fix(parser): don't pop from lexer errors to preserve error after rewind

[ulrichstark]

Closes #12816 and leads to more and better error messages as you can see in the updated coverage snapshots.
I'm not sure why the size of antd.js suddenly changed and I had to update the allocation snaphots as a result.

[Copilot]

Pull Request Overview

This PR fixes an issue where lexer errors were being lost after parser rewind operations by changing from popping errors to referencing the last error. This ensures error messages are preserved and available even after the parser backtracks, leading to better error reporting.

• Changes error handling from destructive pop() to non-destructive last() access
• Adds error cloning to maintain ownership semantics
• Preserves lexer errors for better diagnostic reporting after parser rewind operations

---

_{Tip: Customize your code reviews with copilot-instructions.md. Create the file or learn how to get started.}

[codspeed-hq]

CodSpeed Instrumentation Performance Report

Merging #13460 will not alter performance

_{Comparing ulrichstark:dont-pop-from-lexer-errors (09bc1b6) with main (b92fa4f)}

Summary

✅ 37 untouched benchmarks

[Boshen]

This is producing two identical errors:

e.g.:

  × Unterminated string
    ╭─[test262/test/language/types/string/S8.4_A14_T1.js:16:11]
 15 │ 
 16 │ var str = ";
    ·           ───
    ╰────

  × Unterminated string
    ╭─[test262/test/language/types/string/S8.4_A14_T1.js:16:11]
 15 │ 
 16 │ var str = ";
    ·           ───
    ╰────

[ulrichstark]

This is producing two identical errors:

e.g.:

  × Unterminated string
    ╭─[test262/test/language/types/string/S8.4_A14_T1.js:16:11]
 15 │ 
 16 │ var str = ";
    ·           ───
    ╰────

  × Unterminated string
    ╭─[test262/test/language/types/string/S8.4_A14_T1.js:16:11]
 15 │ 
 16 │ var str = ";
    ·           ───
    ╰────

Ok thanks! You are right. I missed that.

My first approach to fix that issue was to clone all errors when creating a checkpoint and then re-assign them when rewinding instead of storing the count of errors in the snapshot. This way the expected behaviour stays the same and the error is preserved after rewinding even if it was popped from the Vec. The only disadvantage is that it has a slightly worse performance and memory footprint.

Are you okay with cloning the lexer errors into the checkpoint or does it make more sense to figure out how to handle the duplicate errors?

[overlookmotel]

I'm not sure why the size of antd.js suddenly changed and I had to update the allocation snaphots as a result.

This is fixed by #13465.

[Boshen]

@ulrichstark I'm not sure, gut feeling is Kind::Eof should have special treatment ... any token reaching EOF is a fatal error that cannot be recovered?

[ulrichstark]

Closing this PR in favour of #13494, because the other PR is a cleaner fix and the performance impact is unexpectedly neutral instead of negative and even slightly positive on some benchmarks.