fix: testcase failed ococcasionally

[IceCodeNew]

No description provided.

[jpillora]

Thanks for catching this — the flake was real: linux/amd64 resolved to yt-dlp_linux.zip (glibc) or yt-dlp_musllinux.zip (musl) at random, roughly 1 in 7 runs.

Rather than relaxing the assertion, I fixed the root cause in the handler in 2028bfe.

Root cause: "assumed" assets (filenames with no explicit os/arch token) are collected into a map in getAssetsNoCache and then merged into the result via for _, cAsset := range candidates — a randomized Go map iteration. When two assumed assets collapse to the same os/arch slot (glibc yt-dlp_linux.zip → linux/amd64 and musl yt-dlp_musllinux.zip → /amd64, both normalizing to linux/amd64), whichever the map visited first won. Fixed input, random output.

Fix: candidates are now resolved through a total order before gap-filling the index:

• Asset.preferredOver — prefer musl over glibc (portability), then avoid gnu, then break ties by name. Because it's a total order, the winner is independent of map iteration order.
• exact os/arch matches still beat assumed assets, so e.g. linux/arm64 stays yt-dlp_linux_aarch64.

Added TestAssumedAssetDeterministic (100 iterations/run) guarding both invariants; TestYtDlp is now stable over 25+ runs and the strict linux/amd64 → yt-dlp_musllinux.zip assertion is kept, so it continues to guard the musl preference.

Since the handler is deterministic now, this test change isn't needed — closing. Thanks again for the report and the clear repro!