CompletionItem of completions response always have `start = 0` and text prefix inclusion is mixed

先使用報告中的 JDK 21 範例，追蹤 Java 偵錯伺服器對 completions 請求和 CompletionItem 回應的 DAP 補全處理。使用 nvim-dap 等用戶端重現 List. 和 com. 的情況，然後確認 start 和 text 遵循 DAP 規範，且選取每個候選項都會插入預期的補全，不會重複其前綴。

I noticed recently that in nvim-dap, if you complete com. and select an entry you get com.com... inserted, so today I took a closer look and noticed that the responses from java-debug are somewhat odd - and I think incorrect.

With a client that specified columnsStartAt1 = true, and a completions payload like:

{
    frameId = <frameId>,
    text = "List.",
    column = 6
}

The responses include:

  }, {
    label = "of(E e1, E e2, E e3, E e4) : List<E>",
    number = 0,
    sortText = "999999179",
    start = 0,
    text = "of()",
    type = "function"
  }, {

The specification says:

/**

• Start position (within the text attribute of the completions request)
• where the completion text is added. The position is measured in UTF-16 code
• units and the client capability columnsStartAt1 determines whether it is
• 0- or 1-based. If the start position is omitted the text is added at the
• location specified by the column attribute of the completions request.
  */
  start?: number;

The expected result for the user is to have List.of() if the completion candidate is selected. Now, start=0 is already odd given the columnsStartAt1, so a possible interpretation in the client is that it's absent, and that the client should just append .of()

This is kinda what I did in nvim-dap so far, and it works for the List.of case, and also for variables, but with a payload like:

{
  column = 5,
  frameId = <frameId>,
  text = "com."
}

I get responses like:

  }, {
    label = "com.sun.tools.example",
    number = 0,
    sortText = "999999183",
    start = 0,
    text = "com.sun.tools.example",
    type = "module"
  }, {

Opposed to the List. result, here text includes the prefix com. and it's again start=0. This led to com.com.sun.tools.example

I suspect vscode does some kind of prefix matching on the client side again, so this isn't noticable there?
As far as I can tell, based on the specification the current behavior is wrong.

I used JDK 21 in my tests - in case it matters.
I can also provide some sample project if needed - but I tried to use examples that should behave similar with only the JDK as dependency