Optimizations, Hardening, & Clarifications

I've upgraded the dependencies for 2026. Still, neither OpenBLAS nor Eigen issues a single AMX instruction. 8-bit GEMM lands at 98 G against a 48.4 T ceiling. NumKong does 😎

Several stupid mistakes were fixed. The ASIO packet loss was never ASIO's fault. A 256 KB batch against a 208 KB default net.core.rmem_max the kernel drops the excess before any library sees it, and one sysctl takes loss to 0% with no recompilation. NEON kernels accumulated into V8-V15, which AAPCS64 requires to survive a call, silently corrupting a caller's double. LibUnifex was used incorrectly and now scores better than most alternatives...

• Docs: Re-measure on Xeon 4, drop the other CPUs (cbc3cfc)
• Make: Split the Debug and Release trees with presets (a3c8b34)
• Docs: Refresh the Apple measurements on an M5 Pro (545fd42)
• Fix: Stop double-counting threads in the TOPS counter (34f73d7)
• Make: Restore the macOS builds (d61bd64)
• Fix: Enable io_uring setup flags on the 6.0 server (6487b27)
• Docs: Separate fast-math's speed from its danger (7c590eb)
• Improve: Make un-fused and AMX kernels match their commentary (1301aa6)
• Improve: Annotate Assembly kernels for profilers (304f020)
• Improve: Enable inline Assembly lessons on Clang (1cbc3b2)
• Fix: Flush denormals before Arm throughput kernels (c7ccb7f)
• Fix: Correct BFMMLA and AMX operation counts (0823fad)
• Fix: Preserve callee-saved NEON registers (c94a1f0)
• Improve: Drop the locks from the ASIO client (bc33d55)
• Fix: Repair the ASIO RPC client and server (2e27b90)
• Fix: Enable the LibUnifEx pipeline (a82958f)
• Docs: Measure Symmetric Transfer, don't assume it (d36ab87)
• Docs: Salvage AVX-512 mixed-sign multiplies (b6a6017)
• Docs: Correct the NRVO explanation (3c28170)
• Fix: Restore par_unseq sorting sweep on GCC 15+ (3905188)
• Docs: Academic citations (89cfe6d)