Quadratic paths made linear; Headers.get, regex crashes and five wrong results fixed - #96

Came out of a review of the library for slow paths and for null handling. Three groups of changes.

Speed

A string's .length is a strlen call and an array's shift() moves every remaining element; both sat inside loops. Measured under the release JIT (gc, Windows x64):

• String helpers read the length once.
• join builds its parts and allocates once.
• sort is a stable merge sort of positions, applied by exchanges (so -mm=own still sees elements only change places). It was a selection sort.
• flatMap, toSpliced and Array<T>.push move elements out with one reverse and pops.
• MessageQueue.take pops from an outbox refilled from the inbox.
• Array<T>.unshift and splice inserts keep the native move for fewer than 8 items and rotate for more; a single-item insert costs what it did.

Null and crashes

• Headers: the private methods now come after the public ones. lib.d.ts declares only the public ones and a program finds a method by its place, so every call ran a method two places off and get returned garbage. get now returns string | null.
• RegExp.split is declared in lib.d.ts (the same kind of mismatch).
• replaceAll with an invalid pattern throws a SyntaxError; it was an access violation.
• exec and test with lastIndex past the end fail the match and reset it to 0; it was an access violation.
• regexp_exec no longer leaks its results on a failed match.
• NUL terminators are written as <char>0; MatchResults loses an unused field and its null defaults; ArrayBuffer never holds null and throws on a detached buffer.

Wrong results

• search returns where the match starts ("abcdef".search(/cd/) was 3).
• join keeps the separator after a leading empty string.
• fill fills the array itself, as in JavaScript; it filled a copy. Behaviour change: tests/array_class_string.ts expected the copy and is updated.
• findLast returns undefined when nothing matches, not -1.
• lastIndexOf treats a fromIndex at or past the end as the last element, and one before the start finds nothing.
• Math.random is spread over [0, 1) on Windows (it assumed a 31-bit rand()).

Not in this PR

• JSON.parse of one object with thousands of keys is still quadratic (61 ms at 8,000 keys).
• String.match still returns null through string[]: the nullable type broke two existing tests at compile time.
• // @strict-null false stays in lib.ts.

Testing

• Release, Windows x64: the whole suite passes for gc, rc, none and own, compiled and under the JIT, with nine new tests.
• scripts/check_declarations.ps1 passes.
• Not run locally: debug builds, Linux, x86 and wasm.
• The ArrayBuffer changes compile but were not executed (programs cannot name the class yet).

🤖 Generated with Claude Code