Date-boundary unit tests drift when they assert UTC instants against local-day logic#

Problem#

A unit test asserted a hard-coded UTC instant ("2026-01-31T23:30:00.000Z") against a predicate that resolves its day bounds in the viewer's local timezone. The test passed in UTC CI but would fail on any developer machine at a positive UTC offset — a latent, environment-dependent failure that CI can never catch because CI runs in UTC.

Symptoms#

matchesInventoryFilter("2026-01-31T23:30:00.000Z", { preset: "custom", after: "2026-01-01", before: "2026-01-31" }, NOW) returns true on a UTC runner but false at UTC+1 (the instant is 2026-02-01T00:30 local, past the local end-of-day for Jan 31).
The mirror case: "2026-01-01T00:00:00.000Z" returns true in UTC but false at UTC−1 (it is 2025-12-31T23:00 local, before the local start-of-day for Jan 1).
Green in CI, red only on some contributors' laptops.

What Didn't Work#

Trusting the green CI run. The 37-test suite passed unmodified through a hand-rolled → date-fns refactor, which read as "behavior preserved." It only proved the tests agreed with the code in UTC — both the fragile literals and the CI runner shared the same zone, so the bug was invisible.

Solution#

Construct the boundary instants in local time so the test's expectation is computed in the same frame the predicate uses. In src/domain/magazines/__tests__/inventory-filter.test.ts:

// Before — UTC literal, drifts across the day boundary on non-UTC runners:
expect(
  matchesInventoryFilter("2026-01-31T23:30:00.000Z", range, NOW),
).toBe(true);

// After — local-time construction, deterministic on any runner:
expect(
  matchesInventoryFilter(
    new Date(2026, 0, 31, 23, 30).toISOString(), // local Jan 31 23:30 → serialized
    range,
    NOW,
  ),
).toBe(true);

new Date(year, monthIndex, day, h, m) builds the instant in the runner's local zone; .toISOString() then serializes it (as UTC) for the string-typed predicate input. The local calendar day is now fixed regardless of offset, so it always lands inside the local [startOfDay(after), endOfDay(before)] window.

Why This Works#

The production predicate resolves its bounds in local time. In src/domain/magazines/inventory-filter.ts the custom branch does:

const entry = parseISO(lastInventoriedAt);
if (filter.after && isBefore(entry, startOfDay(parseISO(filter.after)))) return false;
if (filter.before && isAfter(entry, endOfDay(parseISO(filter.before)))) return false;

parseISO("2026-01-31") (a date-only string) yields local midnight, and date-fns startOfDay/endOfDay operate in local time — so before covers through 2026-01-31T23:59:59.999 local. The entry, however, is an absolute instant. A UTC literal like ...T23:30:00.000Z is a fixed point on the timeline; which local calendar day it falls on depends on the runner's offset. At UTC+1 that instant is already Feb 1 locally, so isAfter(entry, endOfDay(local Jan 31)) is true and the row is excluded — the opposite of the UTC result. Building the instant from local Y/M/D removes the offset dependency: the assertion pins a local wall-clock time, matching how the predicate reasons.

Prevention#

Match the test's time frame to the code under test. If the SUT interprets dates in local time (parseISO on a date-only string, startOfDay/endOfDay, new Date(y, m, d)), construct test fixtures with new Date(y, monthIndex, d, ...), not UTC ...Z literals or toISOString()-derived day strings. If the SUT works in UTC (Date.UTC, parseISO on a ...Z timestamp), use UTC literals. Never mix frames across the assertion.
A green UTC CI run is not evidence of timezone-correctness. CI almost always runs in UTC, so it structurally cannot catch UTC-vs-local drift. Treat any date-boundary assertion as suspect until it has been reasoned about (or run) at a non-zero offset — e.g. TZ=America/New_York bun test or TZ=Asia/Tokyo bun test locally.
Boundary values are where this bites. Mid-interval fixtures (a date ~15 days inside a month range) are robust to a ±1-day offset shift; only instants within ~1 day of a startOfDay/endOfDay edge flip. When testing inclusivity of a day boundary, that closeness is the whole point — so those are exactly the assertions that must be built in the local frame.

Introduced during the last-inventoried inventory-date filter (PR #70 / #72); the local-frame fix landed in the PR #72 review round.
Related tests-failures learnings (different root causes): docs/solutions/test-failures/bun-test-misloads-playwright-e2e-specs.md, docs/solutions/test-failures/e2e-dotenv-mise-clobbers-launcher-env.md.