---
id: review-wave-6
name: Review gate — wave 6 (mem engine)
status: completed
depends_on: [mem-engine-integration, pg-fix-scheduler-tick-seam]
scope: moderate
risk: low
impact: phase
level: review
tags: [wave-6, review]
---

## Description

The wave-6 review gate (the plan's §Review gates, wave 6; the
born-pinned gate). Code-read the `alkstore-mem` engine against the
contract and the pinned ADR-024 text, and audit:

1. **Conformance read**: engine-vs-contract line by line, mirroring
   the wave-3/4 gate pattern — every trait surface against
   core-contract.md and the pinning ADRs (007/008/009/010/015/019/
   020/021/023), the tx overlay against ADR-024 + engine-mem.md §The
   tx overlay (commit-merge atomicity, wake-at-commit, drop = discard,
   panic-through-unwind).
2. **The third column as the compatibility instrument born pinned**:
   the mem column's wiring read row-by-row against the discharge map
   (which rows carry, which engine-shaped rows omit and why — the
   asymmetry classification, ADR-024 §3); `PayloadTooLarge`
   non-occurrence arm present; three-way arithmetic equivalence
   actually asserted green (backoff curve, opts resolution — the
   ADR-012 §2 one-owner rule's third owner), plus the derived-queue
   naming row composing identically.
3. **wasm32 acceptance audit** (ADR-024 §5): the compile gate clean,
   tokio feature subset verified on the manifest, no blocking calls
   in mem internals, the current-thread flavor actually enforced by
   the column's tests.
4. **engine-mem.md read against the implemented engine**; the
   deployment matrix's mem rows (`deployment.md`, landed with ADR-024)
   still true as implemented (honesty carriers: identity + docs +
   matrix — nothing runtime invented). Gate may flip `engine-mem.md`
   to `stable` if the read is clean.
5. **The rider closed**: `pg-fix-scheduler-tick-seam`'s arms verified
   (review 003 Finding 1 discharged into this window; Findings 2–3
   confirmed still parked for wave 7).
6. **Contract-window discipline**: the class-1 amend-in-place window
   is still open — any wobble the third implementation surfaced must
   appear as resolved ADR text (or be recorded as a finding with its
   disposition), not as a silent divergence between suite row and
   engine behavior.

## Acceptance Criteria

- [x] Full conformance code-read recorded (findings + dispositions
      inline in this task's Notes, mirroring review-wave-3/4 style)
- [x] Suite mem column green on host, 25-row set wired per the
      classification; three-way equivalence rows green
- [x] wasm32 compile gate + feature-subset + current-thread flavor
      verified, recorded
- [x] The rider's arms verified; Finding 1 disposition recorded
- [x] Workspace gates re-run green this session (`cargo build`,
      `cargo test`, clippy `-D warnings`, fmt)
- [x] `engine-mem.md` disposition decided (stable flip or follow-up)
      and `implementation.md`'s wave-6 row updated

## References

- docs/plans/implementation.md §Wave 6, §Review gates (wave 6 entry)
- docs/architecture/decisions/024-mem-engine.md (the charter this
  gate audits code against)
- docs/architecture/engine-mem.md
- tasks/review-wave-3.md / tasks/review-wave-4.md (the gate pattern)
- docs/reviews/003-wave-5-general-review.md (Finding 1's disposition
  trail; Findings 2–3's pre-release-pass park — wave 8 since the
  fuzzing wave's renumber, 2026-10-10)

## Notes

Gate verdict: **passed**. No blocker, no major live defect; three
edge-domain fixes landed inline (below); the remaining observations
are recorded with dispositions, none mem-divergent from the engines.

### Fixed inline at the gate (this session, verified clippy/fmt/test green after each)

1. **Stream cursor read overflow** (`stream.rs` `read_events`): the
   cursor advanced by plain `from + 1`, panicking on debug build for a
   caller-supplied `offset == i64::MAX` (direct `read_since`, or
   indirect via `save_offset(i64::MAX)` + `read_from_consumer`) and
   wrapping to a from-the-beginning read in release. Fixed with
   `Bound::Excluded(from)` — boundary-total at every i64, the
   ADR-023 §2 read-cursor classification this fix encodes.
2. **`trim_to` horizon endpoint**: `horizon.saturating_add(1)` clipped
   `i64::MAX` and left a row at offset `i64::MAX` undeleted — the
   exact-`<=` boundary broke at the one unreachable endpoint anyway.
   Fixed with `Bound::Included(horizon)`; unreachable in practice
   (2⁶³ publishes), fixed because the template cost nothing.
3. **Scheduler register overflow alignment** (conformance divergence
   found by the read): mem registered a *parseable* interval whose
   first boundary overflows (`@every 9223372036854775807s`) with a
   saturating `next_fire_at` — accepted-never-fires — where the SQLite
   register path rejects the same spec with `InvalidSpec`.
   Class-1-window divergence between engines on a register-time
   semantics; aligned mem to reject (`checked_add` → `InvalidSpec`, no
   row stored) with a new pin test distinguishing it from the
   unparsable-too-large spec the parser already refuses. The tick's
   `i64::MAX` overflow sentinel stays identical in both engines.
4. **Scaffold-residue cleanup**: `clock.rs`'s stale
   `#![allow(dead_code)]` on the now-wired accessor + scaffolding
   comment removed; `tx.rs`'s commit-wake comment corrected to what
   the code does (one wake per staged *stream* — receivers drain by
   cursor, the feed is plumbing); `wakes.rs` module doc's "mechanisms
   arrive with their tasks" updated to the wired mechanisms;
   `stream.rs` module doc no longer claims "a fresh stream's first
   event does not carry offset 1" for mem (mem's fresh count *is* 1;
   the ADR-015 §4 note is engine-internal and never compared
   cross-engine — the doc now says exactly that).

### Recorded findings, no code change (dispositions)

- **Multi-thread publish→wake ordering tolerance** (streams): the
  publish path allocates the offset, then inserts under the guard,
  then fires the wake after the guard drops — on a multi-thread
  runtime two concurrent publishes can complete insert+wake out of
  offset-allocation order, and a subscriber's cursor can skip an
  event whose wake fires later (the event is never lost from the log,
  but a pre-insert drain's checkpoint passes it). The SQL engines
  share the shape in principle (offset assigned pre-commit, wake
  post-commit — the "pg-shaped story" the mem docs cite), and nothing
  contract-side pins wake-fire order vs offset order. Disposition:
  this is a   contract-level tolerance nobody pinned, not a mem-local
  divergence — recorded for the pre-release pass (wave 7 at writing —
  wave 8 since the fuzzing wave's renumber) to either
  document as tolerated or amend as ADR text (wake order per stream =
  offset order). Mem's declared posture is current-thread-clean
  (engine-mem.md) and the suite's mem column proves it under the
  current-thread flavor, which is why nothing here fails today.
- **Mem arithmetic is a transplant, not an independent derivation**
  (one-owner-rule note): `resolution.rs`'s backoff curve and opts
  resolution, `scheduler.rs`'s `@every` parser and `skip_forward`, are
  byte-level copies of the SQLite-side implementations (semantics
  re-verified against the ADR text during this read). ADR-012 §2's
  letter — each engine owns its implementation, equivalence
  suite-pinned — is satisfied (mem owns its modules; the
  `backoff_curve_equivalence` and `enqueue_opts_resolution` rows run
  green on all three columns). Recorded so the gate's record doesn't
  overstate the independent-signal value of the third copy; the suite
  *is* the honesty instrument.
- **`heartbeat(extend)` unclassified** (ADR-023 §2): the domain table
  classifies `try_lock`/`renew` durations but not the queue
  heartbeat's seconds-shaped `extend` argument. Both engines treat it
  boundary-total (`saturating_add`/SQL integer add — a negative
  shortens/voids the claim). Recorded as a class-1-window doc
  annotation candidate for wave 7's pre-release pass; no engine
  diverges.
- **`save_offset`'s argument unclassified**: negatives accepted (stored,
  reads resume forward), monotone no-op below the checkpoint — both
  engines, contract pins monotonicity only (ADR-019 §6). Same
  annotation candidate.
- **`EventReceiver::try_recv`'s closed shape unpinned**:
  `Err(Closed)` (mirrors `WakeReceiver::try_recv`, ADR-008 §3);
  taxonomy-consistent, mem-test-pinned, but ADR-021 §5 pins only
  `recv()`'s arms. Same annotation candidate.
- **`get_job` is id-global, not queue-filtered** — uniform on all
  three engines (the handle carries the queue, the lookup the id;
  `get_job_tx` *is* queue-filtered on both engines with tx). Recorded
  so the record doesn't misread core-contract's "queue-scoped" handle
  phrasing as a filter; the ambiguity is contract-prose, not code.
- **Mem outbox `run_once` ignores a trailing ack/retry refusal**
  (deadline lapsed mid-delivery): identical to SQLite; the uniform
  validity predicate makes the refusal the at-least-once honesty —
  conformant, noted for the boolean's silent discard.
- **Wake-bus attach-vs-close race (defense-in-depth note)**:
  `WakeBus::attach`/`fire_commit_wakes` check `closed` before taking
  the registry guard, so a listen racing the engine's drop could in
  principle mint a feed whose sender post-dates the registry clear.
  Unreachable through safe APIs (a live `&MemStore` borrow pins the
  last reference — borrowck; the bus is reachable only through the
  store), so no change; recorded as a hardening candidate if the bus
  ever gains a second entry path.

### Audit record (verified items)

- **Conformance read**: queue state machine (claim predicate/ordering,
  visibility, reclaim-eats-attempt, uniform validity predicate,
  heartbeat, retry/fail/cancel/ack_batch, dead-letter + no-stranded
  sweep, queue-scoped get_job incl. dead rows); scheduler (grammar
  parser, register-time stamp resolution, one-guard fire-and-advance
  tick in the schedules→queues order, 64-cap catch-up + grid-aligned
  skip-forward, leadership acquire/renew/loss/clean-stop on the
  reserved lock with runner-unique owner); outbox (derived reserved
  backing queue unreachable by every consumer entry point, 60/5/5
  stamps, run_once ack/retry curve, tx half stamping identically);
  streams (offset-ASC reads, absent-consumer 0 head read, monotone
  saves, trim exact-`<=` never-renumbering and waking nothing,
  attach-before-read subscribe, engine-drop close arms); locks
  (contention loser `Ok(None)`, `ttl <= 0` refused on try + renew,
  silent lapse, non-owner release = false, renew-past-expiry refused,
  no `lock_tx`); clock (one accessor, all stamps read it, freeze
  seam cfg(test) only); validation (every entry routed through core's
  `validate_shared_name`/`validate_local_name`, no engine-local
  rules); tx overlay (eleven `*_tx` staged with entry validation
  before effect, one clock read, ADR-020 §3 stamps incl. per-job
  override, merged reads queue-scoped + dead-visible + offset-ASC +
  extent-clamped, one-op commit merge + wake-at-commit, drop =
  discard through unwind, stage-time id/offset allocation with legal
  gaps); wake substrate (channel-scoped, per-notify never coalesced,
  attach-after-commit no-replay, close disconnects all, dropped-feed
  detach, commit-ordered FIFO). Guard discipline audited: canonical
  order (schedules → queues → streams) holds at every two-guard site;
  no guard across an `.await` or user code; zero panics/unwrap/
  expect in library code.
- **Third column as compatibility instrument**: the mem target wires
  exactly the 24 engine-shared rows + mem's
  `payload_too_large_never_produced_on_mem` (stamped ADR-016 §5 +
  ADR-024 §3) + the factory-shape test; the reconnection-shaped rows
  stay per-engine wired (absent here, per ADR-024 §3's
  classification); the three-way equivalence (`backoff_curve_equivalence`,
  `enqueue_opts_resolution`) green on all three columns; 25/25 green
  on the host.
- **wasm32 acceptance**: `cargo check --target wasm32-unknown-unknown
  -p alkstore-mem` clean; manifest tokio subset sync+time (dev-deps
  rt/macros only); no `spawn_blocking`/blocking-call in mem source;
  current-thread flavor enforced on every mem-column test (a blocking
  call would hang, not pass); on-target suite execution stays scoped
  out behind the named adapter-crate collapse condition (ADR-024 §5 —
  recorded, not an OQ).
- **Spec + matrix**: `engine-mem.md` read against the implemented
  engine (tx overlay, asymmetry classification, isolation, wasm
  identity — all true as landed; the stable flip records the gate in
  its frontmatter); `deployment.md`'s mem rows read true as
  implemented.
- **Rider closed**: `pg-fix-scheduler-tick-seam`'s two arms
  (`one_faulted_tick_is_retried_and_the_job_fires`,
  `tick_retry_exhaustion_exits_typed_and_leaves_the_lock_to_lapse`)
  verified present against the real retry loop; review 003 Finding 1
  discharged; Findings 2–3 confirmed still parked for wave 7
  (implementation.md §Wave 7); the fire-wake gap
  (`pg-fix-scheduler-fire-wake`) closed pre-wave and its retired gap
  records verified; the wave-5 flake ledger's §6 update recorded.
- **Workspace gates re-run green this session**: `cargo build`; 
  `cargo test --workspace` green (core 25, suite harness 3, mem 101
  lib + 25 suite, SQLite 191 + 25, pg 124 + 25 + 9, pg column
  server-less per the SKIP posture — the pg-side verification was the
  rider tasks' own live runs); clippy `--all-targets -D warnings`
  clean; `cargo fmt --check` clean; `taskgraph validate` green
  (61 tasks).

## Summary

The wave-6 review gate passed and flipped `engine-mem.md` to
`stable`. The conformance read was clean in mechanism substance; the
findings class was edge-domain and hygiene: three inline fixes (the
stream cursor's `i64::MAX` overflow, `trim_to`'s horizon endpoint,
and the scheduler register-overflow alignment with SQLite's
`InvalidSpec` — the one true engine-vs-engine divergence the read
found, resolved into the class-1 window) and the expected residue
cleanups. The suite's third column is born pinned as ADR-024
required — 25 green under the current-thread flavor, the three-way
arithmetic equivalence actually asserted — and the wasm32 acceptance
items verified. The unpinned contract-adjacent observations (the
multi-thread wake-ordering tolerance, three ADR-023 §2 table
classifications, the receiver try_recv closed shape) are recorded
above with their dispositions for the pre-release pass (wave 7 at
writing — wave 8 since the fuzzing wave's renumber); none
blocks. The rider (`pg-fix-scheduler-tick-seam`) closed Finding 1.
Task files audited: all eleven dependency tasks completed with
Notes/Summary filled.