# ollama 0.33.3: tool-call requests that don't match the PEG grammar are cancelled with no `finish_reason`

## Environment
- Package: `ollama 0.33.3-1` (+ `ollama-rocm 0.33.3-1`) from the `[omarchy]` stable repo (`pkgs.omarchy.org/stable`)
- Hardware: AMD RX 7900 XTX (ROCm)
- Model: `gpt-oss-20b:q4`
- Client: any OpenAI-compatible `/v1/chat/completions` consumer (reproduced with Hermes Agent)

## Symptom
When a chat request with `tools` produces tool-call output that the `peg-native` PEG parser does not recognize, the server:

1. logs `unparsed peg-native output`
2. logs `srv stop: cancel task`
3. **closes the SSE stream with no `finish_reason` and no content**

The client therefore sees a stream that simply ends mid-request. Consequence for OpenAI-compatible clients: the missing `finish_reason` is classified as a truncated/partial response, triggering continuation retries that eventually crash the worker.

## Root cause (from binary string analysis)
The `peg-native` grammar expects tool calls framed as:

```
<|channel|>(commentary|analysis) to=functions<|message|>{ ... }
```

but `gpt-oss-20b` emits:

```
<|channel|>functions.<name><|channel|>commentary json<|message|>{ ... }
```

On the mismatch, `common_chat_peg_parse` fails and the task is cancelled instead of the content (or a `finish_reason`) being returned. Notably, `gpt-oss` is not in the fork's known-grammar family (cogito, gemma4, qwen3.5, olmo3, glm-4.6, cohere, lfm2, deepseek, peg-simple, peg-gemma4, peg-minimax-m3), so it falls through to `peg-native` and fails.

## Why this matters
A PEG parse failure should not silently kill the request. For a model the distro ships, the tool-calling path is unusable. This also affects any model whose tool framing doesn't match a known grammar.

## Requested fix
On PEG parse failure, one of:
- **(a)** pass the raw tool-call content through to the client (best-effort), or
- **(b)** emit a `finish_reason` (`stop` or `tool_calls`) with the content,

rather than cancelling the task and closing the stream with no `finish_reason`.

## Reproduction
`POST /v1/chat/completions` with `gpt-oss-20b:q4`, `stream: true`, and a `tools` array. Prompt the model to call a tool (e.g. a `terminal` tool). The stream ends with no `finish_reason` and no content; the server log shows `unparsed peg-native output` + `cancel task`.