GeezerVison

Fixes #3012.

Changes proposed in this pull request:

• Switch to xjb-as which is an improvement over zmij-as
• Comply to the ECMA262 Specification with the exception of a trailing .0

Note: I wrote xjb-as with claude assisting me for an initial port. I then optimized by hand. Everything is carefully checked and passes over a trillion fuzz cases compared directly against Number::parse() in V8. It also reaches 100% code coverage and tests every edge case both I and claude could think of.

It's good quality stuff though, and I'm quite confident in that.

Current implementation measurements from independent review of 0c115329 against b6bda05, including the exact-integer string path. These are diagnostic results for the stated engine, machine, runtime, and workloads.

Generated Wasm size for tiny modules exporting f64.toString(), f32.toString(), or both, compiled with -O --noAssert --runtime stub:

With the default incremental runtime, the both-width module grows from 6,834 B to 13,248 B (+6,414 B). These small formatter-focused modules do not estimate whole-application growth. The generated string wrappers call shared buffered formatters; the exact-integer direct-allocation path adds specialized code.

Throughput on Node 24.22.1, x86-64 AMD EPYC 9V74, default incremental runtime, scalar builds: 4,096 inputs per bucket, 150,000 warmup operations, and medians of seven alternating base/PR samples of 400,000 operations. The timed loops run inside Wasm.

Ordinary decimals are nonintegral values roughly between −1,000 and +1,000, formed by dividing integers by 1,000. Gains are strongest for buffered formatting and integers; complete-string performance varies with workload. The 96.0 → 96.4 ns ordinary-decimal f64 result is effectively flat, not a meaningful measured regression.

A separate SIMD diagnostic run improved random-finite buffered formatting by 2.31× (f64) and 2.75× (f32) against its own interleaved baseline. Ordinary-decimal f64 string output improved by about 6% in that run.

Correctness validation at this revision includes exact BigInt binary32 rounding intervals with ties-to-even, the four formerly double-rounded f32 regressions, all four checked-in debug/release × scalar/SIMD configurations (556,880 exact output comparisons), 28-code-unit canaries, the float-array join allocation budget, and independent verification of all 618 cached powers. Independent review also checked 28,076 cases against C++ std::to_chars(float, scientific) and 12,384 additional boundary comparisons across the buffered, ordinary string, and low-level string paths. These are substantial checks, not an exhaustive proof over all floating-point bit patterns.

• I've read the contributing guidelines
• I've added my name and email to the NOTICE file

P.S. Sorry about the diff. Most of it is generated WAT fixtures.