fix: lazy load node-llama-cpp for no-op embed

[aceclaw826]

Summary

Avoid importing node-llama-cpp until QMD actually needs a local LLM/model operation.

This keeps no-op embed/status-style paths lightweight. In particular, store.embed() already checks for pending documents before calling into the LLM, but the previous static import meant the native node-llama-cpp module was still loaded as soon as llm.ts was imported.

This matters for OpenClaw multi-agent gateways where boot/update cycles may call qmd embed frequently and most runs often have no missing embeddings.

Changes

• Convert node-llama-cpp value imports to type-only imports.
• Add a memoized dynamic loader for node-llama-cpp.
• Load getLlama, resolveModelFile, LlamaLogLevel, and LlamaChatSession only from code paths that actually need them.
• Add a regression test proving a no-op embed does not import node-llama-cpp.

Related

This addresses part of openclaw/openclaw#72144: avoiding heavy native/LLM initialization when all content hashes already have embeddings.

Tests

Passed:

corepack pnpm exec vitest run test/llm-lazy-load.test.ts test/sdk.test.ts -t "lazy node-llama-cpp loading|store.embed rejects invalid batch limits|store.embed forwards batch limit options" --reporter=verbose

Also checked:

corepack pnpm exec tsc -p tsconfig.build.json --noEmit

That currently fails on main with an unrelated existing error:

src/store.ts(2142,22): error TS2339: Property 'transaction' does not exist on type 'Database'.