Write a regression test for a cursor-paginated endpoint that previously skipped rows sharing a timestamp.

Seed more equal-timestamp records than fit on one page in an isolated store. Follow nextCursor until null and compare all returned ids with the seeded set. Insert a newer record after page one and verify it does not shift forward traversal.

Twelve records share one timestamp with page size five. Three pages return all twelve ids exactly once.

• Treat cursors as opaque strings in the client test.
• Verify ids, not just page sizes.
• Bound the page-walking loop.
• Use deterministic isolated seed data.

I test the API as a consumer would: request a page, retain ids, follow the returned cursor, and stop at null. A set comparison catches missing rows while a duplicate assertion catches unstable tie handling.

[Fact]
public async Task Cursor_paging_returns_every_tied_row_once()
{
    await using var factory = await ApiTestFactory.WithTiedEventsAsync(count: 12);
    var client = factory.CreateAuthenticatedClient("user-1", "events.read");
    var seen = new List<Guid>();
    string? cursor = null;

    for (var pageNumber = 0; pageNumber < 10; pageNumber++)
    {
        var url = QueryHelpers.AddQueryString("/api/events", new Dictionary<string, string?>
        {
            ["pageSize"] = "5",
            ["cursor"] = cursor
        });

        var page = await client.GetFromJsonAsync<EventPage>(url);
        seen.AddRange(page!.Items.Select(item => item.Id));
        cursor = page.NextCursor;
        if (cursor is null) break;
    }

    Assert.Equal(12, seen.Count);
    Assert.Equal(12, seen.Distinct().Count());
    Assert.Equal(factory.SeededEventIds.Order(), seen.Order());
}

This test is cheap with deterministic isolated data. If provider translation determines cursor behavior, use a disposable relational database rather than the in-memory provider.

• All records share one timestamp.
• The final page is partially filled.
• A new record arrives after page one.
• The cursor is malformed or paired with changed filters.

• Seed ties larger than one page.
• Assert the loop terminates.
• Compare concatenated pages to a known snapshot.

• Asserting only that each page has at most five rows.
• Sorting by timestamp without id.
• Decoding and rebuilding opaque cursors in the test.
• Using unbounded page walking.