regex_match drops every capture after a non-participating group, and hard-caps at 15 groups

[InauguralPhysicist]

Finding

regex_match stops emitting captures at the first group that didn't participate, and silently ignores groups past 15. Both break positional indexing, with rc=0.

src/builtins.c:1594-1608:

regmatch_t matches[16];
if (regexec(&re, str, 16, matches, 0) != 0) { ... }
...
for (int i = 0; i < 16 && matches[i].rm_so >= 0; i++) {

POSIX sets rm_so = -1 for a group that didn't participate. Having that test in the loop condition turns "this group is empty" into "stop emitting groups".

Reproduction

print of ["(x)?(a)(b) on 'ab' ->", regex_match of ["ab", "(x)?(a)(b)"]]
print of ["(a)(b)     on 'ab' ->", regex_match of ["ab", "(a)(b)"]]

["(x)?(a)(b) on 'ab' ->", ["ab"]]            <- matched, but ZERO groups returned
["(a)(b)     on 'ab' ->", ["ab", "a", "b"]]  <- control: works

Python on the same pattern: re.match('(x)?(a)(b)', 'ab').groups() → (None, 'a', 'b').

The match succeeded — a and b did participate — but one optional group ahead of them deletes both from the result. result[2] is then out of range rather than wrong-but-present.

A 17-group pattern returns 16 elements (full match + 15 groups); group 16 onward vanishes.

Invariant violated

docs/BUILTINS.md:111 specifies regex_match of [s, pattern] → [full_match, group1, ...]. On a successful match, group n must be at index n. Any optional group anywhere in a pattern silently breaks that for every capture after it — and (x)? is not an exotic pattern.

Not load-bearing

Every current use is the easy shape — all groups participate, count ≤ 2: tests/test_regex.eigs:5,37, tests/test_raise_on_bad_input.eigs:207-208 (indexes only [0]). lib/eigen.eigs:1773 re-exports the builtin. Nothing in tests/, docs/, lib/, or examples/ depends on the early break or the 16-cap. Textbook "works only because the examples are too easy".

Fix

Size the array from the compiled pattern and move the participation test into the loop body:

size_t ng = re.re_nsub + 1;
regmatch_t *matches = xmalloc(ng * sizeof(regmatch_t));
...
for (size_t i = 0; i < ng; i++) {
    if (matches[i].rm_so < 0) { list_append_owned(result, make_null()); continue; }
    ...
}

Choose null vs "" for a non-participating group deliberately and pin it in docs/BUILTINS.md — null is the more honest encoding. The heap allocation also drops a fixed stack array, consistent with the repo's no-big-by-value-arrays-in-recursive-paths rule.

Test to add

regex_match of ["ab", "(x)?(a)(b)"] → length 4, [1] null, [2] == "a", [3] == "b"; plus a 17-group pattern asserting length 18.