serve — Streaming response always returns empty finish_reason instead of null/"stop", breaking OpenAI-compatible clients

[Novatoris]

Describe the bug
When using geniex serve with streaming enabled ("stream": true), all chunks in the streaming response have "finish_reason": "" (empty string) including the final chunk. According to the OpenAI streaming specification, finish_reason should be null on intermediate chunks and "stop" on the final chunk. This causes OpenAI-compatible clients such as AnythingLLM to never detect that generation has finished, requiring the user to manually stop generation.

Non-streaming responses ("stream": false) correctly return "finish_reason": "stop".

To Reproduce

1. Start server: geniex serve --host 127.0.0.1:18181 --compute npu
2. Send a streaming request:

Invoke-WebRequest -Uri "http://127.0.0.1:18181/v1/chat/completions" -Method Post -ContentType "application/json" -UseBasicParsing -Body '{"model": "qualcomm/Qwen3-4B-Instruct-2507", "messages": [{"role": "user", "content": "Say hello"}], "stream": true}' | Select-Object -ExpandProperty Content

3. Observe all chunks have "finish_reason": "" including the final chunk before data: [DONE]

Expected behavior
Per the OpenAI streaming specification:

• Intermediate chunks: "finish_reason": null
• Final chunk: "finish_reason": "stop" with empty "delta": {}

GenieX currently sends "finish_reason": "" (empty string) on all chunks including the final one, which is non-compliant and causes clients to never detect completion.

Workaround
Manually stop generation in the UI once the response appears complete. In AnythingLLM, click the stop button after the model finishes generating text — the response is complete but the client keeps waiting due to the missing finish_reason: stop signal.

Desktop:

• OS: Windows 11 ARM64
• Device: Snapdragon X Plus X1P42100
• GenieX CLI Version: v0.3.17
• QAIRT Runtime Version: v2.45.0.260326
• LlamaCPP Runtime Hash: ae9291e
• Model: ai-hub-models/Qwen3-4B-Instruct-2507 (W4A16, QAIRT runtime)
• AnythingLLM Version: v1.15.0 (used for OpenAI-compatible client testing)

[mansiverma897993]

Novatoris I've root-caused this and opened PR #1246.

The server built streaming chunks from openai-go's response structs, whose FinishReason is a plain string so every chunk serialized as "finish_reason": "" and no terminal chunk was ever sent before [DONE]. The PR makes intermediate chunks send "finish_reason": null and ends every stream with a final empty-delta chunk carrying "stop" (or "length" / "tool_calls"), exactly per the OpenAI spec.

Verified: unit tests + the exact JSON output against openai-go v3.17.0, e.g. final chunk:
{"object":"chat.completion.chunk","choices":[{"index":0,"delta":{},"finish_reason":"stop"}]}

Not verified: I don't have a Snapdragon device, so I couldn't run geniex serve end-to-end with AnythingLLM. Could you test with a build of this patch and confirm it unblocks your client? Alan Zhu (@alanzhuly) Mengsheng Wu (@mengshengwu)