Base compiles a regex and runs it on a Pike VM

[phenomenon0]

This follows the regex design in #873, taken in smaller steps. It is the first
step: a regex in Base alone, with no runtime or compiler change.

Regex.compile(p) gives a Result<String, Regex>. On a compiled regex:

• Regex.exec(re, s, at) is search.
• Regex.match_at(re, s, at) is match.
• Regex.fullmatch(re, s) is fullmatch.
• Regex.group(s, m, k) gives a group's text.

The semantics are CPython 3.11 re under re.ASCII, for literals, .,
sets, \d \w \s \D \W \S \b \B, ^ $, groups, (?:, |, and * + ?,
greedy or lazy. Offsets count code points. The VM is a Pike VM with threads
in priority order, so a run is linear in the subject.

A pattern outside the slice fails to compile; it never runs with a different
meaning. The cut constructs each fail this way, and each is a later step:

• flags
• counted repeats (so a lone { fails too, where CPython reads a literal)
• \x \u \0, backrefs, \A \Z
• named groups and lookarounds
• possessive repeats
• */+ on a body that can match empty

Also not in this step: findall/split/sub, Unicode classes, and a
native VM.

The types are Regex.* because a top-level Match would collide with
def Match in evals/hell.glob_backtrack_match.bend.

check	result
bun gates/repo.ts	PASS 46 / 46 at each commit. No cap changed. base.bend 25,423 → 31,449 of 32,000 ttok.
tests/regex, 6 tests	pass check, interp, C and JS (gates/test.ts's steps run locally)
differential vs CPython 3.11	tests/regex/oracle.bend, 160 generated rows: 0 diffs. Offline, 5 × 1,000 rows: 0 diffs on C and JS.
laws	exec = first match_at from at on; match_at(i) starts at i; fullmatch spans all. Each law fails under a VM mutant.
! smoke	one compiled regex over a list with !; on CUDA, --gpu on and off give the pin
everything else	--checkup over all 1,412 tests is byte-identical. The C and JS emitted for the 623 tests that build are byte-identical but for one rename (List.contains's specialization _0 → _1 in tests/base/list_fold).
runtime	comp.ts, bend.ts, main.ts untouched

Not run here: the cluster gate and Metal. The ! smoke has run on CUDA
only.

Size: +1,061 lines over 7 files. Base is +485 of those, and 369 test
lines are oracle data.

Base has 551 ttok left after this, so the next Base step will need room.

🤖 Generated with Claude Code