Test flake: room-invite-preview compares a rendered countdown against a second Date.now(), losing a second at the boundary #1839

Implementation: FlyBot (ours) — PR follows within 24h. Do not pick up without a pickup-ok comment on this issue.

Written-by: FlyBot Lead
Claimed-by: FlyBot Lead
Claim-Id: a497345b7dd544c0b87a56a96d6217cc
Branch: dev/2-invite-preview-clock-once
Base: f245520
At: 2026-10-30T13:51:53Z

What goes wrong

core/test/room-invite-preview.test.js:24-30 renders the card and then compares its deadline line against a clock read after the render:

const tree = RoomInvitePreview({ h, preview, onJoin() {}, onClear() {} })   // render calls inviteCountdown → Date.now()
...
assert.equal(tree.children[3].children.join(''), inviteCountdownAt(preview.expiresAt, 0, Date.now())) // second Date.now()

The render (core/src/client/room-invite-preview.js:78) formats inviteCountdown(expiresAt, clockOffsetMs), which reads Date.now() itself; the assertion then reads Date.now() again. Two reads of a wall clock are not the same value, so when the two calls straddle a second boundary the expected text is one second ahead of the rendered one and the test fails by construction — not from any product defect.

Observed once on f245520c in Dev-1's full verify run of the #1838 branch (that branch does not touch this file): expected 37958984:53 left to join, actual 37958984:52 left to join — same minutes, total % 60 off by one, exactly the boundary. Single-file rerun 60 times: no reproduction (the window is ~1 second), which is why this is a latent flake rather than a hang.

preview.expiresAt is 2099-01-01, so the two readings normally land in the same second; only a boundary hit shows it. Other counts in the same file are safe (:74-76 computes future and the expectation from the same Date.now() window and allows 1:0\d).

Expected

One clock value must drive both the render and the assertion. Freeze it for the test — no tolerance window:

• const real = Date.now, const frozen = real(), Date.now = () => frozen, restore in finally.
• Render, then assert the deadline line equals inviteCountdownAt(preview.expiresAt, 0, frozen) — the same instant the render saw; the render's inviteCountdown reads Date.now(), which is now constant.
• No ±1 s tolerance and no "matches before or after" leniency: either would hide a genuine off-by-one inside inviteCountdown itself.
• No production change: RoomInvitePreview/inviteCountdown keep using the wall clock; only the test pins it. (Injecting a clock parameter into the component is acceptable as an alternative, but then the render must use it, not just the test.)

Deterministic red-first

Two probes on the same file, both restoring Date.now in finally:

1 Current code is clock-sensitive: advance the clock one second per read (let t = Date.now(); Date.now = () => (t += 1000)), render, run the current assertion → it fails deterministically (t+1s vs t+2s), proving the race.
2 The fixed assertion is immune and not vacuous: freeze the clock (Date.now = () => frozen), render, assert against frozen → passes; then mutate the expectation by one second → fails.

Scope

• core/test/room-invite-preview.test.js and its byte-identical mirror patches/ds-bot/files/test/room-invite-preview.test.js (both identical at f245520c; that is the whole repo convention for test mirrors, not a new patch). No .patch regeneration, no topics.json change.
• Test-only, T1: production files untouched, so no PASS is needed and no user-visible behaviour changes.

Validation

• The deterministic probe above: old assertion red, new assertion green.
• The file run 200 times consecutively on a loaded box with the clock frozen, 0 failures; plus one full npm run verify on the committed clean tree with ARCH_DOCS_BASE=<live main base>.
• node scripts/ds-bot-patches.mjs check → the three green lines (mirror is compared byte-for-byte).

Dedupe (2026-10-30)

Grep of live main f245520c: the only two-clock comparison is this file's :28; inviteCountdownAt callers in :74-76 derive both sides from one clock window. Issue search for room-invite-preview / countdown flakes: only #798, #846, #458 (closed, feature defects); no open or closed issue covers this assertion race.