Regex

Embeds the value's already-computed RegexEncoder.Encoded shape, reconstructed by RegexDecoder, instead of regenerating it through the public smart constructors (concat, alt, inter, star, unary_!, lookahead) at every call site - the same spirit as TokenMatcher's ToExpr embedding its DFA tables as plain data. The smart constructors exist to (re-)establish ACI normalization and merge Chars sets when building a tree from scratch; a Regex value reaching this given is already normalized, so re-running them at every macro call site - and again at every class load, since the generated code re-executes them - would just redo that work for a result that's already known. See RegexEncoder for why the encoding is flat data rather than nested constructor calls.

The names of the product elements.

The name of the type.

Alternation of a collection, preferring earlier elements - see Regex.Alt's doc comment. Builds flat straight into an AltBranches (order-preserving dedup, one pass) instead of .toVector.distinct followed by a separate .toSet later for equals/hashCode: the latter would build two different set-like views of the same elements where one suffices.

Smart constructor for a capturing group node - see Regex.Group. Unlike lookahead/ star/etc., this never collapses even when inner is Eps/Empty: the group's identity (its index/name existing at all) matters regardless of what it can match, e.g. (a?) is still a real, numbered group even though it can capture the empty string.

Intersection of a collection.

Convenience: literal char.

Convenience: literal string, one character per code point (a surrogate pair is one character).

Zero-width lookahead assertion: Regex.lookahead(r, true) is (?=r), Regex.lookahead(r, false) is (?!r). Degenerates to Eps/Empty directly when r itself is Eps/Empty, since the assertion is then a constant regardless of context.

Convenience: char range.

All-strings regex Σ*.

Upper bound on quantifier bounds accepted by repeat. Repeat itself stays O(1) regardless of lo/hi (see repeat), but each Brzozowski derivative step through it (Subset.rawDerive) peels one repetition off, so exploring L(r{0,hi}) up to a fixed point (Subset.isEmpty/subset) still visits on the order of hi distinct states - an unbounded value would let a single malformed pattern blow up that exploration. Real-world token patterns never need bounds anywhere near this size.

Embeds the value's already-computed RegexEncoder.Encoded shape, reconstructed by RegexDecoder, instead of regenerating it through the public smart constructors (concat, alt, inter, star, unary_!, lookahead) at every call site - the same spirit as TokenMatcher's ToExpr embedding its DFA tables as plain data. The smart constructors exist to (re-)establish ACI normalization and merge Chars sets when building a tree from scratch; a Regex value reaching this given is already normalized, so re-running them at every macro call site - and again at every class load, since the generated code re-executes them - would just redo that work for a result that's already known. See RegexEncoder for why the encoding is flat data rather than nested constructor calls.