False-red local and GPU tests: memory caps, odd thread counts, GPU OOM #5337

Scope

Tests that fail for reasons outside the code: memory caps on a small host, odd thread counts, and GPU out-of-memory.

Why it matters

A false red costs a rerun and teaches people to ignore red.

Done when

Each test below passes reliably on every host class it runs on, or skips with a stated, logged reason.

Parent epic: #3997.

Rows

• nextest: 5 aprender-serve lib tests SIGKILLed by 16G local cargo-memcap — local runs false-red (heavy-memory group; mega test allocates 34GB) #4343 — nextest: 5 aprender-serve lib tests SIGKILLed by 16G local cargo-memcap — local runs false-red (heavy-memory group; mega test allocates 34GB)
• ThreadConfig: odd rayon pool sizes (RAYON_NUM_THREADS=3) mis-split work (serve, FLAKE-0) #4537 — ThreadConfig: odd rayon pool sizes (RAYON_NUM_THREADS=3) mis-split work (serve, FLAKE-0) (branch: fix/4537-thread-odd-pool)
• yoga cuda-unit: cta64_vs_cta32_vs_cublas_fp16 CUDA_ERROR_OUT_OF_MEMORY under the host GPU lock (flaky, 1 of 2 on yoga-eph) #4593 — yoga cuda-unit: cta64_vs_cta32_vs_cublas_fp16 CUDA_ERROR_OUT_OF_MEMORY under the host GPU lock (flaky, 1 of 2 on yoga-eph) (branch: flake0/4593-cta64-oom)

Each row was a separate ticket, folded here under the 100-open-ticket cap (TRIAGE-100, 2026-10-14) and closed as not planned. Reopen a row's ticket when its work starts; an open PR for it reopens it.

Agent: aprender-clerk