ci: pilot per-arch split + manifest merge for faster-whisper and llama-cpp-quantization

[localai-bot]

Summary

Pilots Phase 2.3 + 2.4 of the CI migration plan: convert two backends from QEMU-emulated multi-arch to native per-arch + manifest-list merge. This validates the split-and-merge pattern end-to-end on real CI before fanning out to the other 34 multi-arch entries (Task 2.5, follow-up PR).

What changes

Two pilot backends split (.github/backend-matrix.yml):

• -cpu-faster-whisper (small Python, fast baseline)
• -cpu-llama-cpp-quantization (heavier compile, stress test)

For each: the single platforms: 'linux/amd64,linux/arm64' matrix entry is replaced with two per-arch entries — amd64 leg on ubuntu-latest, arm64 leg on ubuntu-24.04-arm (native, ~5–10× faster than emulated). Each new entry carries platform-tag: 'amd64' | 'arm64', which the previously-merged Phase 2.1 wires into backend_build.yml to scope the registry cache and the digest artifact name.

Merge-job infrastructure (reused by Task 2.5+):

• .github/workflows/backend.yml and backend_pr.yml forward platform-tag from matrix to backend_build.yml.
• A new backend-merge-jobs job in both workflows consumes a merge-matrix output from generate-matrix and calls the existing backend_merge.yml (already-shipped from PR ci: phase 1-3 of GHA free tier migration (path filter, multi-arch split prep, /mnt disk relief) #9726).
• scripts/changed-backends.js gains a computeMergeMatrix(entries) helper that groups filtered linux entries by tag-suffix, emits an entry only for groups of size ≥ 2, and warns if tag-latest disagrees across legs (cheap insurance for the 34-backend fan-out coming next).
• PR-side merge job is also event-gated on github.event_name != 'pull_request' so the no-op-on-PR run doesn't even start.

What's NOT in here (follow-ups)

• Task 2.5: fan out the same shape to the other 34 multi-arch entries.
• Task 2.6: same pattern for image.yml / image-pr.yml (3 multi-arch entries).
• Phases 4–5: migrate bigger-runner and arc-runner-set jobs to free tier (depends on Phase 3 disk relief, already shipped).

Decisions worth flagging

• Singletons not merged: backends with a single matrix entry (single-arch) push by digest only and don't need a manifest list. The computeMergeMatrix helper skips them.
• tag-latest mismatch guard: cheap warning surfaced if the two legs disagree on tag-latest. Won't fire today (the two pilot legs both say 'auto'); future-proofs the 34-entry fan-out.
• PR variant gating: the merge job's if: is at the job level (github.event_name != 'pull_request'), so the matrix doesn't even instantiate on PRs — saves a runner over relying on backend_merge.yml's internal step gates.

Test plan

• On this PR, the backend_pr.yml generate-matrix job runs and emits an empty merge-matrix (no backend dir touched), so backend-merge-jobs is correctly skipped via has-merges == 'false'.
• After merge, the first push that touches backend/python/faster-whisper/ schedules:
  • 2 per-arch backend_build.yml jobs (amd64 + arm64 native), each pushing by digest under digests-cpu-faster-whisper-amd64 / -arm64.
  • 1 backend-merge-jobs matrix entry for -cpu-faster-whisper that downloads both digests and runs docker buildx imagetools create to produce the final tagged manifest list.
• docker buildx imagetools inspect quay.io/go-skynet/local-ai-backends:master-cpu-faster-whisper shows two platforms (linux/amd64, linux/arm64).
• arm64 native build of faster-whisper finishes faster than the previous emulated multi-arch run (compare wall-clock from before/after).
• Same checks for -cpu-llama-cpp-quantization (the heavier one).
• Weekly Sunday cron (added in PR ci: phase 1-3 of GHA free tier migration (path filter, multi-arch split prep, /mnt disk relief) #9726) still rebuilds the full matrix and the merge-matrix correctly contains the two pilots.

Plan reference: docs/superpowers/plans/2026-05-08-ci-migration-to-gha-free-tier.md (uncommitted working artifact).

Assisted-by: Claude:claude-opus-4-7

[mudler]

can only test on master