Performance of code completion got drastically worse some time in the last month or two

Problems:

- low project context awareness, completions just seem to copy and paste code verbatim from related places, instead of adapting to the local context. I even pointed in comments to the exact file names of where to find related code and how to adapt.
- in particular, performance on filetype `bash` has dropped very strongly. It's basically lobotomized as far as my experience goes.

Just some feedback; maybe someone cares enough for paying subscribers to gather this kind of feedback. I'll be looking for a replacement now asap as it's very disappointing to see performance nosedive that hard. It's really a different world. Feels like the model was severely overfit to a too small scope of use cases.

沒有指定檔案、測試或進入點，而且報告提供的是回饋，而不是可重現的失敗。先確認補全回歸是否仍然可以重現，尤其是針對 bash 檔案；只有在範圍達成共識、存在可重現的案例，並且效能改善已獲得驗證後，才能算完成。

由索引模型根據 Issue 內容生成。