[XBold]

What this is

PR #1424 added the Qwen3.8 CUDA tier (dense trunk int8 in VRAM + hot-expert fp8 tier) in v1.12.0, but left two stale "CPU only" claims behind:

• docs/qwen38.md: one line under the memory table says "There is currently no Qwen3.8 GPU backend." — immediately preceding a 60+ line section titled "GPU: CUDA VRAM expert tier" that describes exactly how to use it.
• README.md: the requirements table marks Qwen3.8 as "not supported; CPU only", and the build command card says "make -C c qwen38 (CPU only)" — both contradicted by the same GPU section on qwen38.md that already shipped in 1.12.0.

This removes just those stale claims. The GPU documentation itself, benchmarks, env vars (Q38_TRUNK_GPU=1, Q38_TRUNK_CPU_INT8=0, etc.), and all three stages of the disk→RAM LRU→VRAM hierarchy are already correct on dev — they were added alongside the code in #1424, only the intro sentence and README entries were not updated.

No code changes; no dependency added; default CPU build path untouched.

Validation

• No C/Makefile/Python code changed; default CPU build on dev is unchanged.
• The GPU section in qwen38.md remains intact with all its commands and benchmarks: verified manually.
• README Qwen3.6 entry already uses the same pattern (CUDA=1 for the VRAM expert tier); this brings Qwen3.8 into alignment.
• No model files, generated binaries, or benchmark artifacts are included.

Checks

• make -C c check — doc-only change; default build on dev is unchanged.
• No model files, generated binaries, or benchmark artifacts are included.
• Default CPU build remains dependency-free (no change at all).