Test coverage missing for `roundingMode: 'halfExpand'` in `since` for negative numbers with 0.5 remainders

[justingrant]

Looks like rounding using the 'nearest' mode isn't doing away-from-zero rounding for negative 0.5 remainders. It seems to be rounding towards zero, i.e. same behavior as Math.round() which isn't correct per the proposal. From #827:

5.4.4 'nearest' - round to nearest integer, with 0.5 rounded away from zero.

Repro:

Temporal.Instant.fromEpochMilliseconds(0).subtract({milliseconds: 500}).round({smallestUnit: 'seconds'}).toString()
// => "1970-01-01T00:00:00Z" - OK (rounding towards beginning of time)

Temporal.Duration.from('-PT0.5S').round({smallestUnit: 'seconds'}).toString()
// => "-PT1S" - OK (rounds away from zero)

Temporal.DateTime.from('2020-01-01').subtract({milliseconds: 500}).since('2020-01-01', {smallestUnit: 'seconds'}).toString()
// actual: "PT0S"
// expected: "-PT1S"

[ptomato]

The existing behaviour seems correct to me. For non-Duration types we are treating "zero" as −∞ for the purposes of rounding (the principle of "zero is the Big Bang, not the Unix epoch", also from #827), effectively considering non-Duration types always positive. and that's why we treat floor and trunc as identical for non-Duration types. Given that, it seems consistent to round ties upwards in nearest mode.

[justingrant]

OK, makes sense re: Instant. I updated the OP to show a repro with DateTime.prototype.since that's still failing, though.

[ptomato]

Still seems correct to me, I believe we made "zero" == "the beginning of time" for all of the non-Duration types?

[justingrant]

Yep, but this is a Duration being rounded, not a DateTime. Duration rounding should round away from zero for negative durations.

[justingrant]

BTW, should I mark these failing tests with skip() in the PR for #993 if a fix for this issue doesn't land before my PR does? Or should I leave them as failing and they'll get fixed when the rounding fix comes in?

[ptomato]

skip would be better, with a comment pointing to this issue.

[justingrant]

It turned out that the failed tests in the #993 fix were my fault and were unrelated to this rounding problem, so I didn't need to skip any tests.

[justingrant]

Haha, looks like the problem doesn't repro anymore after #1023 and was merged. I'm not sure if that PR or @ptomato's earlier refactor of DateTime's add/subtract fixed it, but I'm OK to consider this fixed.

I'll leave this issue open in case someone wants to add a test case to handle this case because it was presumably a test hole before.

[ptomato]

This is a polyfill or test262 thing and not a requirement for Stage 3.