v2.7.2

[2.7.2] - 2026-06-08

Core System Updates

• Audio & STT: Implemented the Cohere STT service supporting FLAC, MP3, OGG, and WAV via the /v2/audio/transcriptions endpoint using the cohere-transcribe-03-2026 model.
• Adapter Lifecycle: Refactored the adapter reload system to resolve initialization deadlocks (init-storm lock re-entry) and stale datasource connections during configuration reloads.
• Config Management: Hardened config_manager.py with improved relative path resolution for nested imports and more robust path-traversal guards.

Chat-app & UI Improvements

• Model Picker: Relocated the model picker from the chat header to an inline pill within the message input bar.
• Threading: Enabled independent model selection for individual thread composers, ensuring each thread reply uses its own selected model state.
• UI Refinements: Extracted the shared model picker component with improved keyboard navigation (Escape dismissal) and focus management; improved mobile visibility and adapter-load synchronization.

Bug Fixes & Technical Improvements

• Language Detection: Enhanced detection accuracy for Persian (preventing misclassification as Arabic) and standardized confidence threshold logic.
• Fault Tolerance: Standardized the FaultTolerantAdapterManager facade by implementing missing delegations for public methods.
• Service Hardening: Decomposed the language detection pipeline into focused heuristic and voting modules; added regression tests for datasource config isolation.

Documentation & Configuration

• Architecture Docs: Updated language-detection-architecture.md to reflect the latest weighted voting formulas and Redis key-namespace logic.