Buildkite Test Engine: fix JSON comma separate bug

[pda]

This pull request fixes a missing-comma-separator bug in the Bootleg JSON writer 🏴‍☠️ which is responsible for producing JSON test results for Buildkite Test Engine.

There was a bug when serializing to JSON array (from AbstractVector) where elements that were === / “egal” to the final element would not have a comma separator placed after them, producing invalid JSON. I stumbled upon this because I work at Buildkite; I've previously contributed a fix for a different issue in `test/buildkitetestjson.jl.

Specifically, this bug occurred when the last item of expanded or backtrace within failure_expanded was an empty string, because elt === last(val) was "" === "" was true, so e.g. ["hello", "", "world", ""] would serialize to:

[
  "hello",
  ""
  "world",
  ""
]

Note this was not happening when the last element was a non-empty string; the strings come from split(string, '\n') which returns a singleton empty string for zero-width splits, but separate strings for non-empty splits:

julia> [[x, objectid(x)] for x in split("\n", '\n')]
2-element Vector{Vector{Any}}:
 ["", 0x37373e5ed3d3af9c]
 ["", 0x37373e5ed3d3af9c]

julia> [[x, objectid(x)] for x in split("hello\nhello", '\n')]
2-element Vector{Vector{Any}}:
 ["hello", 0x8096d78ab9fd726c]
 ["hello", 0x2bec5d6f24b3bdf5]

This patch changes the “is final item” check to be index-based, rather than identity-based.

I've tested it by adding a synthetic failure to an existing test:

@test throw(ErrorException("testing Buildkite JSON\n\ncomma after blank lines\n\nwhen last line is blank\n"))

And then observed the test/results_1.json before and after this patch:

       "expanded": [
         "testing Buildkite JSON",
-        ""
+        "",
         "comma after blank lines",
-        ""
+        "",
         "when last line is blank",
         ""
       ]

Related:

• Buildkite Test Analytics: fix failure_expanded #53706
• cc: @tecosaur for past context

[tecosaur]

Ah yep, this looks like a silly oversight in the original implementation. Thanks for contributing this fix Paul!