Files
alkstore/tasks/review-wave-2.md
T

143 lines
6.6 KiB
Markdown

---
id: review-wave-2
name: Review gate — wave 2 (substrate fork)
status: completed
depends_on: [fork-provenance-and-floor]
scope: narrow
risk: low
impact: component
level: review
tags: [wave-2, review]
---
## Description
Review the substrate fork before wave 3 builds the engine on it.
Primary lens: **diff reviewability against the lineage** (ADR-012 §3's
fidelity posture — the fork's economics rest on future cherry-picks
staying cheap) and **the defect class staying out** (the fork's
reason for existing).
Check:
- **Fidelity**: the kept half's module structure and internal names
match the lineage; renames confined to the table family and hygiene
deltas; the diff vs. `/workspace/honker` @ `f4e53c6` contains
nothing unregistered.
- **Port deltas**: W-1/W-2/dead-man's-switch applied as decided; no
other behavior changes smuggled in; W-4 (RW watcher connection)
deliberately kept.
- **Drops**: cron, experimental watchers + their optional deps,
rate-limit/result tables, superseded queue functions — absent, and
their absence registered.
- **Re-derivation vs. ADR-010**: stamps land per §3a; the claim
statement's ordering and visibility semantics match §1–§2; the
validity predicate is uniform across ack/heartbeat/retry/fail;
savepoint hardening real (test-read it, don't trust the test name);
sweep covers both states; the curve/boundary arithmetic is *not* in
the substrate (contract-blind boundary — ADR-012 §2).
- **Provenance**: register complete and accurate per ADR-018; license
notice correct.
- **Family standards** in the substrate: sync only, no panics, no
error swallows, no comments (doc comments fine).
- **Gates**: full engine-crate test suite, clippy `-D warnings`, fmt
clean.
## Acceptance Criteria
- [x] Lineage diff reviewed; every divergence registered or fixed
- [x] Re-derivation matches ADR-010's pinned semantics (spot-checked
against the ADR text, not just the tests)
- [x] Provenance register accurate
- [x] All gates green
- [x] Findings recorded; wave 3 decomposition may proceed
## References
- docs/architecture/decisions/012-forked-substrate-design.md §3
- docs/architecture/decisions/010-queue-semantics-depth.md
- docs/architecture/decisions/018-provenance-register-and-cherry-picks.md
- docs/plans/implementation.md (Review gates)
## Notes
Review performed over a fresh full-coverage lineage diff (subagent,
honker-core @ `f4e53c6` via `git show` vs the substrate tree) plus an
ADR-010/020/019 spot-check of the re-derivation against the ADR text
(not the tests).
**Fidelity — clean.** Every kept-half divergence is register-addressable:
the table-family rename, W-1 (backoff tick counter verified at
`MAX_RECONNECT_TICKS`; initial-open path lineage-identical), W-2
(fallible spawn; baseline-wait carried verbatim), dead-man's switch
(panic → log+`return`; WatcherDeathGuard byte-faithful), W-4 keep-RW
(open flags identical to lineage), the `pragma_table_info` race re-key,
the D-16 attach-prune cap, `cron_expr` → `spec`, schema/migration
extensions (column-order invariant pinned), and hygiene deltas. All
kept internal names preserved. Drops verified absent (no cron symbols,
no kernel/shm watcher, no rate-limit/result surfaces, no scheduler
pause/resume/list/update). queue_ops.rs smuggle-scan: no `.ok()`
swallows, no lineage-identical defect-carrying code.
**Re-derivation vs ADR-010 — one real defect found and fixed this
session (D-27)**: `scheduler_tick` passed the schedule row's *relative*
`expires_s` directly into the re-derived `enqueue`'s *absolute*
`expires_at` parameter — the lineage resolved `expires_s → now + s`
inside its enqueue, and ADR-020 §2 pins the relative→absolute
resolution at the enqueue instant. A schedule with `expires_s: Some(30)`
would have fired jobs with `expires_at = 30` (expired since 1970):
swept straight to dead, silent work loss. No prior test covered expiry
through the tick. Fixed at the fire site in the tick's transaction
(one clock read, `now + s`); regression test
`tick_resolves_the_fires_expires_sticker_to_an_absolute_row_value` pins
the resolved value and the fired job's claimability. This was the
review's only code-behavior fix.
**Minor, fixed inline (D-28)**: the cross-mechanism pressure test's
lock-churn branch used a literal `'p{producer}'` (no interpolation —
all producers contended on one lock name) and swallowed its result via
`.unwrap_or(0)`; now per-producer names, periodic release, errors
propagate. Test-quality only.
**Minutiae, no action needed (recorded here, not as register rows)**:
`column_present`'s prepare-error → "absent" fallback produces a
different error message than the lineage on a missing table (bootstrap
still errors either way); bootstrap/`poll_data_version` doc-comment
trims (D-15-adjacent); the concurrent-open race test's tuned magnitudes.
None change behavior.
**Register**: D-27 (re-derivation fix) and D-28 (test hygiene) appended
after the review's fixes; header updated. Cherry-picks stay empty.
**Gates**: `cargo build` ok; workspace `cargo test` — 23 core + 3
contract-suite + 84 sqlite (83 inherited + 1 new regression) — green in
one run; `cargo clippy --all-targets -- -D warnings` clean;
`cargo fmt --check` clean. The flake watch-record of
`fork-provenance-and-floor` (one unreproducible transient) did not
recur in this session's runs.
**Boundary arithmetic check (ADR-012 §2)**: the equal-jitter curve's
arithmetic is absent from the substrate — `retry` takes a caller-
resolved literal delay; no `backoff`/curve symbols exist. The schedule
boundary math present (`parse_every_interval`, `next_every_boundary`,
`skip_forward`) is the ADR-010 §6 / fork-rederive-queue-ops-Notes
sanctioned re-derivation surface (`@every`-only numeric advance per
ADR-009 §2; cron machinery and local-TZ arithmetic absent).
**Wave 3 watch-item (no action this wave)**: the W-2 `Result<_, String>`
error is untyped — the engine-layer mapping into the `Database` variant
is wave-3 work per fork-port-connection-watcher's Notes; nothing here
blocks decomposition.
## Summary
Wave-2 review gate complete. A fresh full-coverage lineage diff
(honker-core @ `f4e53c6`) found every kept-half divergence
register-addressable — no unregistered divergence. The ADR-spot-check
of the re-derivation found one real contract-semantics defect
(scheduler fire's relative `expires_s` passed through as an absolute
`expires_at`; silent expiry-at-1970 work loss under
`expires_s`-bearing schedules) — fixed inline with a regression test,
registered D-27; one test-quality wart fixed and registered D-28.
Gates: build, workspace test (110 green), clippy `-D warnings`, fmt —
all clean. Wave 3 decomposition may proceed.