[emeryberger]

Follow-up to #129 (now merged), rebased onto master.

Summary

• End-to-end regression test, //src:high-offset-mesh-test. Reserves an untouched 4 GiB placeholder so the next two spans start more than 4 GiB into the backing file, meshes them, then scavenges. It checks that:

  • the meshed object keeps its contents (on current master it reads back as an empty string instead of its ZZZ… contents);
  • after resetSpanMapping, writes to the old span don't show up in the placeholder. This check also fails on its own when only resetSpanMapping has the old 32-bit shift.

  It's a separate target because of the 4 GiB address-space reservation, which uses almost no real memory. make test and both Bazel CI jobs run it alongside unit-tests. The CMake job doesn't run it.

• Cleanup. Merges the __APPLE__ and Linux branches of finalizeMesh, which were identical apart from hard_assert(_fd >= 0); that assert now runs on all platforms. Also widens span.length before the shift in the MADV_DONTDUMP call, which is dead code today since kAdviseDump = false.

Testing

• CI (run 39354360163): all jobs pass. The new test passes on Linux (4K pages) and macOS (16K pages), in both debug and optimized builds.
• Locally (macOS arm64): passes with the fix; fails at the s2 content check with master's meshable_arena.{h,cc}; fails at the alias check when only resetSpanMapping has the old shift.

Questions for review

• Is a 4 GiB arena reservation acceptable on all CI and developer machines? It needs kArenaSize > 4 GiB plus the spans after it. That holds for 64 GB on Linux and 32 GB on macOS.
• Is it OK for the test to write to freed arena pages after scavenge (it zeroes them again afterward)?

🤖 Generated with Claude Code