Story: Hotfix: two macOS test failures that assert Linux-only facts

With the macOS build compiling again, its test phase runs and fails in two places, both of them facts about Linux that the tests assert rather than defects in the product:

• filesystem_scoped_temp_file_tests.cpp compares sut.path().parent_path() with std::filesystem::temp_directory_path() as strings. macOS returns the system temp directory with a trailing separator – confstr's _CS_DARWIN_USER_TEMP_DIR ends in one – so the same directory has two spellings and the comparison fails: "/var/folders/.../T" == "/var/folders/.../T/".
• systemd_notify_tests.cpp binds an abstract unix datagram socket, whose name begins with a NUL byte, to prove that systemd's leading @ is not a literal file name. The abstraction is a Linux extension; macOS has pathname sockets only, so bind fails with No such file or directory and takes the whole ores.service.tests binary with it.

Neither failure can be fixed by changing the product: the temp file is already created in the directory the platform names, and systemd is Linux-only. The assertions are what move.

• The two temp-directory assertions pass for both spellings of the temp directory, and still assert that the file and the directory are the platform's temp directory.
• The abstract-socket case is skipped on platforms without an abstract namespace, with a message that says why, and still runs on Linux.
• ores.platform.tests and ores.service.tests pass on Linux, and the macOS job's test phase gets past both.

• The temp assertions compare directories rather than strings (std::filesystem::equivalent), not normalised spellings. lexically_normal() was tried first and does not strip a trailing separator – path("/tmp/").lexically_normal() is still "/tmp/" – so the assertion now states what it means: the file landed in the system temp directory, however the platform spells it. It also survives macOS resolving /var through /private/var, which a string comparison would not.
• The abstract-socket case is skipped, not rewritten to a pathname socket. The case exists to pin the @ to NUL translation, which only systemd needs and only Linux can express; a pathname socket would test a different thing and pass while proving nothing.
• Both fixes are in tests. Nothing in the product reads a path as a string or assumes an abstract namespace, and the scan recorded on the task found no other test that does either.

• The remaining macOS-only facilities the tests skip by design (the /sys/class/net cross-checks, for instance). They already skip with a message.