Input consumption timing technical review — 2026-10-24
Historical source last updated, 2026-10-24. Runtime and implementation claims below describe that dated source; see current app notes and source for subsequent changes.
Verdict: REJECT candidate 01da39b4471d7f2827be043804b755aa9bb6b141. Its parent and the observed remote main are
adb2852f0ad9944625d809fd9b73c50dc3d4514e. This review branch starts from
that accepted parent and publishes review evidence only. It does not integrate
the rejected implementation or change main.
The trace labels raw batch-clock state as lastGuestTick, despite the production
startup clock mapping guest-visible ticks. In test/run.js:4422, the startup
wrapper maps the return of batchClock.getTicks() through startup dilation. The raw clock still writes its unmapped value to tickState.lastTick. The new snapshot at test/run.js:4473 emits that raw value as lastGuestTick. The candidate documentation explicitly describes this field as the last tick
read by the main guest. It is a different clock domain in this supported mode.
The source-only reproducer executes the exact production startup wrapper and
trace setup, using the real batch, startup, guest-clock and trace implementations. With factor 0.01, maxMs=30000, batch step 200, batch 7, and one guest
tick read, the guest receives 14. The observation says:
[input-schedule] scheduledBatch=2 action=keydown correlation=unavailable batch=7 guestMs=14 clock=batch lastGuestTick=1400
The raw 1400 is not a stale guest return: the only guest read returned 14. This is a timing correctness defect, not formatting. The registered Comanche
Gold startup configuration uses this factor; the CLI also supports explicit
startup-clock configuration. Recording an actual guest-visible last return, or
making its unavailability/raw domain explicit, needs a bounded producer repair
and a regression for the production wrapper before reconsideration. This review
does not modify the candidate to make its oracle pass.
Controls and scope
The candidate's test/test-cli-tick-ms-input.js --source-only,
test/test-batch-clock.js, and test/test-guest-clock-source.js pass. The first
executes production schedule/enqueue/poll/route/DispatchMessage hook bodies with
real clocks and controlled MSG memory; it covers schedule batch 2, dequeue batch
7, and dispatch batch 11, rather than only mocking the trace helper.
The adjacent input-consumption-independent-20261024/independent-controls.js
adds independent controls. Its normal invocation exits 0 after confirming the
defect; --assert-correct-last-tick exits 1 on 1400 !== 14. Both receipts are
retained. Other independent controls pass:
- The full production renderer input module is installed and its actual
takeInput/checkInputmethods retain the raw event object, so WeakMap enqueue identities survive dequeue. Identical balanced key and mouse events retain distinct observation IDs, FIFO order, and full message fields. The producer's simpler renderer queue test alone did not establish this identity property. - Real scaled/mapped publish observations spend no guest tick probes. Batch snapshots do not change clock state. Ordinary tracing enabled/disabled returns, routes, legacy logs, FIFO and batch state agree.
- Duplicate DispatchMessageA/W tuples at a later batch carry
correlation=unavailable, with no inferred enqueue/dequeue identity. - Mixed record types produce 1000 records and one notice, with exactly 1000 snapshot calls. No further clock samples occur after the cap.
- Empty polls add no observation; a route request before any dequeue has no
timing identity. After a consumed event, the existing last-event state remains
available by design. The inspected WAT call sites invoke routing only after
nonzero
host_check_input; calling route after an empty poll is not evidence of another input consumption. Existing focus/explicit-hwnd behavior is preserved.
The cap bounds output and snapshot calls, not every operation: enqueue identity bookkeeping and dispatch MSG decoding still occur after suppression. The identity map uses weak keys and is not an independently retained event queue. No correctness failure was established in that behavior.
Reproduction and evidence
Run with a source tree at the exact candidate (no build or guest needed):
node --max-old-space-size=256 docs/re-notes/input-consumption-independent-20261024/independent-controls.js /path/to/candidate-source
node --max-old-space-size=256 docs/re-notes/input-consumption-independent-20261024/independent-controls.js /path/to/candidate-source --assert-correct-last-tick
The complete renderer-input module runs in a Node VM context, without browser, canvas, Wasm or guest construction. Unused renderer region metadata is supplied as a minimal placeholder; the tested FIFO/key-state path uses no guest memory. Production API hook bodies use a controlled 1 KiB MSG buffer.
Adjacent receipts include candidate Git blob IDs, SHA-256 identities, exact test
commands, memory admission, output hashes, elapsed times and outputs. Producer
tests.json, publication.json, and cleanup.json were read from immutable
input-consumption-timing-oct24 and their evidence is not elevated to native
or gameplay qualification.