feat(latex): add sourceOffsets to parser-generated Error nodes

[zojize]

closes #311

AI usage disclosure: majority of implementation, human reviewed

Summary

The LaTeX parser reports parse errors as MathJSON Error nodes that say what went wrong but not where in the input. This PR populates the existing MathJsonAttributes.sourceOffsets field on parser-generated Error nodes with the character range of the offending LaTeX, so consumers can map a parse error back to its source span — e.g. to highlight the exact invalid token inline in a MathfieldElement.

// LatexSyntax().parse('\\foo')
{
  "fn": ["Error", { "str": "unexpected-command" }, ["LatexString", { "str": "\\foo" }]],
  "sourceOffsets": [0, 4]
}

Live demo: https://zojize.github.io/compute-engine/ — type an invalid expression and the offending token is colored; empty \sqrt/\frac slots get a red placeholder box. (Closes #.)

What's exposed

• sourceOffsets on parser Error nodes — zero-based, end-exclusive character offsets into the serialized LaTeX (tokensToString). For input that round-trips through the tokenizer unchanged (no comments, Unicode normalization, or macro expansion — i.e. editor-generated LaTeX), they equal offsets into the original input string.
• Missing-token errors (empty \sqrt{}, empty \frac{}{}) use a zero-width range at the parser position.
• Parser.sourceOffsets(startToken, endToken?) — public helper converting a token range to that character range, so custom dictionary entries can attach offsets to errors they raise. Parser.error() uses it internally.
• Boxed path — offsets propagate through boxing/serialization, so they survive ce.parse(latex).toMathJson() (selectable via the metadata option, included with metadata: 'all').

Implementation

Deliberately minimal — tokenizer.ts is not touched. Offsets are derived from the existing tokensToString via prefix length, so no per-token source map is threaded through the tokenizer.

Area	Change
latex-syntax/parse.ts	Parser.sourceOffsets() (= latex(0, token).length); error() returns { fn, sourceOffsets }
latex-syntax/dictionary/definitions-arithmetic.ts	empty \sqrt{} / \frac{}{} operands emit a positioned missing error instead of the position-less MISSING sentinel
latex-syntax/types.ts, math-json/types.ts	Parser.sourceOffsets signature + documented sourceOffsets semantics
boxed-expression (box, serialize, boxed-function, abstract-boxed-expression) + math-json/utils.ts, types-expression.ts, types-kernel-serialization.ts	propagate sourceOffsets through ce.parse().toMathJson()

~150 lines of source + test-snapshot updates.

⚠️ Behavior change (please review)

Parser Error expressions now use the object form { fn: ["Error", …], sourceOffsets } instead of the bare array ["Error", …] whenever a range is available. Both are valid MathJSON, but a consumer matching Array.isArray(expr) && expr[0] === "Error" must also handle expr.fn?.[0] === "Error". This is the bulk of the test-snapshot churn in this PR. If you'd prefer this gated behind a parse option rather than on by default, happy to adjust.

Limitation (documented on the field)

Offsets are measured on tokensToString(tokenize(input)), so they drift from the original string only when that round-trip is lossy — comments, Unicode normalization, or multi-codepoint ZWJ emoji. There the range stays valid and monotonic but isn't byte-aligned to the original. Editor-generated LaTeX never hits those paths, so it's exact in practice. A heavier implementation could thread true source spans through the tokenizer if byte-exact ranges for arbitrary pasted input are ever needed.

Testing

• latex-syntax/* and incomplete-expressions suites pass; snapshots updated for the object-form errors. New assertions in errors.test.ts cover unknown command (\foo → [0,4]), nested error through ce.parse().toMathJson(), unexpected delimiter, and zero-width missing operands.
• Full suite: 10,496 passing. The only failures are two pre-existing, environment-dependent flakes unrelated to this change — arithmetic.test.ts (last-digit machine-float drift) and compile-performance.test.ts (a ~1MB memory-threshold assertion).
• tsc --noEmit clean.

Notes

• src/api.md is not regenerated here — running typedoc/concat-md locally reformats the whole file (toolchain/version drift, ~3k lines of noise). The API documentation lives in the source doc-comments included in this PR; happy to regenerate api.md if you'd like it in the same PR or prefer to do it at release time.
• I can add a CHANGELOG.md entry under ## [Unreleased] → ### New Features if that's the expected flow — let me know.