PR discovery startup cutoff — server service proof
Original: https://github.com/pingdotgg/t3code/pull/13757
Base: 1451e7cb23034d29774338608368a15d8845df36
Replacement branch: maintainer/pr-discovery-backfill-week
Fixture clock: 2026-09-20T00:00:00.000Z

Runs use the actual v2 ThreadPullRequestService.make/start/drain with mocked external dependencies. No git/gh subprocess or provider is started, and no live database is used. Deferred activation, a snapshot-read queue, service drain, and Effect TestClock provide deterministic completion.

Before (main production, new regression present):
11 existing tests pass; one regression fails.
Actual startup branch lookups:
  auto-settled, boundary, old, old-override, old-unsettled, recent, recent-override
Expected:
  boundary, old-unsettled, recent, recent-override
The three extra calls are seven-days-plus-one-millisecond old settlement; old activity auto-settled with a fresh updatedAt; and explicit settled override with old updatedAt and null settledAt.

After (integrated replacement):
All 12 service tests pass.
Startup branch lookups:
  boundary, old-unsettled, recent, recent-override
One minute later:
  old-unsettled
Boundary is inclusive. Recent explicit override uses updatedAt only when settledAt is null.
Four focused server suites: 52/52 passing.
Server typecheck, targeted lint/format and diff check exit 0. Existing lint/Effect advisories remain.

Production scope:
A private seven-day Duration and startup eligibility comparison using Effect Clock and DateTime.toEpochMillis. Existing five-attempt retries, per-thread events, grouping, archived filtering and saved-link behavior remain unchanged.

Limits:
This is headless service/test-level proof, not a full installed-server restart or a 5,000-thread subprocess benchmark. No macOS/Windows, Electron shell, iOS/Android or remote/relay/tunnel clients were run. No UI changes, so screenshots are not applicable. The old contribution's scale benchmark numbers are not asserted for this head.