine-seam-tx.md b/tasks/pg-engine-seam-tx.md
new file mode 100644
index 0000000..ef1aa33
--- /dev/null
+++ b/tasks/pg-engine-seam-tx.md
@@ -0,0 +1,129 @@
+---
+id: pg-engine-seam-tx
+name: Postgres engine — `begin_tx`, `TxHandle` (pooled-object handle), all eleven `*_tx` methods
+status: pending
+depends_on: [pg-engine-open-opts]
+scope: moderate
+risk: high
+impact: component
+level: implementation
+tags: [wave-4, postgres-engine, seam]
+---
+
+## Description
+
+Implement the transactional seam — the pg arm of ADR-007, the shape
+every other mechanism task rides, and the wave's highest-risk piece
+(the wave-3 review's writer-slot findings show exactly the defect
+class seam error arms breed):
+
+- **No bridge, no `spawn_blocking`** — the pg handle holds the pooled
+  object directly (tokio-postgres `Client` is `Send + Sync`, POC #2
+  compile-probe; ADR-007's pg arm). `begin_tx`: pool checkout +
+  `BEGIN`, return the boxed `PgTxHandle` holding
+  `Option<deadpool::Object>`; ops are straight `.await`s; commit/
+  rollback return the object to the pool.
+- **Drop = rollback** (ADR-021 §4): the handle's `Drop` issues
+  `ROLLBACK` and re-pools the object — panic-safe (no panic escapes
+  drop). A failed rollback's disposition is pinned (the object is
+  discarded, not re-pooled — its state is unknowable; the pool
+  accounting stays exact, deployment.md's ADR-021 note). Ops after
+  commit/consumption fail closed with `Error::Closed` (the wave-3
+  fail-closed guard shape).
+- **All eleven `*_tx` methods, for real** (nothing stubs — the schema
+  exists, the SQL is this task's to write):
+  - Writes: `enqueue_tx` (stamps the derived plain-queue defaults
+    300/3/5/none — the no-handle-open shape, ADR-020 §3a),
+    `publish_tx`, `publish_with_key_tx` (empty-`Some`-key →
+    `InvalidName`, ADR-015 §2), `notify_tx` (`SELECT pg_notify($1,
+    $2)` inside the caller's tx — delivers at commit, native
+    commit-atomicity, ADR-007; the 8000-byte client-side check fires
+    here and on the auto-commit twin — typed `PayloadTooLarge` with
+    the limit in the variant, before any round trip, ADR-016 §5),
+    `save_offset_tx` (monotone upsert), `outbox_enqueue_tx` (derives
+    the `__alkstore_outbox:{name}` backing queue name engine-side,
+    stamps the outbox's 60/5/5 set — ADR-014 §1).
+  - Reads (ADR-021 §1, read-your-own-writes through the same held
+    tx): `get_job_tx` (queue-scoped — a found row's queue column
+    must match the argument, the wave-3 seam's scoping rule),
+    `get_offset_tx`, `read_since_tx`, `read_from_consumer_tx`
+    (offset ASC; the extent guard `limit <= 0` → empty `Vec` at
+    trait-impl entry, ADR-023 §2).
+  - `commit(self: Box<Self>)`: `COMMIT` + re-pool. A failed COMMIT
+    discards the object (unknowable state) — the wave-3 review's
+    finding (c), pre-empted here.
+- **Name validation at every entry point** via core's `validation`
+  helpers (shared-namespace kinds vs consumer-local; ADR-008 §4) —
+  before any round trip.
+- **Payload encoding**: core's fallible `encode_payload` at the seam;
+  `?` the `Codec` error (ADR-023 §1). Decode via the `#[doc(hidden)]`
+  `from_row` constructors; malformed rows → `Codec` (the posture
+  guard).
+- **The engine-side arithmetic lives here first** (one owner per
+  formula, ADR-012 §2 — the pg engine's own `resolution.rs`, the
+  SQLite one is reference not dependency): `now` (the single clock
+  read — `now()` at the SQL layer or one `std::time` read; pick one,
+  document it), `plain_queue_default_stamps()` (300/3/5/none),
+  `outbox_default_stamps()` (60/5/5), `outbox_backing_queue_name()`
+  (the reserved-prefix derivation), `resolve_enqueue_opts`
+  (delay-over-`run_at`, relative-`expires`, one clock read per
+  enqueue). The auto-commit mechanism tasks reuse these directly —
+  same shape as the SQLite engine's `resolution.rs`.
+- **`pg_notify` payload carriage**: the notify payload is the
+  serde_json serialization's UTF-8 text (≤ 8000 bytes checked
+  client-side). Large payloads do NOT ride a side-table in v1 — the
+  typed rejection is the contract's answer (ADR-016 §5's occurrence
+  asymmetry; the outbox shape is the documented large-payload
+  pattern, not engine machinery).
+
+The mechanism tasks then add what the tx path deliberately lacks: the
+auto-commit handles, the forwarder-driven receivers, the backoff
+curve, and the scheduler runner.
+
+## Acceptance Criteria
+
+- [ ] `begin_tx` checks out + `BEGIN`; concurrent `begin_tx` draws
+      distinct pooled objects (pool-bounded); pool exhaustion blocks
+      (deadpool's acquire semantics) — tested
+- [ ] All eleven `*_tx` methods implemented with entry-point
+      validation; writes commit-atomic, reads see the tx's own writes
+      (read committed, both directions POC-pinned); opts resolution
+      (delay-over-run_at, relative-expires, 300/3/5/none and 60/5/5
+      stamp sets) pinned by tests
+- [ ] `commit` commits + re-pools; drop (without commit) rolls back +
+      re-pools — no-ghosts tested for job rows, stream events,
+      notifications (a listener hears nothing from a rolled-back
+      `notify_tx`), and offset saves
+- [ ] Drop is panic-safe; failed COMMIT/ROLLBACK arms discard the
+      object (pool accounting exact); ops-after-consume → `Closed`
+- [ ] `notify_tx`/`notify`'s 8000-byte check: typed
+      `PayloadTooLarge { limit: 8000 }` client-side, before any round
+      trip; ≤ 8000 crosses
+- [ ] `with_tx` (core's provided method) works end-to-end on this
+      engine: Ok ⇒ commit, Err ⇒ rollback — tested
+- [ ] `encode_payload`'s `Codec` error propagates; extent guard on the
+      tx stream reads; `get_job_tx` queue-scoped
+- [ ] `cargo test -p alkstore-postgres` (harness server), clippy
+      `-D warnings`, fmt clean; gates green server-less
+
+## References
+
+- docs/architecture/decisions/007-transactional-seam.md (pg arm)
+- docs/architecture/decisions/021-tx-reads-and-value-shape-fixes.md §1, §2, §4
+- docs/architecture/decisions/014-outbox-tx-enqueue.md §1
+- docs/architecture/decisions/020-enqueue-opt-semantics-and-bridges.md §1–§4
+- docs/architecture/decisions/016-deployment-honesty.md §5
+- docs/architecture/decisions/023-fourth-review-round.md §1, §2
+- docs/architecture/engine-postgres.md (Mapping the contract: begin_tx, handle ops)
+- docs/research/poc-pg-posture-findings.md (Sub-module T — the shape
+  (a) evidence, read-your-writes, notify-at-commit)
+- alkstore/src/tx.rs (the trait to implement)
+- alkstore-sqlite/src/tx.rs, resolution.rs (the structural twin)
+
+## Notes
+
+> To be filled by implementation agent
+
+## Summary
+
+> To be filled on completion
\ No newline at end of file
diff --git a/tasks/pg-engine-streams.md b/tasks/pg-engine-streams.md
new file mode 100644
index 0000000..967a170
--- /dev/null
+++ b/tasks/pg-engine-streams.md
@@ -0,0 +1,126 @@
+---
+id: pg-engine-streams
+name: Postgres engine — streams (`StreamHandle`, reads, subscribe, trim)
+status: pending
+depends_on: [pg-engine-seam-tx, pg-engine-notify-listen]
+scope: moderate
+risk: medium
+impact: component
+level: implementation
+tags: [wave-4, postgres-engine]
+---
+
+## Description
+
+Implement the stream mechanism's pg arm — the full `StreamHandle` trait
+and the durable `subscribe` receiver, over the schema task's events/
+offsets tables and the forwarder's wake substrate (ADR-015's depth,
+engine-postgres.md's streams mapping row):
+
+- **`publish` / `publish_with_key`** (auto-commit): pool checkout,
+  `INSERT` into the events table (bigserial id = the assigned offset;
+  nullable key column — carried metadata, empty-`Some`-key →
+  `InvalidName` per the tx path's rule), `encode_payload` at the seam
+  (`Codec` propagation), return to pool. Then **wake**: `pg_notify` on
+  the stream's wake channel (the mechanism-name-is-the-channel
+  realization pinned in the plan's Wave 4 section) — a separate
+  statement on the pooled connection (auto-commit: its own implicit
+  tx; the wake is best-effort, not commit-atomic with the insert —
+  the no-replay hole covers a gap between them; the durable row is
+  the truth).
+- **`read_since` / `read_from_consumer`**: pool reads, `ORDER BY id
+  ASC` (global FIFO — bigserial offsets, immutable, never renumbered;
+  ADR-015 §4's ordering row), the **extent guard** (`limit <= 0` →
+  empty `Vec` at trait-impl entry, ADR-023 §2 — never `LIMIT -1`-class
+  dialect semantics; pg's `LIMIT` with a negative is an error, which
+  the guard makes unreachable). `read_from_consumer` anchors at the
+  consumer's stored offset.
+- **`save_offset` / `get_offset`**: the monotone upsert (a save below
+  the stored checkpoint is a silent no-op — `GREATEST`-guarded upsert
+  or `WHERE offset > stored`; one owner of the monotonicity rule) and
+  the checkpoint read. Both save forms' `Err` arm carries `Database`
+  only (ADR-021 §5).
+- **`trim_to(horizon)`**: `DELETE … WHERE id <= horizon` on a pool
+  connection (wakes nothing — ADR-015 §5's posture; the boundary
+  argument is total: a negative horizon deletes nothing,
+  idempotently), returning the deleted count. Surviving events keep
+  their offsets (gaps legal).
+- **`subscribe(consumer)`** — the durable receiver, the structural
+  twin of the SQLite engine's bridge with the pg deltas:
+  - Attach-before-read ordering (subscribe to the stream's wake
+    channel *before* the initial cursor read — a commit cannot fall
+    between attach and wait).
+  - Initial cursor read of the consumer's *stored* offset; page-drain
+    to the tail (`offset ASC`); park on the wake feed (the forwarder
+    fanout for this stream's channel); re-drain on wake
+    (wake-driven with table re-read — events never require polling
+    to become visible, core-contract.md's streams section).
+  - **Close semantics (the pg arm, ADR-021 §5)**: the receiver stays
+    open across forwarder reconnects — reconnect-wakes arrive on the
+    wake feed and simply trigger re-drains (which is the recovery:
+    the no-replay hole's gap is healed by the re-drain reading the
+    durable rows); closes only at engine shutdown (terminal `None`).
+  - **Position semantics**: `position` = the last-*yielded* event's
+    offset (0 before any yield ⇒ `save_offset()` before any yield is
+    a no-op, ADR-021 §5); the bridge's internal cursor advances on
+    drained-from-storage events — a stalled consumer's
+    `save_offset()` never checkpoints unobserved events. Backpressure:
+    bounded channel + backpressure-aware send.
+  - `recv()`'s `Err` carries `Database` only (ADR-021 §5 — remap
+    non-`Database` read errors with the original as source, the
+    SQLite `database_only` helper's shape); `try_recv` arms mirror
+    the wake receiver (idle `Ok(None)`, `Err(Closed)` after
+    shutdown-close).
+  - The receiver's `read_since` (anchored, extent-guarded) and sync
+    `save_offset` (the last-yielded offset through the same monotone
+    op) per ADR-019 §6.
+- **Validation**: stream names via `validate_shared_name` at
+  `store.stream()` and (defensively) the handle's reads; consumers
+  via `validate_local_name` (non-empty only — reserved prefixes legal
+  for consumers).
+- **Closed-store posture**: every handle op checks the store's closed
+  flag first and fails closed with `Database` (the wave-3 shape).
+
+## Acceptance Criteria
+
+- [ ] `publish`/`publish_with_key` return the assigned bigserial
+      offset; key round-trips (`None` vs `Some`); `encode_payload`'s
+      `Codec` propagates; empty-`Some`-key rejected `InvalidName`
+- [ ] Reads yield `offset ASC` (global FIFO); extent guard
+      (`limit <= 0` → empty `Vec`) at trait-impl entry; offsets
+      immutable (trim leaves gaps, never renumbers — tested)
+- [ ] `save_offset` monotone (regression save = silent no-op);
+      `get_offset` reads the checkpoint; save `Err` = `Database` only
+- [ ] `trim_to` exact-boundary (`<=`), returns deleted count, wakes
+      nothing, negative horizon deletes nothing; reads resume at the
+      horizon's first remaining row; saved offsets below the horizon
+      stay valid
+- [ ] `subscribe`: replay from stored offset, wake-driven delivery,
+      position = last-yielded offset, backpressure holds, receiver
+      stays open across forwarder reconnects (reconnect-wake triggers
+      re-drain — the gap-healing property tested), terminal close at
+      engine shutdown only
+- [ ] Restart durability: publish → close store → reopen →
+      `subscribe` resumes from the saved offset
+- [ ] Validation + closed-store fail-closed on all entry points
+- [ ] `cargo test -p alkstore-postgres` (harness server), clippy
+      `-D warnings`, fmt clean; gates green server-less
+
+## References
+
+- docs/architecture/core-contract.md (streams section)
+- docs/architecture/decisions/015-streams-depth.md
+- docs/architecture/decisions/019-mechanism-handle-surfaces.md §6
+- docs/architecture/decisions/021-tx-reads-and-value-shape-fixes.md §5
+- docs/architecture/decisions/023-fourth-review-round.md §2 (extent guard)
+- docs/architecture/engine-postgres.md (Mapping the contract: streams)
+- alkstore-sqlite/src/stream.rs (the structural twin)
+- alkstore/src/stream_handle.rs, event_receiver.rs (the traits)
+
+## Notes
+
+> To be filled by implementation agent
+
+## Summary
+
+> To be filled on completion
\ No newline at end of file
diff --git a/tasks/review-wave-4.md b/tasks/review-wave-4.md
new file mode 100644
index 0000000..80d8848
--- /dev/null
+++ b/tasks/review-wave-4.md
@@ -0,0 +1,94 @@
+---
+id: review-wave-4
+name: Review gate — wave 4 (Postgres engine)
+status: pending
+depends_on: [pg-engine-integration]
+scope: moderate
+risk: low
+impact: phase
+level: review
+tags: [wave-4, review]
+---
+
+## Description
+
+Review the Postgres engine before wave 5 decomposes. Primary lens per
+the plan's review gates: **engine-vs-contract conformance** — the same
+discipline as the wave-3 gate, now against the second engine, with
+the pg-specific machinery added. Wave 5 (the contract suite) pins
+*cross-engine equivalence*; this gate checks the pg implementation
+against the pinned contract text line by line.
+
+Check:
+
+- **Contract conformance**: every trait method's behavior against
+  core-contract.md's per-mechanism text — validation at entry points
+  (auto-commit AND tx paths), error arms (`Database`-only receivers,
+  `Codec` at payload seams, `LeadershipLost` vs clean stop), value
+  shapes (`Job`/`StreamEvent` field-for-field), the no-work-is-a-value
+  vocabulary (`None`/`false`/empty-`Vec` where pinned, never errors).
+- **ADR-023 follow-through**: extent guards at trait-impl entry
+  (claims + all four stream reads); duration guards on both lock call
+  sites; `encode_payload` consumed fallibly at every call site.
+- **ADR-016 §5**: the 8000-byte check is client-side, typed, before
+  any round trip, on both `notify` and `notify_tx`; the limit constant
+  has one owner.
+- **Forwarder integrity** (the pg engine's own high-risk machinery —
+  the analogue of the wave-3 gate's seam-integrity lens): the two
+  POC-pinned deadlock pitfalls structurally excluded (poll loop
+  before first query; `Client` alive for the listener's lifetime);
+  reconnect backoff + re-LISTEN from the live channel set; the
+  synthetic reconnect-wake on the reserved string broadcast to every
+  subscriber; lag surfaced, not silent; receiver close semantics
+  exactly the pg arm (open across reconnects, close at shutdown
+  only, terminal, never reopens).
+- **Seam integrity**: drop = rollback under panic; failed
+  COMMIT/ROLLBACK arms discard (not re-pool) the object — pool
+  accounting exact under error paths (deployment.md's ADR-021 note);
+  ops-after-consume fail closed; no `spawn_blocking` anywhere (the
+  natively-async posture — a stray bridge would be a finding).
+- **Backoff curve + boundary math**: the engine-side formulas match
+  ADR-010 §3's pinned range/cap and ADR-009 §4's boundary guarantee
+  exactly (wave 5 pins cross-engine equivalence; this gate checks
+  this implementation against the text).
+- **Schema posture**: one engine-owned schema; payload columns
+  `bytea` (exact bytes — no JSONB normalization); no per-queue
+  tables; bootstrap idempotent.
+- **Backlog column**: the engine's suite rows present, stamped, green
+  against the pg factory; the `PayloadTooLarge` pg arm row present.
+- **Family standards**: no panics, no unwrap/expect outside tests, no
+  inline comments (doc comments fine), module-per-file.
+- **Gates**: full workspace build/test/clippy/fmt; the pg suite green
+  against the harness server.
+
+## Acceptance Criteria
+
+- [ ] Contract conformance spot-checked against the spec text (not
+      just the tests)
+- [ ] ADR-023's numeric-domain + typed-encoding follow-through
+      verified at their sites
+- [ ] Forwarder + seam integrity verified (code-read, not
+      test-name-trust)
+- [ ] Backoff curve + scheduler boundary math match the pinned ADR
+      text
+- [ ] Backlog column present, stamped, green against the pg factory
+- [ ] All gates green (incl. against the harness server)
+- [ ] Findings recorded; wave 5 decomposition may proceed
+
+## References
+
+- docs/plans/implementation.md (Review gates)
+- docs/architecture/core-contract.md
+- docs/architecture/engine-postgres.md
+- docs/architecture/decisions/004-postgres-driver.md
+- docs/architecture/decisions/016-deployment-honesty.md §5
+- docs/architecture/decisions/023-fourth-review-round.md
+- tasks/review-wave-3.md (the gate discipline's precedent)
+
+## Notes
+
+> To be filled by implementation agent
+
+## Summary
+
+> To be filled on completion
\ No newline at end of file