Maelstrom bounded select lifecycle observer — 2026-10-24

Offline implementation based on candidate 164ec9fe3e42d5d1fdf21fdefae1e52588a4cc16 (accepted main 8fe3b51b94f59a17e3af58e08e0cb87cdbb4165a). No main integration, compilation, WebAssembly instantiation, native acceptance, guest, download, deployment, worker launch or box operation occurred. Claude remained idle.

Behavior and use

tools/maelstrom-select-lifecycle-observer.cjs is an optional companion to the unchanged accepted event observer. Its transform(source) is disabled by default and returns the identical source. createObserver(emit) is likewise disabled; its wrapInstance returns the identical instance without reading exports or memory. The disabled CLI does not read its unused export/scheduler arguments.

The enabled CLI has the accepted tool's static transform shape:

node tools/maelstrom-select-lifecycle-observer.cjs --enable-lifecycle \
  OWNED_RUN_JS PINNED_EXPORTS_WAT PINNED_THREAD_MANAGER_JS NEW_ADJACENT_RUN_JS

This only writes runner source; it does not execute it. Never write into the readonly donors. Use --disabled for a byte-identical output. The JS API accepts transform(source, {enabled:true, exportsSource, schedulerSource, lifecycle:{...}}). Existing runner, exports, scheduler, module and configuration pins remain active. The enabled setup validates lifecycle configuration before the instantiate site.

Default lifecycle window: batches 36164–36180 inclusive, at most 512 invocation pairs, one summary of at most 4096 bytes. Allowed caps are 1–4096 invocation pairs, 3072–8192 output bytes, and a batch endpoint difference of at most 4096. Unknown options and invalid limits reject before guest access. Each snapshot reads a fixed number of getters, two thunk words, and at most six stack words through a fully checked direct-memory span. No fd-set traversal, setters, clear_yield, vlan_pump, g2w, guest-read exports or extra run calls are introduced.

There is no per-invocation output or retained sequence. The summary stores fixed counters plus first/last/arrival points and one identity. windowInvocation counts only observed invocations within this window, not the accepted runner's global invocation number. parkedSnapshots counts snapshots, not handler executions. Retry pairs require the same original thunk, metadata, ESP and six-word stack, reason 0 before, reason 8 after, and a non-guard halt after. A reason-8 pair ending at halt 4 is separately counted as a pending guard; it is not a handler retry. Other uncleared parked pairs are counted without claiming execution. Tick comparisons distinguish advancing, equal, backward and unavailable samples. They compare the cached exported tick_count, which select does not update; they do not sample the actual host_get_ticks values used by select. Equal cached ticks can coexist with advancing host time. Neither clock freeze nor expiration follows: host tick at select and the select deadline are unavailable.

The first qualified reason-8 snapshot anchors the original call only. A window starting already parked explicitly leaves onset unobserved. Tracking closes after the first unmatched transition, so another call cannot inherit its retries. Unrelated net waits and unknown/callback-shaped transitions have separate counts. After cap, finish or outside the window, this companion performs no guest reads. Finish emits once; a sink failure cannot replace a guest return or exception. When composed, the accepted observer retains its independent 16-record/8192-byte budget; combined output is at most 17 records/16384 bytes (12288 default bytes). This does not claim a runtime overhead measurement.

Evidence and interpretation

The exact retained events file is committed as test/fixtures/maelstrom-select-arrival-20261024.json (SHA-256 0a176485a70ed9e4a7a896f8b8e5fa276eff806e5000b564557888e18c820560). Replay retains the actual guest register values and stack words, while rebasing synthetic linear backing to avoid allocating a guest arena. The real record is arrival after batch 36164, global invocation 104910, EIP 0x679753 → 0x75040f8, reason 8, thunk [0x80000012,934], stack return 0x7a3362, timeval [1,0]. Its EAX 0x53000001 is not an attributed result. Deduplication explains why absence of later records cannot establish a parked stall. The final 972/972 counters do not establish execution. No original runtime was reproduced.

Pinned handle_select advances ESP by 24, pumps readiness, retains a deadline, and calls vsock_block(24) while waiting. That restores ESP, selects the current thunk EIP, sets reason 8/yield flag and steps 0. The real completion path clears waiting state, rewrites fd sets and stores total into EAX (winsock lines 2745–2778). Pinned runner lines 10697–10731 clear reason 8, await setImmediate, then pump the wire; the next invocation re-enters the handler. Scheduler lines 2640–2647 and 3460–3474 similarly clear/pump and yield the slice. Runner line 7527 assigns tickState.batch; get_tick_count is a readonly global getter. Observed tick progress is reported rather than assuming wall or batch clock deadline behavior. The accepted original run was natural exit 0, 42000 batches, 11.124 s; this remains retained evidence, not new acceptance.

The retained actual argv selects --tick-ms-per-batch=200 and no --real-ticks, startup-clock or time-scale flag. Runner lines 4431–4465 bind h.get_ticks to createGuestClockSource(...).ticks and publish the same source through context. The manifest-matching lib/guest-clock.js selects batchClock.getTicks() in this mode. lib/batch-clock.js derives a base from batch/epoch/paused time and consumes a bounded per-call step, with a monotonic floor. Runner line 7527 advances the batch and line 7530 resets calls within the batch. Reading this host function would itself spend a step, so this passive observer never does so. Its cached tick getter is a separate readonly observation; synthetic advancing/frozen tick tests exercise that observation, not a reconstructed deadline or clock freeze.

The runtime manifest's selected runner/scheduler/module entries match the pins, but it contains no winsock source entry. Separately, the historical Git source b41a7c5f4b752d1f0bdab1141a11d4fed4dc6f9a:src/09d-winsock.wat is byte-identical to the supplied source SHA-256 c281da4763624a29910386d2d6b05a66f98ca92b2a9194827d880c0234ba89ab. This distinction is retained in the selected identity receipt.

Even EIP equal to the saved return address with ESP equal to saved ESP + 24 does not prove a select return at a coarse invocation boundary: intervening code, callbacks or another call may have run, and EAX may have been reused. Such a tuple increments returnCandidates and closes tracking as unknown. witnessedReturns stays zero, apiCompletion stays unobserved, and apiResult stays unavailable.

The missing exact hook must capture an invocation-owned call token, thread/instance, API ID 934, current thunk, pre-pop ESP and return address at select entry, and a matching non-park completion after handle_select has written total (line 2778), before any other guest instruction or callback can overwrite EAX. It must carry the result and post-pop ESP through the dispatcher return/continuation decision; a blocked $vsock_block path must never produce completion. Both thunk-zone and decoded inline CALL/JMP dispatch paths must be covered (win32_dispatch_inner establishes current_thunk_eip, lines 164–180). No existing pinned export supplies that hook. This task deliberately implements honest retry/unknown observation.

Cheap validation

All commands ran under ops/guard.py: Node heap 256 MiB, CPU/wall 60 s, stdout/stderr 131072 bytes, file 24 MiB, process-group TERM/KILL/reap within 5 s. Full command receipts and pass/fail evidence are adjacent in maelstrom-select-lifecycle-observer-20261024-evidence/.

node --check tools/maelstrom-select-lifecycle-observer.cjs
node --check test/test-maelstrom-select-lifecycle-observer.cjs
node test/test-maelstrom-select-lifecycle-observer.cjs RUN_JS EXPORTS_WAT THREAD_MANAGER_JS
node test/test-maelstrom-event-boundary-observer.cjs RUN_JS EXPORTS_WAT THREAD_MANAGER_JS
git diff --check

Syntax, retained-event replay, source-shaped pre/post sequences and existing suite pass. The transformed nested observer setup is executed with a fake instantiate sentinel; no WebAssembly object is created. Return/exception identity and exactly one original run per invocation are checked. The existing suite rejects six mutations; the lifecycle suite rejects seven: dedup-as-stall, stale-EAX-as-result, unrelated-call conflation, cap bypass, disabled wrapping, direct-window guard removal and an extra yield clear. CLI checks establish enabled syntax and disabled identity, including nonexistent unused pin paths.

Two initial test-fixture failures remain recorded: synthetic backing accidentally fell below the direct-window floor, then a bounds fixture still exercised a second guard after the targeted guard was removed. Correcting those fixtures preserved the production guards and rejecting oracle. An early metadata inspection printed the first 30 manifest entries instead of selected entries; no whole manifest was printed, and subsequent reads used selected fields. An attempted historical resource receipt used an absent filename; no unchanged retry occurred. The first staged whitespace check rejected spaces after numbered blank lines in rendered source excerpts. Their committed display copies were trimmed; raw excerpts remain in ops and full-source identities remain unchanged.

Resource admission retains the first 60 MiB projection inside the 64 MiB allocation and its 8 MiB reserve. The correction includes full external launcher/session/WAL, positive SQLite growth against an earlier conservative baseline, future files and publication copies. Initial per-file SQLite sampling was omitted and is explicitly qualified; the earlier baseline overcharges pre-task growth instead of resetting charges. All six pins were freshly hashed before use. Git uses dependency-first flattened readonly alternates and explicit sparse paths: no fetch, clone, repack, full-tree checkout or recursive object-store publication. Historical 24 MiB pack failure and all prior accounting defects remain historical evidence.

No box was assigned. bx_qa9ex5bu shutdown belongs to the coordinator; hostbx_69aem736 was untouched. State/expiry were not queried. Publication is to the task branch only; final commit/remote verification and cleanup are recorded in /tmp/maelstrom-select-lifecycle-observer-oct24/ops/result.json.