`-i` matches multi-character sequences via Unicode case folding where GNU matches one

[sylvestre]

Under -i, uu_grep lets a single-character pattern match a multi-character sequence when that sequence is a Unicode case-folding of one character — e.g. st (folds from the ligatures ſt/st), ss (from ß), or ff/fi/ffi (from ff/fi/ffi). GNU under LC_ALL=C folds case 1:1 and matches a single character. The input here is plain ASCII, so this is not the locale/encoding limitation — the extra matching comes from the case folder, not from byte-vs-codepoint handling.

Found by the differential fuzzer (fuzz_grep).

Rust (incorrect)

$ printf 'st\n' | ./target/release/grep -o -i '[[:alpha:]]'
st
# one match spanning two characters

GNU (correct)

$ printf 'st\n' | LC_ALL=C /usr/bin/grep -o -i '[[:alpha:]]'
s
t
# two separate single-character matches

More cases (Rust → GNU): ss → ss vs s/s; ff → ff vs f/f; ffi → ffi vs f/f/i. It is order-sensitive (st merges but ts does not) and changes match counts, not just -o output.

[lhecker]

This is a shortcoming of us not handling LC_* at all just yet. It always operates as if it was en_US.UTF-8 and I believe in that case grep folds it as per Unicode, right?

[wondr-wclabs]

Good point that LC_* is the larger missing seam, but I do not think GNU grep's UTF-8 behavior is full Unicode case folding in the sense Oniguruma is applying here.

GNU's own manual calls out this exact class of case: ß normally capitalizes to the two-character string SS, but under grep -i it does not match SS (and even matching capital ẞ is documented as locale/implementation-sensitive): https://www.gnu.org/software/grep/manual/html_node/Matching-Control.html

That makes the issue a little different from just “we behave as en_US.UTF-8 until locale support exists.” The current behavior lets one regex atom match a multi-character input sequence ([[:alpha:]] matching st, ss, ffi, etc.), which changes -o output and counts. GNU appears to avoid those multi-character fold expansions even outside the C-locale discussion.

So I think the eventual fix should probably be: keep broader LC_* support as the long-term locale seam, but make the current ignore-case matcher avoid multi-codepoint/full-fold expansions for GNU compatibility. I have not opened a PR yet because I want to check whether Oniguruma exposes a clean option/syntax flag for “simple folds only”; if it does not, a hand-rolled workaround around character classes and -o boundaries could get brittle quickly.