Function split_text_chunks 

pub fn split_text_chunks(text: &str, max_chars: usize) -> Vec<String>

Splits text into chunks of at most max_chars chars, on paragraph boundaries wherever packing allows, and never inside a fenced code block.

The contract, in the order it matters:

• Round trip. The concatenation of the returned chunks is text byte-for-byte, always. Separators travel with blocks: a paragraph or fenced block keeps the blank-line run that follows it, and a blank run at the very start of the document rides in front of the first chunk’s content whenever the two fit one chunk together (a run that cannot stands alone as a whitespace-only chunk — TextModelChunkedExt::complete_chunked never asks the model about one). \r\n input round-trips too — the split works on byte offsets, and a \r\n blank line counts as blank.
• The bound. Every chunk is at most max_chars chars (characters, not bytes), with exactly one exception: a single fenced code block longer than max_chars is one oversized chunk, because chunks are never split inside a fence.
• Paragraph preference. Boundaries fall on blank-line runs whenever packing allows, so a chunk is a whole number of paragraphs and blocks. (The blank-line runs themselves are packing items: a run that no longer fits a chunk starts the next one, still never inside a paragraph.)
• Fences. A line whose leading-whitespace-stripped form starts with three or more backticks or tildes opens a fence; it closes at a line whose trimmed form is a run of the same marker at least as long, with no info string. A blank line inside a fence stays inside it, an unclosed fence runs to the end of the input as one (possibly oversized) chunk, and a four-backtick fence does not close at an inner three-backtick line. Indented (four-space) code blocks are not fence-tracked: they are ordinary lines and may split at line boundaries.
• Fallbacks. A plain block longer than max_chars splits at line boundaries, greedily packing whole lines; a single line longer than the bound is hard-cut every max_chars chars, always on a char boundary — multibyte input never splits mid-character.

Empty input returns an empty Vec; whitespace-only input is a single chunk holding it, however long — there is nothing in it to model.

Panics

Panics if max_chars is zero: a zero-char chunk cannot hold anything.

Examples

use cratefield_core::{reassemble_text_chunks, split_text_chunks};

let text = "first paragraph\n\nsecond paragraph";
let chunks = split_text_chunks(text, 20);
assert_eq!(chunks, vec!["first paragraph\n\n", "second paragraph"]);
assert_eq!(reassemble_text_chunks(&chunks), text);