[BinaryPlatforms] Change "shortest match" algorithm to "best match"

My assertion in the previous attempt to fix this issue was incorrect:

> We define a simpler match as one that has fewer tags overall.
> As these candidate matches have already been filtered to match the
> given platform, the only other tags that exist are ones that are in
> addition to the tags declared by the platform.  Hence, selecting the
> minimum in number of tags is equivalent to selecting the closest match.

This is demonstrably false, by my own test case:

```
platforms = Dict(
    Platform("x86_64", "linux") => "bad",
    Platform("x86_64", "linux"; sanitize="memory") => "good",
)
select_platform(platforms, Platform("x86_64", "linux"; sanitize="memory")) == "good"
```

In this case, because there exists a candidate that is _more general_
than the provided platform type, the shortest match is no longer the
best match.

This PR performs a more rigorous matching that works more reliably in
all cases.

base/binaryplatforms.jl

@@ -1068,21 +1068,28 @@ function select_platform(download_info::Dict, platform::AbstractPlatform = HostP
     end
 
     # At this point, we may have multiple possibilities.  We now engage a multi-
-    # stage selection algorithm, where we first choose simpler matches over more
-    # complex matches.  We define a simpler match as one that has fewer tags
-    # overall.  As these candidate matches have already been filtered to match
-    # the given platform, the only other tags that exist are ones that are in
-    # addition to the tags declared by the platform. Hence, selecting the
-    # minimum in number of tags is equivalent to selecting the closest match.
-    min_tag_count = minimum(length(tags(p)) for p in ps)
-    filter!(p -> length(tags(p)) == min_tag_count, ps)
-
-    # Now we _still_ may continue to have multiple matches, so we now simply sort
-    # the candidate matches by their triplets and take the last one, so as to
-    # generally choose the latest release (e.g. a `libgfortran5` tarball over a
-    # `libgfortran3` tarball).
-    p = last(sort(ps, by = p -> triplet(p)))
-    return download_info[p]
+    # stage selection algorithm, where we first sort the matches by how complete
+    # the match is, e.g. preferring matches where the intersection of tags is
+    # equal to the union of the tags:
+    function match_loss(a, b)
+        a_tags = Set(keys(tags(a)))
+        b_tags = Set(keys(tags(b)))
+        return length(union(a_tags, b_tags)) - length(intersect(a_tags, b_tags))
+    end
+
+    # We prefer these better matches, and secondarily reverse-sort by triplet so
+    # as to generally choose the latest release (e.g. a `libgfortran5` tarball
+    # over a `libgfortran3` tarball).
+    ps = sort(ps, lt = (a, b) -> begin
+        loss_a = match_loss(a, platform)
+        loss_b = match_loss(b, platform)
+        if loss_a != loss_b
+            return loss_a < loss_b
+        end
+        return triplet(a) > triplet(b)
+    end)
+
+    return download_info[first(ps)]
 end
 
 # precompiles to reduce latency (see https://github.com/JuliaLang/julia/pull/43990#issuecomment-1025692379)

test/binaryplatforms.jl

@@ -330,8 +330,7 @@ end
     # Ambiguity test
     @test select_platform(platforms, P("aarch64", "linux")) == "linux3"
     @test select_platform(platforms, P("aarch64", "linux"; libgfortran_version=v"3")) == "linux3"
-    # This one may be surprising, but we still match `linux3`, and since linux3 is shorter, we choose it.
-    @test select_platform(platforms, P("aarch64", "linux"; libgfortran_version=v"3", libstdcxx_version=v"3.4.18")) === "linux3"
+    @test select_platform(platforms, P("aarch64", "linux"; libgfortran_version=v"3", libstdcxx_version=v"3.4.18")) === "linux5"
     @test select_platform(platforms, P("aarch64", "linux"; libgfortran_version=v"4")) === nothing
 
     @test select_platform(platforms, P("x86_64", "macos")) == "mac4"
@@ -343,13 +342,21 @@ end
     # Sorry, Alex. ;)
     @test select_platform(platforms, P("x86_64", "freebsd")) === nothing
 
-    # The new "prefer shortest matching" algorithm is meant to be used to resolve ambiguities such as the following:
+    # The new "most complete match" algorithm deals with ambiguities as follows:
     platforms = Dict(
-        # Typical binning test
-        P("x86_64", "linux") => "good",
-        P("x86_64", "linux"; sanitize="memory") => "bad",
+        P("x86_64", "linux") => "normal",
+        P("x86_64", "linux"; sanitize="memory") => "sanitized",
+    )
+    @test select_platform(platforms, P("x86_64", "linux")) == "normal"
+    @test select_platform(platforms, P("x86_64", "linux"; sanitize="memory")) == "sanitized"
+
+    # Ties are broken by reverse-sorting by triplet:
+    platforms = Dict(
+        P("x86_64", "linux"; libgfortran_version=v"3") => "libgfortran3",
+        P("x86_64", "linux"; libgfortran_version=v"4") => "libgfortran4",
     )
-    @test select_platform(platforms, P("x86_64", "linux")) == "good"
+    @test select_platform(platforms, P("x86_64", "linux")) == "libgfortran4"
+    @test select_platform(platforms, P("x86_64", "linux"; libgfortran_version=v"3")) == "libgfortran3"
 end
 
 @testset "Custom comparators" begin