fix(server): close mkstemp fd in _tokenize_prompt to prevent fd leak

[sanastasiou]

The bug

In dflash/scripts/server.py, _tokenize_prompt() calls:

tmp = Path(tempfile.mkstemp(suffix=".bin")[1])

tempfile.mkstemp() returns (fd, path). This code keeps only the path (index [1]) and discards fd. Python's GC does not close integer file descriptors, so the fd stays open in the OS until the process dies. Every chat request leaks exactly one fd.

Under sustained load the process hits RLIMIT_NOFILE (default 1024 on a systemd service without explicit LimitNOFILE=). Once the wall is reached, socket.accept() starts throwing OSError: [Errno 24] Too many open files and the API stops accepting new HTTP connections — even though the model process itself is still healthy.

Reproduction

In our deployment we have Lucebox sitting behind ~6 concurrent client services (4 Hermes agents + a Hindsight memory server + a Honcho peer-model service + a Firecrawl extractor, all OpenAI-compatible). Within roughly an hour of normal aux traffic, FD count reached 1023/1024 and new requests were rejected.

/proc/<lucebox-pid>/fd showed the composition: only ~5 actual sockets and pipes, and ~1000 orphan /tmp/tmp*.bin file descriptors — exactly the pattern of a leaked mkstemp fd.

Journal log excerpt:

OSError: [Errno 24] Too many open files
socket.accept() out of system resource
socket: <asyncio.TransportSocket fd=8, family=2, type=1, proto=6, laddr=('0.0.0.0', 8001)>

The fix

Use os.fdopen(fd, "wb") which takes ownership of fd and closes it when the with block exits:

fd, path = tempfile.mkstemp(suffix=".bin")
tmp = Path(path)
with os.fdopen(fd, "wb") as f:
    for t in ids:
        f.write(struct.pack("<i", int(t)))

The on-disk temp file is still cleaned up by the existing prompt_bin.unlink() in chat_completions (both streaming and non-streaming paths), so there is no change to disk behaviour — only the fd lifecycle changes.

Verification

• Before fix: FD count delta across 3 probe chat requests = +3 (one fd leaked per request)
• After fix: FD count delta across 3 probe chat requests = 0 ✓

Also raised LimitNOFILE from the default 1024 to 65536 on our systemd unit as a defence-in-depth measure (unrelated to this PR). The code change alone is the correct fix; the limit bump is independent and up to the operator.