fix(gallery): correct meta-backend definitions for platform auto-selection

[localai-bot]

Follows the same root cause as #10291 (opus): backends that ship per-platform images must be meta backends (a capabilities: map and no uri) so the right variant is auto-selected per platform. An audit of backend/index.yaml found several that were misdefined, so on the affected platform the wrong (or a non-existent) backend is selected.

Fixes

Category 1 - base was a concrete with a metal- sibling (identical to the opus bug)

silero-vad, piper, kitten-tts, local-store (each + -development): the anchor carried a uri and a separate metal-<name> existed, so on macOS Load("<name>") got the CPU build / failed. Converted each anchor to a meta ({default: cpu-<name>, metal: metal-<name>}), added the cpu-<name> concrete (keeping the original image tag), renamed <name>-development -> cpu-<name>-development, and added the <name>-development meta.

silero-vad is the VAD-on-Mac failure in #9813.

Category 2 - mlx family had uri AND capabilities (capabilities silently ignored)

mlx, mlx-vlm, mlx-audio, mlx-distributed (+ -development): because the anchor had a uri, IsMeta() was false and the capabilities map was ignored - Load("mlx") always got the metal-darwin-arm64 image (broken on CUDA/Linux), and the metal-<name> capability target did not even exist. Removed the uri/mirrors from each anchor (now real metas) and added the missing metal-<name> / metal-<name>-development concretes. Verified every other capability target exists.

Relevant to the mlx-vlm install problems in #9810.

Category 3 - dangling capability targets (install fails on those platforms)

• diffusers nvidia-l4t-cuda-12: repointed nvidia-l4t-arm64-diffusers (missing) -> nvidia-l4t-diffusers (exists).
• kokoro nvidia-l4t-cuda-12: repointed nvidia-l4t-arm64-kokoro (missing) -> nvidia-l4t-kokoro (exists).
• coqui nvidia-cuda-13: removed the key (cuda13-coqui does not exist) so cuda13 falls back to nvidia/default.

Left unchanged (flagged for maintainer)

• locate-anything: its capabilities reference *-locate-anything-cpp concretes that do not exist anywhere (no images, no entries) - it appears to be a broken/unpublished WIP entry. Not touched; please confirm whether its concretes should be published or the entry removed.

Verification

• Referential-integrity audit: from 62 problems down to the documented residue (the only remaining flags are locate-anything, and opus/opus-development which are fixed in fix(gallery): make opus a meta backend for platform auto-selection (#9813) #10291 and absent from this branch).
• go test ./core/gallery/... green.
• backend/index.yaml parses; every meta capability target (except documented locate-anything) resolves to an existing concrete.

Net change is backend/index.yaml only.

Assisted-by: claude:claude-opus-4-8 [Claude Code]

[localai-bot]

Registry verification: queried quay.io/go-skynet/local-ai-backends for every image tag these restructured metas resolve to (all metal-/cpu-/cuda12-/cuda13-/nvidia-l4t- variants of silero-vad, piper, kitten-tts, local-store, mlx, mlx-vlm, mlx-audio, mlx-distributed - latest and master - plus the diffusers/kokoro repoints). All 82 distinct tags are published. No meta now points at a missing image; the change only re-points tags that already existed in the index.

locate-anything confirmed: no image is published under *-locate-anything or *-locate-anything-cpp for any platform - it is genuinely unpublished/WIP, hence left untouched here.

[localai-bot]

Update: locate-anything is now wired up in this PR (the earlier "left for maintainer" note is superseded - confirmed with @mudler it was an accidental omission).

The meta locate-anything existed but its concrete entries were never added, so it was un-installable on every platform. Added the full concrete set (8 platform variants x release+development = 16 entries) plus the locate-anything-development meta, mirroring rfdetr-cpp exactly.

Image tags were grounded against the published quay.io tags, not assumed:

Capability	Concrete	Image variant	master- (dev)	latest- (release)
default	cpu-locate-anything-cpp	cpu-…	published	at next release
nvidia / nvidia-cuda-12	cuda12-…	gpu-nvidia-cuda-12-…	published	at next release
nvidia-cuda-13	cuda13-…	gpu-nvidia-cuda-13-…	published	at next release
nvidia-l4t / …-cuda-12	nvidia-l4t-arm64-…	nvidia-l4t-arm64-…	published	at next release
nvidia-l4t-cuda-13	cuda13-nvidia-l4t-arm64-…	nvidia-l4t-cuda-13-arm64-…	published	at next release
intel	intel-sycl-f32-…	gpu-intel-sycl-f32-…	published	at next release
vulkan	vulkan-…	gpu-vulkan-…	published	at next release

(Plus an intel-sycl-f16 concrete - the image is published and rfdetr-cpp ships the same orphan variant.) All master- images verified present on quay; latest- will publish with the next tagged release, exactly like every other backend. The -development meta resolves to the working master- images today.

Referential-integrity audit now reports zero missing capability targets across the whole index.