Pages lay out faster, and even more JavaScript becomes machine code

Two threads of work in one release. First, rendering speed: we profiled real pages instead of guessing, and found that most of the time went into styling and layout, not scripts. Fixing what the profile showed - selectors were being re-parsed over and over, style lookups used a needlessly heavy hash, text was re-measured character by character - made an encyclopedia-length article about 38 percent faster to render, with pixel-for-pixel identical output. Along the way the profiler exposed a real layout bug: a box could report a narrower natural width after layout had already wrapped its text, which is backwards. Fixing it visibly improved real pages - on a link-sharing site, story titles and their domain now share one line exactly as the reference browser renders them. Second, the JavaScript engine: four more loop shapes now compile to machine code. Loops that call a small helper function inline the helper into the generated code and run about 27 times faster. Text scanning with charCodeAt runs about 81 times faster than the interpreter. Loops that declare their own variables - an everyday pattern that previously never got optimized at all - now reach machine code too, as do loops that write to several arrays per pass. Every change shipped the same way as the JIT itself: multi-million-case bit-exact fuzzing, a six-way conformance gate, and byte-identical render comparisons.