Kristofer Karlsson was checking which CI tests take the most time and found that t5608 does not create the >2GB pack its name promises. The setup writes only 24 objects, below the default fastimport.unpackLimit of 100, so fast-import turns the pack into loose objects. He traces the breakage to d9545c7f46 (fast-import: implement unpack limit, 2016-04-25). The test's original comment had anticipated that the setup might stop working.
While adding assertions, he found that the unpack-objects variant of the >4GB test never runs unpack-objects. Once it does, it fails on Windows because stream_blob() takes the blob size as "unsigned long", which is 32 bits there. The size is truncated and writing fails with "fatal: write stream object 4294967304 != 8". The first patch passes the size as size_t, following f2063855fb (index-pack, unpack-objects: use size_t for object size, 2026-05-08).
Patrick Steinhardt agreed the test had never exercised what it was designed to. He asked whether the first patch fixes an assertion that would otherwise fail, and noted as a nit, not worth a reroll, that he would have folded a helper into the test itself. He was also surprised that one behavior applies to only one command rather than both git-clone and git-fetch, and asked whether that is intentional or another bug.
Karlsson said it appears to be intentional and old, and that he would inline the helper if a reroll is needed. Steinhardt otherwise said the series looks good.