[Bug] OpenAI-compatible streaming tool_calls omit function.name after first chunk, breaking OpenCode with Qwen3.6-27B on llm-scaler-vllm 0.14.0-b8.3.2

## Summary

Using `intel/llm-scaler-vllm:0.14.0-b8.3.2` with `Qwen3.6-27B` on a single Intel Arc Pro B70, streaming tool calls from the OpenAI-compatible `/v1/chat/completions` endpoint are emitted in a way that breaks strict OpenAI-compatible clients such as OpenCode Desktop v1.17.15.

The first `tool_calls` delta includes `id`, `type`, and `function.name`, but follow-up deltas only include fragments of `function.arguments` and do not re-emit `function.name`. In some parser configurations I also observed `function.name: null` in follow-up chunks.

This causes OpenCode to fail intermittently with:

`Expected 'function.name' to be a string`

The issue is intermittent in OpenCode because it only happens on turns where the model actually emits streamed tool calls in multiple chunks.

## Environment

- Hardware:
- 1 × Intel Arc Pro B70 32GB
- Ryzen 7 9700X
- 64GB DDR5 RAM
- OS:
- Ubuntu 26.04
- Container / build:
- `intel/llm-scaler-vllm:0.14.0-b8.3.2`
- Model:
- `/llm/models/Qwen3.6-27B-int4-AutoRound`
- served as `Qwen3.6-27B`
- Client:
- OpenCode Desktop v1.17.15
- Reproducible both with:
- `--tool-call-parser qwen3_xml`
- `--tool-call-parser qwen3_coder`

## Server launch command

```bash
vllm serve /llm/models/Qwen3.6-27B-int4-AutoRound \
--served-model-name Qwen3.6-27B \
--port 8000 \
--host 0.0.0.0 \
--dtype float16 \
--enforce-eager \
--trust-remote-code \
--gpu-memory-util 0.94 \
--max-model-len 245760 \
--max-num-batched-tokens 8192 \
--swap-space 16 \
--block-size 32 \
--max-num-seqs 2 \
--allow-deprecated-quantization \
--reasoning-parser qwen3 \
--enable-auto-tool-choice \
--tool-call-parser qwen3_xml \
--chat-template /llm/models/qwen3.6_chat_template.jinja \
--default-chat-template-kwargs '{"max_tool_response_chars": 15000, "max_tool_arg_chars": 2000}'
```

I also tested `--tool-call-parser qwen3_coder` and saw the same root problem, but with follow-up chunks omitting `function.name` instead of sending `"name": null`.

## Reproduction

```bash
curl -N http://127.0.0.1:8000/v1/chat/completions \
-H "Content-Type: application/json" \
-d '{
"model": "Qwen3.6-27B",
"stream": true,
"messages": [
{
"role": "user",
"content": "Lee el archivo api/informes.php usando una herramienta."
}
],
"tools": [
{
"type": "function",
"function": {
"name": "read_file",
"description": "Lee un archivo",
"parameters": {
"type": "object",
"properties": {
"path": { "type": "string" }
},
"required": ["path"]
}
}
}
],
"tool_choice": "auto",
"temperature": 0
}'
```

## Observed streaming output with `--tool-call-parser qwen3_coder`

### First tool-call delta

```json
data: {"choices":[{"index":0,"delta":{"reasoning_content":null,"tool_calls":[{"id":"call_7b2e5a158a144cb4b6930b67","type":"function","index":0,"function":{"name":"read_file","arguments":""}}]},"finish_reason":null}]}
```

### Follow-up deltas only carry argument fragments and do not re-emit `function.name`

```json
data: {"choices":[{"index":0,"delta":{"reasoning_content":null,"tool_calls":[{"index":0,"function":{"arguments":"{"}}]},"finish_reason":null}]}

data: {"choices":[{"index":0,"delta":{"reasoning_content":null,"tool_calls":[{"index":0,"function":{"arguments":"\"path\": \"api/informes.php\""}}]},"finish_reason":null}]}

data: {"choices":[{"index":0,"delta":{"reasoning_content":null,"tool_calls":[{"index":0,"function":{"arguments":"}"}}]},"finish_reason":null}]}
```

### Final chunk

```json
data: {"choices":[{"index":0,"delta":{"content":"","reasoning_content":null},"finish_reason":"tool_calls"}]}
data: [DONE]
```

## Observed streaming output with `--tool-call-parser qwen3_xml`

In that configuration, follow-up deltas included:

```json
"function":{"name":null,"arguments":"..."}
```

instead of omitting `name`.

## Expected behavior

For better compatibility with strict OpenAI-compatible clients, follow-up `tool_calls` deltas should continue to include:

- `id`
- `type`
- `function.name`

on every tool-call delta for the same `index`, not only on the first one.

## Actual behavior

The server emits only the initial `function.name` and later emits partial argument deltas without `function.name` (or with `name: null` depending on parser). OpenCode Desktop v1.17.15 then fails with:

`Expected 'function.name' to be a string`

## Additional notes

- This is not a pure OpenCode-side hallucination issue: the broken chunk pattern is directly observable with `curl`.
- The problem is intermittent in OpenCode because it only happens when a tool call is streamed in multiple chunks.
- Switching from `qwen3_xml` to `qwen3_coder` changes the exact chunk shape, but does not solve the compatibility issue.