fix(config): only inject llama.cpp serving options on the llama.cpp path

[localai-bot]

Problem

Starting a non-llama backend model (e.g. longcat-video-avatar-1.5) in distributed mode fails at LoadModel:

unknown model option(s): cache_reuse, parallel

SetDefaults injects the llama.cpp server options cache_reuse (ApplyServingDefaults) and parallel (ApplyHardwareDefaults, re-applied per selected node by the distributed router) onto every model config, backend-blind. Every other backend silently ignores options it doesn't understand, so this was harmless until longcat-video, which is the only backend that strictly validates its options and hard-rejects unknowns.

Fix

Core - gate the injections by backend. New UsesLlamaCppServingOptions(backend) allow-list (llama-cpp + the empty/auto-detect case that resolves to llama.cpp from a GGUF file, mirroring how llamaCppDefaults is registered). It follows the existing UsesLlamaSamplerDefaults precedent for llama-only defaults. Applied at all three sites:

• cache_reuse in ApplyServingDefaults (single-host)
• parallel in ApplyHardwareDefaults (single-host)
• parallel in applyNodeHardwareDefaults (distributed router, per selected node) - backendType threaded through

The typed NBatch field is intentionally left ungated: it's a proto field every backend simply ignores, which is exactly why batch never appeared in the error alongside the two option strings.

Backend - defense in depth. longcat-video now warns-and-ignores unknown model options and request params via a testable select_known_options helper, matching the other LocalAI Python backends, so a future server-injected option can't break loading again.

Tests

TDD throughout (each new behavior watched fail first):

• core/config: new specs asserting cache_reuse / parallel are not injected for a non-llama backend, and still are for llama-cpp
• core/services/nodes: new spec asserting the node-tuning path skips parallel for a non-llama backend
• backend/python/longcat-video: new select_known_options unit tests

All affected suites pass locally (304 config, 336 nodes, 9 longcat util specs); golangci-lint --new-from-merge-base clean.

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