fix: sample tagging returns same tags for all samples

Commit notes

Root causes fixed: 1. Invalid OpenAI model name 'gpt-5-nano' changed to 'gpt-4o-mini' - The invalid model caused API calls to fail, falling back to rule-based tagging

2. Fixed interpretEmbedding function in ai-tagging.ts: - Changed truthiness check for spectralCentroid to nullish check (0 is valid) - Fixed attack sharpness calculation to use features.attackTime directly instead of L2-normalized embedding values - Fixed harmonic content to derive from spectral bandwidth/centroid ratio

3. Added attackTime and decayTime fields to AudioFeatures interface - These temporal envelope values are now stored as direct features - Previously only available in L2-normalized embedding vector

The rule-based tagging now correctly uses the actual audio features instead of falling back to default values, producing varied tags based on each sample's unique characteristics.

Co-authored-by: armin.naimi <[email redacted]>