QVAC-23327 fix: use the mmap weights path for CPU SmolVLA loads

[DmitryMalishev]

🎯 What problem does this PR solve?

• Four VLA tests (vla-hparams-shape, vla-run-synthetic-shape, vla-run-stats, vla-invalid-img-size) fail intermittently on iOS during a full SDK e2e run with Failed to load SmolVLA model. Reproduced on a local iPhone 16e and on iPhone 16 Pro in CI.
• smolvlaLoadModel prefers mapping the GGUF and wrapping it with ggml_backend_dev_buffer_from_host_ptr, falling back to one large buffer plus a copy. The choice reads capabilities off the device that owns the default buffer type — but ggml declares the CPU buffer type with /* .device = */ NULL (see the FIXME in ggml_backend_cpu_buffer_type), so ggml_backend_buft_get_device returned nullptr, both gates stayed false, and the mmap path was unreachable on every CPU load. qvac-fabric's own loader resolves the CPU device by type for exactly this reason (create_backend_buffers, src/llama-model.cpp); this port dropped that fallback.
• Consequence on mobile: the e2e loads the 1.9 GB smolvla-libero-vision-q8 GGUF with backend: 'cpu', so every load committed the whole weight set to anonymous memory. Device samples from a passing iPhone 16 Pro run show footprint and resident rising together by 2028 MB and peaking at 2693 MB footprint / 3010 MB resident — within ~10% of the per-process jetsam ceiling. (A file-backed mapping would instead widen the resident-minus-footprint gap by ~1.9 GB; that gap stayed flat at ~320 MB.) Once earlier tests leave the process near the ceiling the allocation is refused outright: footprint never moves and the load fails ~25 ms after the backend comes up. Hence "intermittent", and hence more frequent in a full-suite run than in a filtered one.
• Nothing said so. A missing device produced no log line at all, so the fallback looked like the intended route, and the thrown error named only the GGUF path.

📝 How does it solve it?

• smolvlaResolveDevice() falls back to ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU) when a buffer type carries no device, matching qvac-fabric's loader. CPU loads now map the file: weights become clean, evictable, file-backed pages, so only the compute buffers (~577 MB in the run above) count against the jetsam limit. Projected peak footprint drops from ~2.7 GB to ~1.2 GB.
• Log why the fast path was skipped, with the device name and both gate values, so this can never regress silently again.
• Reject a mapping whose tensor data is not buffer-type aligned. The CPU backend asserts on an unaligned host pointer ("buffer pointer must be aligned"), and an assert aborts the process instead of returning NULL; that code only becomes live for CPU devices with this change, so it needs a guard rather than a crash.
• Carry the failing step in model.load_error, so the JS error reads failed to load SmolVLA model from: <path>: out of memory: 1889 MB weights buffer instead of naming only the path.

Verified that no weight tensor is written during inference (every ggml_backend_tensor_set target is a graph input), which is what makes the read-only mapping safe. Behaviour on Metal/Vulkan/CUDA is unchanged — those buffer types already carry a device. Windows is unaffected (the mmap path is #ifndef _WIN32).

No version bump or changelog here, per the repo's split between change PRs and chore: bump vla-ggml release PRs.

🧪 How was it tested?

• New test/unit/test_smolvla_weights_device.cpp pins both fast-path gates for a CPU load, so swapping the helper back for a plain ggml_backend_buft_get_device fails the suite.
• cpp-tests-vla dispatched on this branch ahead of the PR: linux-x64, darwin-arm64, darwin-x64, windows-x64.
• Labelled for run-cpp-addon-tests, run-desktop-addon-tests and run-mobile-addon-tests: the desktop integration tests load the real SmolVLA GGUF on both the preferred backend and forced CPU, which is the path this PR makes live, and the mobile leg exercises it on iOS/Android hardware.

🔌 API Changes

No API surface change. One behaviour change worth noting for consumers: a failed SmolVLA load now appends the reason to the existing error message.

[github-actions]

License compliance — clean

No new dependency license findings in this PR.

Warn-only (shadow) mode — this check does not block merges yet.

Updated automatically by the canonical license compliance workflow.

NOTICE presence (advisory)

Missing NOTICE (advisory, does not block):

• ./.github/actions/release-merge-guard
• ./docs/website
• ./packages/ggml-coload-smoke
• ./packages/fabric/test/integration
• ./packages/inference-addon-cpp/mobile
• ./packages/sdk/e2e
• ./packages/llm-llamacpp/benchmarks/performance
• ./packages/llm-llamacpp/benchmarks/server
• ./packages/vla-ggml/sim/server
• ./packages/embed-llamacpp/benchmarks/performance
• ./packages/embed-llamacpp/benchmarks/server
• ./packages/asr-ggml/benchmarks/server

[github-actions]

Mobile integration tests — @qvac/vla-ggml (Android)

Result: passed

metric	value
Devices passed	9
Devices failed	0
Test cases total	27
Test cases passed	27
Test cases failed	0
Test cases skipped	0

View workflow run

[github-actions]

Mobile integration tests — @qvac/vla-ggml (iOS)

Result: passed

metric	value
Devices passed	6
Devices failed	0
Test cases total	18
Test cases passed	18
Test cases failed	0
Test cases skipped	0

View workflow run

[github-actions]

Review Status

Current Status: ❌ PENDING
Approvals so far: none

Pending reviews: Needs 1 Management or Team Lead, and 1 more from Management, Team Lead, or Member.