v9.3.32

v9.3.32 — Research-Driven Hardening

Added

• Schema-validated structured outputs for all LLM classify calls (Ollama format + OpenAI response_format)
• Centralized ai_request() helper with retry, backoff, API key redaction, and normalized errors
• Full tag library backup/restore with SHA-256 manifest (Settings Hub + CLI)
• TagSpaces sidecar import/export wired into the Tag Library GUI
• Windows Shell properties merged into file scan metadata
• Qt translator at startup + language picker in Settings Hub
• Accessible names/descriptions on major widgets for screen readers
• Async debounced Tag Library search (300ms settle, QThread worker)
• Bad-extension cleanup analyzer (content/extension mismatch detection)
• FTS5-backed tag name and entry filename search with ilike fallback
• Zone.Identifier download provenance auto-populated on Windows
• Mark-of-the-Web risk label in Tag Library file preview
• Release audit tooling (SBOM, license inventory, vulnerability report)
• Rendered UI smoke screenshot tests (3 new panels)
• Windows-friendly Python dev task runner
• Frozen-build Qt-binding isolation preflight

Changed

• [full] extra no longer pulls cmake/dlib/face_recognition/nexaai — use [face] or [nexa]
• Bootstrap no longer tries --break-system-packages
• All SQLite connections use WAL mode + busy_timeout

Removed

• Tracked debug outputs from repo root