A focused explanation of the top-layer benchmark contract in REX: same runtime stack, same inputs, normalized correctness, benchmark-appropriate timing, clause-preserving fairness, and results interpreted as evidence rather than a simplistic scoreboard.

A focused case for keeping REX's GPU benchmark layer narrow: benchmarks are the final reality check, but they are the wrong place to first detect parser drift, semantic normalization bugs, or simple lowering-structure regressions.

A focused case for why REX still needs full GPU benchmark runs after parser, AST, lowering, and CPU-equivalence tests already pass: only real applications exposed misplaced offload init, bad timing proxies, baseline drift, and full-stack integration issues.

A focused walkthrough of the correctness side of REX benchmark validation: stripping non-semantic output lines, using reduced-output modes when benchmarks hide their results, and distinguishing current native-versus-REX agreement from stale-baseline drift.

A focused methodology post on fair GPU benchmark comparisons between REX and native LLVM: same runtime stack, clause-preserving launch policy, benchmark-appropriate timing sources, and correctness rules that do not confuse drift with regression.

A focused walkthrough of the host offload-entry table in REX: how `__tgt_offload_entry` objects are emitted, why REX uses synthetic host identity symbols, how entry names stay aligned with generated device kernels, and how the table is joined with the CUBIN during registration.

A focused debugging guide for REX GPU offloading failures that appear only at runtime: checking CUBIN presence, offload-entry integrity, kernel-name matching, init ordering, and map-array correctness.

A focused walkthrough of REX's offload init/fini policy: why the lowerer eagerly inserts `rex_offload_init()` near the top of `main`, why it deliberately does not auto-insert `rex_offload_fini()`, and what the lowering tests verify about that decision.

A focused walkthrough of the wrapper layer in `rex_kmp.h`: vendored runtime ABI structs, asm aliases for the real `libomptarget` symbols, direct and safe REX wrappers, macro rewriting of generated `__tgt_*` calls, and the `ident_t *` bridge for `__tgt_target_kernel`.

A focused walkthrough of REX's device-image registration path: why it uses standalone CUBIN files, how `register_cubin.cpp` builds `__tgt_device_image` and `__tgt_bin_desc`, how one-time registration is synchronized, and why `rex_offload_init()` is explicit.