This one doesn't make any sense to me either.
Some quick tests show that:

1. citeproc-js indeed behaves this way (as one would hope given the test)
2. This is unrelated to the numbers in the title: Title A, Title B, Title C behave the exact same way

<sort>
  <key variable="title"/>
  <key macro="issued-date"/>
</sort>

Sorts as

Title 1
Title 2
Title 3

as expected (with or without the title being part of the issued-date variable).

It'd be great if @fbennett could chime in, but I'd currently call this as an incorrect test and suggest changing it (likely by just removing the title variable from the macro since the test is about dates)

Ah, I remember this one. Rintze and I discussed the case in 2013, back in the BitBucket days. In the discussion Rintze also questioned the discarding of the non-date text from the macro when generating keys, and noted that doing so is an off-spec invention.

As I note in the linked thread, intermixing date and string values in a single key can't be made to produce intuitive sort results. Dates are a bit special: keys for non-date variables are simple strings or numbers that can be derived directly from the variable; but date keys must generate a machine-readable sort key that takes account of the date-parts element of the rendered form. As a result, there are cases in which a macro containing a condition must be used to produce the date key (APA is mentioned in the discussion).

The need to express date keys in a macro can be addressed in one of two ways. One would be to assume that such macros will always contain exactly one date. The other would be to extract a date from an existing macro, and ignore everything else it contains. citerproc-js adopts the latter approach. While it (obviously) can create confusion, it has the advantage that, if a style designer mistakenly assumes that a macro combining two elements (creator and date, say) will treat the contents as two atomic keys, the attempt will throw bad sort results immediately. (There is an example of this in the repo, in the initial history of deutsche-sprache.csl, where the initial checkin of the style is followed immediately by an amendment to the sort keys.)

So that's the reasoning behind it. On a quick check of the repository (based on macro names alone), it looks like a few styles would need to be amended to add and call date-only macros if the former approach is preferred.

beltz-padagogik.csl
berlin-school-of-economics-and-law-international-marketing-management.csl
biuletyn-polskiego-towarzystwa-jezykoznawczego.csl
deutsche-sprache.csl
hochschule-hannover-soziale-arbeit.csl
keel-ja-kirjandus.csl
zeitschrift-fur-qualitative-forschung.csl
wirtschaftsuniversitat-wien-abteilung-fur-bildungswissenschaft.csl
wirtschaftsuniversitat-wien-unternehmensrechnung-und-controlling.csl
wirtschaftsuniversitat-wien-wirtschaftspadagogik.csl

As I note in the linked thread, intermixing date and string values in a single key can't be made to produce intuitive sort results

I still don't understand this, even after reading the linked discussion. It's perfectly coherent to sort first on a text field, then on a structured date field (whose fields are determined by some kind of condition).

The test is muddled, and should be divided into two separate fixtures for clarity. The reason for boiling the macro down to the date alone is that sorting on a string concatenation of title and date can produce different results than sorting atomically on the title and then on the date. Things can get messed up if the title ends in a number. Here are three entries sorted by title, then by date:

Title     | 2002-12-01
Title 1   | 2002-11-01
Title Two | 2002-10-01

If the same entries are sorted on a concatenation of title and date, the order comes out differently:

Title 1 2002-11-01
Title 2002-12-01
Title Two 2002-10-01

It's not a common case, but it does come up. Date extraction went into the processor in the early days, and I don't have the records on it to hand, but I'm pretty sure it was prompted by a user complaint about a badly sorted bibliography.

Thanks -- the reason now makes more sense to me, though I'm still not clear what's particular about dates here:

Title | 99 Red balloons
Title 1 | 1 Flew over the cuckoo's nest

Also reverses sort order when concatenated into a single string.

My inclination would therefore be to have processors/tests behave according to spec and to discourage using multi-variable macros for sorting, potentially even in the specs (discourage rather than forbid because there may be cases for them, though only rarely).

That's the best for processor simplicity. The styles listed above should be amended to assure consistent results, though.

The styles listed above should be amended to assure consistent results, though.

I'm working on that -- seems like a good idea anyway in case other processors decide(d) to ignore the test & follow the letter of the specs and it won't affect output given current citeproc-js behavior, so we don't need to coordinate timing.

Will also create a PR for changing the test by removing the title from the issued-date macro.

@bdarcus would you create a PR for the specs that adds something along the lines of:

Except in rare cases, macros that render more than one element should be avoided as sort keys, as they can lead to unexpected sorting in edge cases. Instead, use their elements as individual, consecutive sort keys.

(wrote this quickly, happy for edits)

@jgm I think that should settle this, you can assume that the test will change -- thanks for bringing this up.

Shouldn't we just remove this test?

The file name, and purpose, of the test would change too, if we remove the macro call, and there should be other tests that cover this sorting?

If not, then I'm misunderstanding the resolution, and so what the change would be.

I interpret the name & content of the test to test sorting by a macro where the date format is conditional on the item type and the test would still do this without the title in the macro.

Edit:
This appears to be Frank's view as well, see above:

The test is muddled, and should be divided into two separate fixtures for clarity.

So instead of two separate fixtures, we remove one of the things the fixture tests for