[Jemmy (JemmyIndex)]

Fixes #65078.

The bug

In a .js file, var R: ; reported TS8010 at column 8 and then crashed:

a.js:1:8 - error TS8010: Type annotations can only be used in TypeScript files.

1 var R: ;
         panic: strings: negative Repeat count

jsErrorAtRange built its span as
NewTextRange(SkipTrivia(sourceText, loc.Pos()), loc.End()). When the type annotation is
missing its location is empty (pos == end == 6), and skipping the whitespace before ;
moved the start to 7, so the span had length -1. --pretty false showed the resulting
wrong column, and the pretty writer passed the negative length to strings.Repeat.

TypeScript 6.0.3 reports TS8010 at column 7 with an empty span, because the classic
compiler creates the diagnostic through createDiagnosticForNodeInSourceFile →
getErrorSpanForNode, which skips trivia only when the node is present:

const isMissing = nodeIsMissing(errorNode);
const pos = isMissing || isJsxText(node) ? errorNode.pos : skipTrivia(sourceFile.text, errorNode.pos);

The fix

jsErrorAtRange was doing the span computation itself instead of using the parser's
existing getErrorSpanForNode, which already implements the missing-node rule. Add
jsErrorAtNode and use it for the 21 diagnostics that are anchored at a node; the three
diagnostics anchored at a node list (type parameters, parameter modifiers, type
arguments) keep the range-based helper, because createDiagnosticForNodeArray also skips
trivia unconditionally for lists.

Nothing changes for present nodes: getErrorSpanForNode still skips trivia for them.

func (p *Parser) jsErrorAtNode(node *ast.Node, message *diagnostics.Message, args ...any) {
	p.jsDiagnostics = append(p.jsDiagnostics, ast.NewDiagnostic(nil, getErrorSpanForNode(p.sourceText, node), message, args...))
}

Verification

• var R: ; in a .js file now reports TS8010 at 1:7 in both modes and --pretty
  renders the empty underline instead of panicking:

a.js:1:7 - error TS8010: Type annotations can only be used in TypeScript files.

1 var R: ;
        ~

a.js:1:8 - error TS1110: Type expected.

• Other JS-only diagnostics keep their spans: var a: number → TS8010 at the type,
  class C { x: string }, function f(p: string) {}, interface I {} → TS8006 at the name.
• Replaced the parser unit test with TestTscJSDiagnosticsForMissingNodes in
  tsc/internal/execute/tsctests/tsc_test.go. The CLI baselines cover the missing
  annotation and a present annotation with leading trivia, each with
  --pretty true and --pretty false.
• go -C tsc test ./internal/execute/tsctests -run TestTscJSDiagnosticsForMissingNodes -count=1 passes.
• Verified the regression test against the pre-fix parser: pretty mode reproduces
  panic: strings: negative Repeat count, while non-pretty mode fails the baseline
  because TS8010 is reported at column 8 instead of 7.
• The full parser, CLI, compiler/conformance and language-service suites pass locally,
  as do lint, formatting and tidy checks. Native and WASM API tests pass with Node 24.
  The full validation run is limited by this Windows session's missing symlink
  privilege in TestSubscribeSymlinkCreate and TestSubscribeSymlinkDelete.

AI assistance

This PR was written with AI assistance, and the change is deliberately scoped to how the parser computes the span of JS-only diagnostics that are anchored at a node. I reproduced the crash and verified the fix manually: --pretty no longer panics, TS8010 on var R: ; is now reported at 1:7 with an empty span like TypeScript 6.0.3, and var a: number, class C { x: string }, function f(p: string) {} and interface I {} keep their existing spans. The CLI regression tests cover missing and present annotations in both output modes and pass locally. They reproduce the original panic and wrong column against the pre-fix parser. I understand the implementation and am happy to answer questions about it.

[Copilot]

🟢 Approval recommended

No active review issues remain for the node-aware JavaScript diagnostic span changes.

No open issues 🎉

What changed in this PR

Fixes #65078 by using node-aware spans for JavaScript diagnostics, preventing missing nodes from producing inverted ranges and crashing pretty diagnostic output.

Changes:

• Adds jsErrorAtNode and migrates node-based diagnostics to the existing span helper.
• Preserves range-based handling for node-list diagnostics.
• Adds regression tests for missing type annotations and leading trivia before present annotations.

File	Description
tsc/​internal/​parser/​parser.go	Uses node-aware diagnostic spans to handle missing nodes correctly.
tsc/​internal/​parser/​parser_test.go	Verifies diagnostic positions for missing and present type annotations.

_{Review effort: Balanced}

[Jake Bailey (jakebailey)]

Why a unit test and not a compiler test or tsctest?

[Jemmy (JemmyIndex)]

I used a parser unit test to assert the exact diagnostic spans for missing and present nodes. A tsctest would cover the reported CLI crash more directly. I'll move this coverage to a tsctest, including the missing-annotation repro in both pretty and non-pretty modes and a present-annotation case.