tsx: JSX text content gets code token classes (keywords, numbers, types)

[mehdibha]

Version: @tanstack/highlight 0.0.9

Prose inside JSX text nodes is tokenized as if it were code: English words that collide with keywords get keyword, digits get number, and Capitalized words get type.

Repro

import { createHighlighter } from '@tanstack/highlight/core'
import { tsx } from '@tanstack/highlight/languages/tsx'

const h = createHighlighter({ languages: [tsx] })
console.log(h.tokenize('<Trigger>How do I get started?</Trigger>', { lang: 'tsx' }).tokens)
console.log(h.tokenize('<Item>Option 1</Item>', { lang: 'tsx' }).tokens)

Actual:

["<", ["Trigger","tag"], ">", ["How","type"], " ", ["do","keyword"], " ", ["I","type"], " ", ["get","keyword"], " started?</", ["Trigger","tag"], ">"]
["<", ["Item","tag"], ">", ["Option","type"], " ", ["1","number"], "</", ["Item","tag"], ">"]

So a sentence like "How do I get started?" renders with do and get in keyword red and How/I in the type color, and "Option 1" gets a colored digit.

Expected: text between > and the next < is inert prose with no token classes (this is also how TextMate-grammar highlighters treat JSX text).

I realize the Capitalized-identifier→type heuristic is a deliberate trade-off elsewhere, but inside a JSX text node there's no ambiguity to trade against — filing since FAQ/accuracy docs suggest corpus-driven reports are welcome. Found while migrating dotui.org's docs from shiki, character-diffing both engines over ~733k chars of tsx; button/FAQ/label copy hits this constantly.