[zhangzhenghao]

Summary

• Replace hard-coded dummy CLI embeddings with native GGUF model inference for text and audio; preserve prompt/file batching, output formats and true cosine similarity.
• Load a real model once for clap serve, expose real text embeddings and projection dimensions, validate requests and serialize inference on the shared model handle.
• Remove the standalone clap-model example executable, migrate its callers to clap embedding, and update documentation.
• Add deterministic stdlib-generated GGUF fixtures with independent numerical oracles so default CLI/HTTP tests exercise actual inference without downloads or PyTorch.

Verification

• Rebased onto current main.
• Static Release build and CTest: 26/26 passed.
• Shared Release build and CTest: 26/26 passed.
• Default CLI/HTTP integration tests verify numerical outputs, different inputs, batching, dimensions, concurrent calls and failure handling. Two optional model-fixture CLI cases are skipped in default CTest.
• Prior implementation validation also exercised optional model-fixture CLI cases and staged installation; fresh latest-main validation above covers both complete default suites.
• CUDA/ROCm inference is not claimed; these interfaces use the existing native CPU/optional Vulkan model runtime.