- Fatbin parser: stop when an entry header runs past the declared files
size instead of wrapping the unsigned remainder and reading outside the
image.
- Output kernel table: keep the latest 64 lookups and resolve launches
newest first under its own lock, so a full table no longer drops new
lookups and a handle reused by a later output-kernel lookup resolves
with that lookup.
- Resolve a relative D4R_ZLUDA_NATIVE_DIR before linking its kernels into
the per-process directory; the relative links dangled.
- cuSurfObjectCreate: destroy the surface and return the error when its
output redirect cannot be written.
- Serve native kernels from D4R_ZLUDA_CACHE_HOME, XDG_CACHE_HOME or
~/.cache, never /tmp, where the stale-PID cleanup could follow links
planted by another user.
