Disk tier server (read_write mode).

One server per disk root directory. Owns no ETS tables; the disk itself is the source of truth for slabs. Reads, writes, and deletes all funnel through the gen_server, which keeps file ordering sequential per directory.

Save pipeline (save/3):

1. Build framed bytes via erllama_cache_kvc:build/2.
2. Open writer-unique temp file <hex>.kvc.<writer_id>.tmp with O_EXCL.
3. prim_file:write/2 an iolist of [Prefix, Payload]. Multi-GB payloads are not concatenated in BEAM memory; the IO subsystem uses writev under the hood.
4. prim_file:datasync/1, then close.
5. prim_file:make_link/2 to publish at <hex>.kvc. EEXIST is handled: if the existing file is a valid KVC for this Key, we adopt it (delete our temp). Otherwise we delete the corrupt file and retry once.
6. erllama_nif:fsync_dir/1 on the root.
7. Reopen the published file and parse it (validation belt-and- braces against FS bugs).

Returns {ok, Header, Size} on success. The caller is responsible for the meta_srv:reserve_save -> check_reservation -> mark_published -> announce_saved protocol around this call (the writer pool in step 10 wires that up).

On startup, the server scans its directory: deletes any *.tmp files (interrupted writes; safe to drop), parses every <hex>.kvc header, and registers each valid file with the meta server. Files that fail to parse are deleted.