fix(handshake): add missing vllm_version field to EngineCoreReadyResponse payload

[iamagenius00]

Summary

The Rust vllm-rs frontend added in #457 expects 5 fields in the EngineCoreReadyResponse handshake payload, but engine_core_ready_payload in Python only emits 4 — missing vllm_version. As a result, any Mac/MLX worker crashes at handshake time with:

Error: failed to connect to engine core
Caused by:
    messagepack decode failed for vllm_engine_core_client::protocol::handshake::EngineCoreReadyResponse:
    missing field `vllm_version`;
    value fallback: {"max_model_len": 7168, "num_gpu_blocks": 0, "dp_stats_address": nil, "dtype": "bfloat16"}

This blocks all inference (/v1/chat/completions, /v1/completions, etc. all return 500) — the HTTP layer never gets a chance to route requests because the frontend ↔ engine IPC never establishes.

Root cause

The shipped vllm-rs binary's strings show:

struct EngineCoreReadyResponse with 5 elements
  max_model_len, num_gpu_blocks, dtype, vllm_version, dp_stats_address

But src/parallax/server/engine_core_protocol.py::engine_core_ready_payload only includes 4 of those, omitting vllm_version. The Python error fallback dump ({max_model_len, num_gpu_blocks, dp_stats_address, dtype}) confirms this exact set.

Fix

Add vllm_version as a keyword-only arg with default "0.0.0+parallax", and include it in the dict. The Rust side appears to only check field presence (no version comparison logic visible in the binary), so any string suffices; the "0.0.0+parallax" marker makes it clear in debug logs that this came from the Python MLX path rather than upstream vLLM. Callers can override if needed.

 def engine_core_ready_payload(
     *,
     max_model_len: int,
     dtype: Optional[str],
     num_gpu_blocks: int = 0,
     dp_stats_address: Optional[str] = None,
+    vllm_version: str = "0.0.0+parallax",
 ) -> bytes:
     """Build the registration payload sent by the engine DEALER socket."""
     return _pack_msgpack(
         {
             "max_model_len": int(max_model_len),
             "num_gpu_blocks": int(num_gpu_blocks),
             "dp_stats_address": dp_stats_address,
             "dtype": dtype,
+            "vllm_version": vllm_version,
         }
     )

Test plan

Single-node mac mini (M4 Pro, MLX backend), HEAD 328c99f with this patch applied:

1. parallax run -m Qwen/Qwen3-0.6B -n 1 (scheduler tab)
2. parallax join (worker tab)

Before patch:

• Worker logs missing field vllm_version and exits with code 1 shortly after model load
• curl localhost:3001/v1/chat/completions → internal server error (500)

After patch:

• Worker logs bootstrapped engines connected engine_count=1
• Worker logs starting OpenAI server bind_address=[::1]:3000 model=Qwen/Qwen3-0.6B
• curl localhost:3001/v1/chat/completions and curl localhost:3000/v1/chat/completions both return valid OpenAI-shaped JSON with model output:

  {"id":"chatcmpl-94ef2c77-...","model":"Qwen/Qwen3-0.6B",
   "choices":[{"index":0,"message":{"role":"assistant","reasoning":"..."}}],
   "usage":{"prompt_tokens":17,"total_tokens":67}}

Notes

• This affects every Mac/MLX user following the README quickstart on current main — it's not a configuration issue.
• Default "0.0.0+parallax" is semver-compatible (build-metadata suffix) and grep-able in logs.
• Happy to adjust the default string or to source the value from parallax.__version__ if maintainers prefer.