test(graalpy): three failures on the GraalPy experimental leg after #2105 and #2111 #2134

main's Python graalpy-25.4 (experimental) leg warns at 8fdc2fb33 (run 40052089224, reported in the experimental: main push runs thread). This is the first GraalPy run to finish since #2105 and #2111 merged; earlier ones were superseded. Three tests fail:

• tests/corekit/test_io_buffer_save_tempfile_2098_unit.py::TestBufferFile::test_collection_removes_the_temporary_file_of_an_unclosed_reader, from fix(corekit): let Windows reopen SeekableReader's buffer file (#2098) #2105, with AssertionError: True is not false. The temporary file survives "collection" of an unclosed reader. GraalPy's GC is not reference-counting, so this is probably a test that assumes CPython's immediate finalisation. Unconfirmed.
• tests/protocols/misc/test_fromtimestamp_windows_range_2099_unit.py::TestFromTimestamp::test_errors_are_cpythons, from fix(protocols): convert capture timestamps without the platform's fromtimestamp limits (#2099) #2111, with ValueError: year 0 is out of range. The expected error classes may be CPython-specific. Unconfirmed.
• …::TestMHOnWindows::test_mesg_id_before_1970, from fix(protocols): convert capture timestamps without the platform's fromtimestamp limits (#2099) #2111, subtest interval=1950-06-01 12:00:00.500000. It got 12:00:01.500000. That is a one-second difference for a pre-1970 instant with a fraction. The cause could be our fromtimestamp's split of a negative float, or GraalPy's own datetime.fromtimestamp or math.modf, if the test compares against it. This one may be a real defect.

Wanted: root-cause each on GraalPy itself, fix any library defect, and gate or adapt tests that encode CPython-only behaviour. Experimental legs must stay green with warnings, and this one already does, so these failures are the defect.