An end-to-end Ubuntu 26.04 runbook for making CosyVoice3 generate Furina-style speech on Intel Arc Pro B50 with PyTorch XPU, with a verified NVIDIA Quadro M4000 CUDA path for old Maxwell hardware.

How REX used private CMake precompiled headers to reduce clean build time without changing installed headers, downstream usage, ABI, or runtime behavior.

How REX turned upstream ROSE synchronization into a repeatable, verified workflow: read-only upstream evidence, per-commit decisions, one adapted REX commit per useful change, complete ledger coverage, and full CTest validation.

A reflection on the twelve-day REX repair session: why Codex with GPT-5.5 crossed a practical threshold for persistent engineering work, and why verification discipline mattered more than raw generation speed.

A practical guide for reviewing the large REX Clang frontend stabilization PR by evidence, subsystem, and risk instead of reading the diff linearly.

A technical retrospective on the REX Clang frontend stabilization PR: major failure families, root causes, key changed subsystems, and the risks that remain after the full CTest suite went green.

How a twelve-day REX cleanup used Codex, GPT-5.5, and a strict no-bounce repair loop to move roughly one thousand historical CTest failures to a full-suite green result.

A case study from the REX Clang frontend migration: why the generated Cxx_Grammar translation test timed out, how the failure exposed an over-eager header traversal policy, and how the fix avoided unparser hacks and CTest masking.

The final REX OpenMP Journey post: what happened when the optimized REX GPU suite was rebuilt against source-built LLVM 22, which margins moved, and why no new correctness or performance regression was confirmed.

The final LLVM 21 REX GPU performance wrap-up: why REX wins or ties every benchmark in the fair suite, which compiler mechanisms explain each result, and what native LLVM still does better even when it loses.