fix(core): drop NUL-byte separator from the when-clause regex tokenizer

[thorwhalen]

Why

Surfaced during the v1.14–v1.18 pre-wiring review (#51) via a byte-level scan: packages/core/src/when.ts used a NUL byte (\x00) as the regex pattern | flags separator — the tokenizer joined pattern + <NUL> + flags (was line 180) and the parser split on the same NUL (was line 348).

It was self-consistent (all tests passed), but:

• The file — and the built dist — were binary to git. No diff, blame, or PR-review rendering for when.ts.
• It was fragile. The separator is invisible; a maintainer "tidying" it to a space would silently break any regex whose pattern contains a space (/foo bar/ → split(' ') mangles it). That exact case was never tested — it worked only because the separator was a NUL.

What

The regex token now carries pattern and flags as distinct Token fields — no separator at all. This is behavior-identical for every input (the NUL was correctly doing its job; this just removes the fragile packing) and removes the non-printable byte from source and dist.

• Token gains an optional flags?: string (regex-only).
• Tokenizer: value: pattern, flags instead of value: pattern + <NUL> + flags.
• Parser: reads t.value / t.flags instead of t.value.split(<NUL>).
• New regression test: a regex literal whose pattern contains a space (/foo bar/, /foo bar/i).

Verification

• when.ts is now valid UTF-8 (0 NUL bytes); built dist verified NUL-free.
• Core: typecheck clean, 96 tests pass (when.test +1).
• Full monorepo typecheck + tests green (core is a widespread dependency; nothing downstream changed).

acture core patch bump (changeset included). No public API or runtime-behavior change.

https://claude.ai/code/session_01A1PykLHv2BjQRQhspjvPWy