Environment

md4x 0.0.25, Bun 1.3.11, CachyOS

Reproduction

import { renderToAnsi } from "md4x"

const md = `# Test

Text before.

\`\`\`lua
local a = 1
\`\`\`

Text between.

\`\`\`lua
local a = 2
\`\`\`

More text.

\`\`\`lua
local a = 3
\`\`\`

Final text.

\`\`\`lua
local a = 4
\`\`\`
`

const blocks = []
renderToAnsi(md, {
  highlighter: (code, block) => {
    blocks.push({ code, lang: block.lang, start: block.start, end: block.end
})
  },
})
console.log(blocks)

This gives the following output:

     [
       {
         "code":
     "local a = 1
     \x1B[22m
     S"
     ,
         "end": 86,
         "lang": "lua",
         "start": 61,
       },
       {
         "code":
     "2m  local a = 2
     \x1B[22m
     Mor"
     ,
         "end": 139,
         "lang": "lua",
         "start": 114,
       },
       {
         "code":
     "local a = 3
     \x1B[22m
     And s"
     ,
         "end": 182,
         "lang": "lua",
         "start": 157,
       },
       {
         "code": "local a = 4",
         "end": 222,
         "lang": "lua",
         "start": 205,
       },
     ]

Describe the bug

Code content includes leaked (partial) ANSI sequences and surrounding text:

• Block 0: "local a = 1\x1b[22m\nS" — DIM_OFF + next text leaks in
• Block 1: "2m local a = 2\x1b[22m\nMor" — starts with tail of DIM, ends with
  next text
• Block 2: "local a = 3\x1b[22m\nAnd s" — DIM_OFF + next text
• Block 3: "local a = 4" — correct (last block, nothing after it)

The error accumulates: each block's end drifts further by ~2 chars. The issue
is probably that meta->start is captured before ANSI_DIM is rendered and meta->end is
captured after ANSI_DIM_OFF, so the offsets include the DIM/DIM_OFF escape
sequences rather than pointing to the actual code content.

From my testing it also seems that only the second block in the markdown starts with a partial
DIM and includes the prefix. All other blocks typically don't have the prefix included.

Additional context

No response

Logs