[ashhart]

On Linux, core/direct_io's O_DIRECT read of a file that ends inside the aligned span it asked for retries from the last 4 KiB boundary, reads the same tail again, makes no progress, and reports ReadFailed. It is the end of the file, and callers handle that: Nemotron's CUDA checkpoint reads stop on EndOfFile, and DeepSeek's pack reader (in #1924) names it ShortRead. Nothing was copied wrongly and nothing hung before; the error was wrong. macOS has no O_DIRECT, so a Mac never took this path.

What changes

• core/direct_io.readAt: an aligned retry that makes no progress returns error.EndOfFile. Only an aligned retry can make no progress (an unaligned read continues from got), so no other case changes. A unit test pins it with a scripted pread.
• families/nemotron/cuda_source_test.zig (new): Nemotron's CUDA checkpoint reader (cuda_source.zig), which no test reached, against a fake CUDA driver that logs each host-to-device copy and stream wait. Two shards' tensors at unaligned offsets and sizes, one split over two slot reads, more uploads than slots, through both the io_uring ring and the thread fallback; every device byte must be the file's. Each test asserts which path ran and skips with a printed reason where io_uring or O_DIRECT is refused. A third test shortens a shard after open so a tensor runs past its end, and wants EndOfFile with nothing copied, on every path and read mode. It runs in test-cuda-host (and zig build test on Linux and Windows).

Checks (Linux aarch64, ext4)

• Main plus only the new test: the shortened-shard test fails on both paths, "io_uring reads through O_DIRECT gave ReadFailed past the shard's end, want EndOfFile" and the same for threads.
• This branch: it passes, and core/direct_io's tests pass (5/5, including the two that only run on Linux).
• test-cuda-host otherwise: the only failures are two drama_voice_tool directory-permission tests, which fail identically on main there because that machine runs tests as root.
• Native Windows (x86_64): test-cuda-host 335/335 steps, 0 failed. The thread-path test passes through unbuffered reads (FILE_FLAG_NO_BUFFERING, the aligned path), the io_uring test skips (no io_uring), and the shortened-shard test skips (Windows can't shorten a mapped file), the same on main plus the test. The full zig build test shows no new failures.
• On a Mac the Linux paths skip, as expected; zig fmt --check clean.