resolve_folder looped classify_file per file -> one CLAP forward per opaque
file. Now: run L1 (filename) for all files first, then classify ONLY the opaque
ones in a single batched forward. Extracted classify_files() as the one place
per-file audio prediction lives (CLAP batched, PANNs per-file); classify_file
and classify_folder both route through it so batching can't be lost by accident.
Validated on the jazz kit (per-index BD/HH/SN via filename, CB/P1/P2 via audio).
