[Bug]: Context compression accepts embedding-only models and silently stops working

### Issue Checklist

- [x] I searched existing open and closed Issues and did not find the same behavior.
- [x] This report describes one reproducible defect.

### Platform

Windows

### Version

2.0.12

### Bug Description

The context-compression model selector accepts a model whose only capability is `embedding`. When compression is evaluated, `resolveCompressionModel` rejects that provider/model combination and skips compression. The failure is written only as a warning, so the conversation continues while automatic context compression silently stops working.

### Steps To Reproduce

1. Configure a local embedding-only model such as `qwen3-embedding-0.6b`.
2. Select that model as the context-compression model.
3. Use a conversation long enough to trigger compression evaluation.
4. Observe `compression model resolution failed` in the logs while the UI shows no error and compression is skipped.

### Expected Behavior

The selector should filter models by generation capability, or reject an unsupported selection with an actionable UI message. A compression failure must not remain invisible.

### Diagnostic Evidence

- Diagnostic ID: `bfbea426-00a1-442b-8048-5d88137ac1b7`
- [Open the diagnostic report](https://api.cherry-ai.com/diagnostics/bfbea426-00a1-442b-8048-5d88137ac1b7)
- [Raw diagnostic package (core maintainers only)](https://mcnnox2fhjfq.feishu.cn/file/CjDsbnvsOoSI35xcPqZcmxLDnlh)

### Related Issues

#18737 concerns a provider factory failure. #19651 concerns exhausted credit during compression. Neither covers selecting an embedding-only model or the silent failure state.