Berry AI Studio v0.2.1

Large-Library Performance & Gallery Virtualization (Milestone 13)
Version 0.2.1 delivers comprehensive performance optimizations across gallery rendering, thumbnail caching, database queries, filesystem indexing, and application bundles to ensure fluid navigation with tens of thousands of local images:

⚡ Virtual Gallery & Layout Optimization
Responsive Fixed-Width Columns: Maintained stable card widths during viewport resizing, dynamically adjusting column counts instead of stretching images.
Frame-Coalesced Scrolling: Coalesced scroll events through requestAnimationFrame, limiting reactive updates to at most one per rendered frame.
Per-Column Binary Search: Replaced full dataset filtering in Waterfall view with per-column binary search for visible items.
Filter-Aware Stack Summaries: Kept stack counts, heroes, and expanded members aligned with active text, structured, and semantic search filters.
📦 Incremental Data Delivery & Deferred Payloads
Bounded 400-Item Paging: Replaced full table materialization with bounded 400-item SQLite queries returning exact filtered window counts.
Deferred Raw Metadata: Omitted heavy prompt strings and ComfyUI workflow JSON from gallery page IPC; loaded detailed records on-demand with a 64-entry LRU cache.
Stale Response Invalidation: Discarded in-flight page queries when navigation filters, folders, or sort orders change.
🖼️ Smart Multi-Tier Thumbnail Pipeline
Persistent Thumbnail Manifest (Schema v11): SQLite-backed manifest tracking image IDs, modification times, size tiers, and codecs with rate-limited access tracking.
Configurable 2 GB Cache Budget: Automated LRU eviction maintaining disk usage within the configured budget during generation and background migration.
Rendered-Size Tier Selection: Selected the smallest sufficient thumbnail tier covering rendered card dimensions across Grid, Waterfall, and compact Table rows.
Existing Tier Reuse: Reused existing larger cached tiers for the same revision to eliminate redundant decoding.
Cancelable Decode Priority: Prioritized visible thumbnails before speculative look-ahead, using monotonic viewport generations to drop stale decoding jobs.
📂 Streaming Ingestion & Filesystem Watcher
Native Directory Watcher: Registered long-lived platform watchers with a durable SQLite journal, 750ms quiet-period coalescing, and targeted path reconciliation.
Streaming Folder Scans: Streamed media entries into bounded database batches without keeping full directory trees in memory.
Throttled Scan Progress: Coalesced scan progress IPC events to at most one per 64 files or 100 ms.
Startup Scan Cooldown: Added a six-hour default cooldown preventing repetitive full-tree walks on app launch.
🚀 Bundle & Resource Reduction
Async Modal Bundle Splitting: Dynamically imported 18 infrequent modals and drawers, reducing the entry JavaScript bundle by ~103 KB and CSS by ~87 KB uncompressed (~25 KB gzipped).