l3i stable 1.1.0 → 1.2.0

Added

• @dream/intern (feature intern): pools that turn strings or buffer spans into dense number tokens under exact or ascii-nocase equality, folding case inside the hash and the comparison, so no lowered copy is made and a repeated identity allocates nothing. find, resolve to the first spelling seen, and interner() for interpreted code.
• Under jit, an annotated dream_intern_Pool's intern and find lower to native code: a case-insensitive duplicate costs 315 instructions against 1378 for string.lower and a table lookup, and an insert or a bad argument runs the bound method without leaving native code.
• IrBuilder::namecall_call: the namecall and call Luau would have emitted, as a userdata namecall hook's slow path.
• l3i::VERIFIED_TOOLCHAIN: true exactly when build.rs found the verified toolchain, whether or not L3I_UNVERIFIED_TOOLCHAIN was set, for hosts that publish performance numbers.
• @dream/bytes's zstdCompress, through ruzstd's one implemented level; any other level is refused by name.
• Runtime::registered_require_modules: every path registered with register_require_module, the plan's included, in registration order.

Changed

• Q:fromXYZW lowers to native code: 93 instructions against the binder's 459.
• String, buffer and string-or-buffer arguments are read straight from Luau's value layout (54 instructions instead of 293), and a trailing optional argument is read once (two optional integers cost 335 fewer instructions).

Fixed

• A lowered rotation packs with the omitted component's own sign, so a quarter turn whose two largest components tie with opposite signs no longer packs a different rotation.