The unified-memory record cites a test that reads the CPU backend, not CUDA

[localai-bot]

A post-hoc review of #1620 (merged as part of the VT-REFTIER-HOST-ADDRESSABLE record) verified its conclusion but found the evidence for it is wrong, and one copy of that wrong evidence is permanent.

The defect

#1620 claims that tests/vllm/platforms/test_platform.cpp "pins for GB10" the CUDA backend's DeviceMemoryIsHostAddressable() default. It does not.

FakeUnifiedAddressablePlatform reports device_type() == kCUDA (:284), but its backend() returns vt::GetBackend(DeviceType::kCPU) (:285). So the assertion at :329:

CHECK_FALSE(gb10.backend().DeviceMemoryIsHostAddressable());

reads the CPU backend, not the CUDA one. The test's own comment says as much. Nothing in the tree pins the real CudaBackend.

What survives

The conclusion is still true, but by absence rather than by that test: CudaBackend has no override, so it inherits the base default false at include/vt/backend.h:77. Re-derived and confirmed alongside it — Vulkan returns literal true (vulkan_backend.cpp:135), Metal returns MetalContext::Get().unified_memory() (metal_backend.mm:123), ROCm returns unified_memory_ (rocm_backend.hip:371), CPU has no override. Both AdoptDeviceBytesAsHost branches are correctly gated (qwen3_5_weights.cpp:345, :365). A sixth backend (Tenstorrent) exists, is unmentioned, and is correctly inert.

Why this needs its own issue

The wrong citation landed in three places:

• docs/ENVIRONMENT.md:210 — editable
• the commit message — immutable, but git history is expected to be a record of what was believed at the time
• .agents/issue-index.md:552 — append-only. It can never be edited.

Per CLAUDE.md, the index is append-only and carries merge=union; a row is never edited and never deleted. So the correction cannot be applied in place and has to exist as a new row pointing here.

Owed

• Correct docs/ENVIRONMENT.md:210 to cite the base-class default rather than the test.
• Either give the real CudaBackend an actual pinning test, or state plainly in the record that the default holds by absence of an override and is unpinned.

Found by post-hoc review, not by a gate. No gate could have caught it: the assertion passes, and it passes for a reason unrelated to what it claims to measure.