Assertion failure in year-and-a-bit durations relative to leap day

[ptomato]

I observed this bug with the reference polyfill:

const d = new Temporal.Duration(1, 0, 0, 0, 1);
const relativeTo = new Temporal.PlainDate(2020, 2, 29);
console.log(d.round({ smallestUnit: 'years', relativeTo }).toString());
console.log(d.total({ unit: 'years', relativeTo }));

This fails the following assertion:

proposal-temporal/polyfill/lib/ecmascript.mjs

Line 3261 in 857d12d

if (sign === 1) assert(startEpochNs.leq(destEpochNs) && destEpochNs.leq(endEpochNs), `${unit} was 0 days long`);

It occurs because destEpochNs (2021-02-28T01:00Z) > endEpochNs (2021-02-28T00:00Z). If I understand the problem correctly, startEpochNs ought to have the latter value and endEpochNs ought to be 2022-02-28T00:00Z, probably because r1 and r2 should be 1 and 2 instead of 0 and 1.

I'm hopeful that this is a polyfill-only bug and not a problem with the spec text, because three other implementations (SpiderMonkey, temporal_rs, Fullcalendar polyfill) all print the correct results:

P1Y
1.0001141552511414

I caught this while writing snapshot tests. This means that it isn't currently covered by test262, or it would have been caught already.

[ptomato]

I'm hopeful that this is a polyfill-only bug and not a problem with the spec text, because three other implementations (SpiderMonkey, temporal_rs, Fullcalendar polyfill) all print the correct results:

I'm less optimistic about this now. I examined the values in a debugger in SpiderMonkey and temporal_rs and they are the same as I observed above. This code also fails an assertion in SpiderMonkey, which I didn't notice before because I was using a release build. temporal_rs doesn't have any assertions here.

One possible outcome is just that the assertions are too strict and we should remove them, since the algorithm does seem to produce correct results (in this particular case.) But I am still suspicious of the bounding window being between 0-1 years and not 1-2 years, since the total is >1 year. So it needs more investigation.

I'm not familiar enough with the code in fullcalendar/temporal-polyfill to say whether they have this same suspicious bounding window. I believe the place to look would be nudgeRelativeDuration or clampRelativeDuration. (cc @arshaw)

[catamorphism]

Same assertion failure happens with:

const d = new Temporal.Duration(1, 0, 0, 0, 1);
const relativeTo = new Temporal.PlainDateTime(2020, 2, 29);
relativeTo.until(relativeTo.add(d), { smallestUnit: 'years' });

[catamorphism]

This triggers a similar assertion failure:

const d = new Temporal.Duration(0, 1, 0, 0, 1);
const relativeTo = new Temporal.PlainDateTime(2025, 3, 31);
d.round({smallestUnit: 'month', relativeTo})

Same result as the original, but with "month" instead of "year" in the assertion failure message.

[catamorphism]

#3172 should fix this in the polyfill, but it also needs a spec change.