fix comparator bisector

[mbostock]

Fixes #249. Previously #227.

If the bisector was given an asymmetric comparator that takes (object, value) as arguments, then the test for whether the search value is orderable was failing because it would pass the comparator (value, value), which the comparator likely did not expect, and thus would typically return NaN or undefined.

So instead, we now perform the bisection without first testing for orderability, but if the final comparison at the end of bisection returns NaN, we know that the search value was not orderable. In the case that the search value is not orderable, this is obviously slower since we still perform the bisection, but we don’t need to optimize for this case and it’s better to be correct and backwards compatible. We can’t do this; see comment below.

So instead, we now limit the search value orderability test to when the bisector uses an accessor or a known symmetric comparator (specifically d3.ascending and d3.descending, which are used by d3.bisectLeft et al.). It’s a bummer that this means we can’t enforce that non-orderable values go to the end in the general case, but I don’t think it’s possible given the constraints, at least not without breaking backwards compatibility and requiring the use of symmetric comparators (or accessors).

[mbostock]

Hmm, I just realized that this won’t work because the array could only have non-orderable values in it, and thus the last comparison could return NaN even if the search value is orderable. For example, this now fails:

const values = [null, undefined, NaN];
assert.strictEqual(bisectLeft(values, 1), 0);
assert.strictEqual(bisectLeft(values, 2), 0);

If the comparator is specified as a function that takes (object, value), there’s no way for us to test just an object for orderability (because we can’t construct a synthetic value to test against), just like we can’t test just a value for orderability (because we can’t construct a synthetic object to test against). So… I think we might be stuck here, although we could special case specific comparators that we know to be symmetric. 🤔