[zhangzhenghao]

Summary

• Add clap search DIRECTORY -m MODEL (--text TEXT | --audio FILE) for recursive audio retrieval.
• Rank WAV/MP3/FLAC files (case-insensitive extensions) by cosine similarity and emit a JSON array of {path, score}, highest first, with deterministic path tie-breaking.
• Support -k/--top-k (default 10), -t/--threads, and --device; load the model once and retain only the best k scores rather than corpus embeddings.
• Skip undecodable candidates with stderr warnings; reject empty/unsearchable directories and propagate model/inference failures. Do not follow directory symlinks.

Usage

clap search ./audio -m model.gguf --text "a dog barking" --top-k 5
clap search ./audio -m model.gguf --audio query.wav --top-k 5

Text search requires a model with text/audio embeddings in a shared space. Audio-only models can be used with --audio. Each invocation scans and encodes the directory; no persistent index or embedding cache is introduced. README is unchanged.

Verification

• Removed the newly added search CLI test and its CTest registration at user request; existing tests are unchanged.
• Rebuilt the Linux static configuration after removal: 35/35 existing CTest tests passed.
• The search implementation is unchanged from the previously reviewed commit, whose cross-platform CI passed. CI for the new head is tracked separately.
• git diff --check passes.