Completion for errors

[psznm]

This adds completion for errors.

There are issues demonstrated by tests

Issue test 11

"switch on error set - 11" is catch |err| switch (err) construct.

I suspect that scope detection might be wrong which is why "switch on error set - 11" does not work. innermostScopeAtIndex in lookupSymbolGlobal returns scope that looks garbage when printed out.

This:

    var current_scope = innermostScopeAtIndex(document_scope, source_index);
    const node = document_scope.getScopeAstNode(current_scope);
    if (node) |nod| {
        const src = handle.tree.getNodeSource(nod);
        std.debug.print("\nSRC=---\n{s}\n---\n", .{src});
    }

Prints out

       SRC=---
       {
         idk

Trying to tackle the issue is intimidating to me as I don't even know where I would start. Maybe someone more experienced could give some leads how to solve the issue?

---

Issue test 13

"switch on error set - 13" is completion in a function in a structure.

This also feels like issue with scopes. I started digging into how scopes are created. I found DocumentScope where it walks the document. It seems like structure declarations are completely ignored, seemingly on the std side, because the walkContainerDecl just walks the const std = @import("std") and then just stops. When I print out container_decl.ast.members.len , it gives 1. When I delete the std import, there is 0. Nothing gets walked.

---

After some more digging, it seems both issues are basically the same one. The 13 is structure member that is not given to zls from std. The 11 seems like a block statement that is not given to zls from std (the walker sees test_decl and its block, but nothing else). Is there any chance of fixing this from zls side, or does this need to be fixed on std side?

EDIT: Okay, I found out that the Ast parser is not error-prone enough for zls. Both issues are caused by that the parser refuses to give us these specific members/blocks that contain invalid syntax. So this is to be improved on the stdlib side. But here we can at least add completions for "empty" context to auto-complete errors. Those should work with both of the problematic cases, and honestly, even be the most ergonomic completions. I'll see if I can do that.