fix(ci): unbreak the e2e build and the darwin vllm-metal pin

[localai-org-maint-bot]

Two independent breakages on master are making every open PR red for reasons unrelated to the change under review. Both are fixed here.

1. The e2e backend suite does not compile

9e831d770 (#11840) landed this in tests/e2e-backends/backend_test.go:476:

Expect(strings.ToUpper(res.GetMessage())).To(ContainSubstring(needle),

Reply.message is bytes in backend/backend.proto:393, so res.GetMessage() returns []byte and strings.ToUpper wants a string. Every other call site in the same file already wraps it (lines 708, 711).

tests/e2e-backends/backend_test.go:476:26: cannot use res.GetMessage() (value of type []byte) as string value in argument to strings.ToUpper
FAIL github.com/mudler/LocalAI/tests/e2e-backends [build failed]

This fails tests-llama-cpp-smoke on every PR whose merge ref includes current master. tests/e2e-backends is behind a build tag, so a normal go build ./... never compiles it and the breakage reached master unnoticed.

2. The darwin vllm build cannot resolve its wheel

VLLM_METAL_VERSION="v0.3.0.dev20260818075955" in backend/python/vllm/install.sh points at a release upstream has deleted. Verified today: the release API returns 404 and so does the wheel.

Upstream re-versioned vllm-metal so its version tracks the vLLM release it targets, which broke three things at once:

• the pinned tag no longer exists (current releases are v0.28.0 and v0.28.0.dev*)
• the coupled vLLM release moved out of upstream's install.sh into .github/vllm-release-tag.commit, so the extractor returns nothing even given a valid tag
• the wheel platform segment moved from macosx_11_0 to macosx_15_0

Rather than just moving the pin, the wheel name is now read from the release's own asset listing instead of being composed from a hardcoded platform segment, so a future platform-tag change cannot silently 404. The vLLM version resolves from the new metadata file with a fallback to the legacy installer. .github/bump_vllm_metal.sh and scripts/lib/extract-vllm-metal-version.sh learn the same two-source lookup, so the next nightly bump converges on the pin checked in here rather than reintroducing the break.

v0.28.0 (the stable release, which /releases/latest returns) is chosen over the newer dev tag deliberately: the dev tags are the ones upstream deletes and re-cuts.

Verification

$ go vet -tags e2e ./tests/e2e-backends/
(exit 0)

$ bash scripts/build/extract-vllm-metal-version_test.sh
(exit 0, 12 assertions)

$ tag=v0.28.0 wheel=vllm_metal-0.28.0-cp312-cp312-macosx_15_0_arm64.whl
wheel HTTP 200

The wheel URL check runs the same resolution logic install.sh now performs, end to end.

🤖 Generated with Claude Code

https://claude.ai/code/session_01UxkheFa72CEpeEJbijyjxe