fix: keep relative timestamps on just now for the first minute

[Astro-Han]

Summary

Relative-time suffixes under one minute were counting seconds (1秒钟前, 2秒钟前, …), which made sidebar task rows and other RelativeTime surfaces flicker. Keep a single just-now label (刚刚 / just now) for the whole first minute, then switch to minute buckets, and delay the ticker until that boundary.

Verification

• packages/core build + node --test dist/__tests__/relative-time.test.js (pass)
• biome format / biome lint on the touched files (pass)
• Did not run repository-wide tests; leave full coverage to CI

AI use

Select exactly one:

• No generative tool made a substantive contribution
• Generative tooling made a substantive contribution

Tool(s) and scope:

Maka implemented the formatter/ticker change, tests, and this PR.

Checklist

• Tests cover the change and fail without it
• Lint, format, typecheck and the affected suites pass locally

Does this PR entail a change in behavior?

• Yes — described under Summary above
• No

[jackwener]

Automated review of exact head b2a51c66b208b2da10a690d51c641deaf62ff71f against current main@d62857a8357e9160926726a2a13096bc2dc2b91d.

The defect remains on main: sub-minute labels count seconds and RelativeTime schedules a render every second. This revision puts the first-minute boundary in the shared pure formatter, reuses it for compact and wide labels, and schedules exactly one timer at the boundary instead of adding a UI-local special case. Boundary coverage at 0/59,999/60,000ms would fail under main. I found no actionable P0-P2 defect; required test is green.

Required conclusions:

1. Optimal for the actual problem: yes; one shared JUST_NOW_MS authority drives both formatting and refresh cadence.
2. Production code to delete: none beyond the superseded second-level branch/comments already removed.
3. Tests to delete/replace: none identified; the two focused tests are behavioral and protect the original defect.
4. Deeper refactor: no.
5. Ready to merge: code-wise yes, but the PR is still Draft and this automated review is not approval.
6. Residual risks/gaps: no rendered React timer test, though the pure delay contract covers the load-bearing behavior; independent human review is still required for this user-visible behavior change.

[jackwener]

Independent review of exact head b2a51c66b208b2da10a690d51c641deaf62ff71f.

No P0–P3. First-minute just-now is a single JUST_NOW_MS authority for both labels and ticker delay. Boundaries 0 / 59,999 / 60,000 ms and future/clock-skew (diffMs < JUST_NOW_MS) hold.

Authorization: Raft #PR-Review-Kabi:8bb877df via dm:@Opus-Qronos-AstroHan:c1beef44 (AstroHan-delegated). Approve only; not merging.

Dismissing: Kabi GitHub-action authorization was withdrawn. Independent Raft GO on b2a51c6 still stands; this APPROVE was submitted under a revoked auth chain.

[jackwener]

Independent review of b2a51c66b.

1. Problem. Sub-minute labels counted seconds, so sidebar rows flickered every second.

2. Solution. One just-now label (刚刚 / just now) for the first 60s, then minute buckets. Ticker waits until that window ends (JUST_NOW_MS - diffMs) instead of ticking every second. Compact path uses the same formatter.

3. Occam. This is the right cut. Tests pin the first minute and the delay.

Merge caveat after #3397: packages/core/src/__tests__/relative-time.test.ts is a new file with no ASF header. The listed test SUCCESS is from 2026-08-21, before that gate. Rebase onto current main and run npm run write:asf-headers before merge, or check:asf-headers will fail.

Not requesting changes on the behavior.

[jackwener]

Correction on the header caveat: only the new test file packages/core/src/__tests__/relative-time.test.ts lacks an ASF header.

relative-time.ts and relative-time.tsx already have headers on current main (added by #3397). This branch is from before that. On rebase, keep main's header and put the new JSDoc under it — do not overwrite the header with the branch's first-line JSDoc.

[jackwener]

Re-review of e65809320 after the header fix.

Behavior still holds: first minute is one just-now label; ticker waits for that boundary. The new test file now has the ASF header. Required test on this head started 18:42Z, after #3397's header gate, and is SUCCESS.

Approve. Merge stays with the author.

Withdrawing: nextRelativeRefreshDelay() treats future timestamps (diffMs < 0) as just-now remaining time, so delay = JUST_NOW_MS - diffMs can exceed Chromium 32-bit setTimeout. Overflow fires immediately; RelativeTime effect reschedules every render. Hold until a new head clamps age to >= 0.

[jackwener]

Withdrawing the previous APPROVE on this head.

P1: nextRelativeRefreshDelay computes JUST_NOW_MS - diffMs for every diffMs < JUST_NOW_MS. That formula is only valid for 0 <= diffMs < 60s. Future timestamps (diffMs < 0) — already documented as clock skew → just-now — produce an unbounded delay. Around 25 days ahead it exceeds INT32_MAX; Chromium/Node setTimeout overflow-fires in ~1ms. <RelativeTime>'s effect has no deps and reschedules every render, so that becomes a tight loop.

This is a regression from main, which used a constant 1000 in the same branch (1Hz tick, no overflow).

Do not clamp at setTimeout. Age should never be negative:

const diffMs = Math.max(0, now - ts);

Future records stay on the just-now label and refresh on the existing 60s cadence until the clock catches up. That also survives clock correction; a 24-day timer would not.

Please add tests for ts > now and ts = now + 30d asserting the delay stays in the cadence table, not JUST_NOW_MS - diffMs.

[jackwener]

JUST_NOW_MS - diffMs assumes diffMs >= 0. Future ts makes this unbounded and can overflow Chromium's 32-bit timer.

Minimal/clean fix is the same line: const diffMs = Math.max(0, now - ts) so age is a non-negative duration. Then this branch returns JUST_NOW_MS for future timestamps, which matches the documented just-now display and cannot overflow.

[jackwener]

Independent re-review of a59cbb0.

The future-timestamp timer overflow is closed: one private relativeAgeMs() = Math.max(0, now - ts) feeds all three consumers. Delay for +30d is 60s, not an unbounded remaining window. Required test on this head is SUCCESS. Non-blocking: the new test name still says "finite" while the sample includes NaN/±Infinity.

Approve. Merge stays with the author.