D055 - The Linux path was broken, and only running it showed that

decided · 2026-08-19

D027 says neither platform ships an unverified implementation with a claim attached. Windows passed all four reservation tests; Linux failed all four, and would have shipped with a false claim had it not been run in the multipass VM.

Two bugs, the second worse than the first:

• MAP_PRIVATE was missing. An mmap with neither PRIVATE nor SHARED is EINVAL, so every reservation failed. A pure oversight, invisible on Windows because that path is separate code.
• Every mmap error was reported as Conflict. So the EINVAL surfaced as "range taken", sending a reader hunting for a phantom occupant instead of at the wrong argument that actually caused it. Errors are now distinguished: EEXIST and ENOMEM are conflicts, everything else reports the errno.

The second is the one worth remembering. A wrong error message is worse than no error message, because it is actively misleading - the same failure D010 exists to prevent, arriving through error mapping rather than through a stub.

Verified: 8/8 on Linux and 8/8 on Windows, same source.

Method note for future platform work. Only orbistoun-mem was built in the VM, with CARGO_TARGET_DIR pointed at VM-local storage so it could not collide with the Windows build directory over the mount. The VM was unmounted and its build artifacts removed afterwards. It is a working CI runner, not a scratch box.