Deterministic bounce oracle: pread the record (no O_DIRECT) into a host
buffer, copy_h2d into a per-slot device buffer, stream-synced. This is the
reference every other tier must match byte-for-byte.
The zero-copy path: O_DIRECT the record straight into the pinned arena slot;
the returned addresses point INTO the arena (GPU-addressable at the same
VA), no copy_h2d.