Onboarding digest: DST boundary yields a zero-width date window (reports 0 members)

Summary

ONBOARDING_DIGEST computes its date window by mixing UTC normalization with local
getDate/setDate. On a DST boundary the window collapses to zero width (since === until)
and the digest reports 0 new members while appearing to succeed.

Found during a code review of an unrelated PR. The bug shipped in #133, and the tests added
there did not catch it because they only assert instanceof Date on the boundaries, never
their values.

Reproduction

packages/outpost/queue/src/handlers/onboarding-digest.ts — window computation. Under
TZ=Europe/Berlin on 2026-03-29 (spring-forward), since and until come out equal, so the
joinedAt: { gte, lt } query matches nothing.

Test-hygiene items in the same file

packages/outpost/queue/src/__tests__/onboarding-digest.test.ts:

• Cleanup (vi.unstubAllGlobals(), mockRestore()) sits at the end of test bodies rather
  than in afterEach, so a failed assertion leaks a stubbed global fetch and a silenced
  console.log into every later test.
• The console-fallback test leaves fetch unstubbed entirely, so a regression in that branch
  would issue a real network call instead of failing.
• The date-range test asserts only instanceof Date — this is precisely the gap that let the
  DST bug through.
• The current-date default is compared against a freshly recomputed new Date(), so it flakes
  across UTC midnight; freeze with vi.setSystemTime.
• vi.clearAllMocks() keeps implementations, so a persistent mockResolvedValue([]) leaks and
  would silently satisfy an unexpected third findMany call.
• computeFunnelMetrics arguments are never asserted, so swapping the two query results stays
  green.
• The DISCORD_TOKEN-missing-while-channel-set branch (handler throws) is untested.
• mock.calls is annotated number[] instead of the [number] tuple, defeating
  signature-change detection.

Acceptance criteria

• Window computation is timezone-consistent (all UTC or an explicit tz), with a test that
  pins TZ=Europe/Berlin on 2026-03-29 and fails against the current implementation
• The date-range test asserts the actual boundary values, not just their type
• Stub cleanup moves to afterEach
• No test can issue a real network call
• Clock frozen via vi.setSystemTime where "today" is under test

Start with packages/outpost/queue/src/handlers/onboarding-digest.ts and packages/outpost/queue/src/tests/onboarding-digest.test.ts. Read the existing window and test setup, then run the onboarding-digest tests while checking the listed DST, cleanup, network, clock, mock, argument, error-branch, and tuple cases. Done means every acceptance criterion is covered without real network calls or leaking test state.