Lexicon v0.5.17

v0.5.17 - Model Memory Mapping, Context Bounds & Paragraph-Level LRU Caching

• Model Context Window Bounds (n_ctx = 4096): Reduced local LLM context window from 8,192 to 4,096 tokens, cutting KV-cache RAM/VRAM allocation by 50% (~300–600 MB savings) and doubling prompt evaluation speed.
• Memory Mapping (use_mmap = True): Configured OS-level memory mapping for GGUF model weights for near-instant model loading (< 0.5s) and instant physical RAM page reclamation during Tier 1 offloads.
• Frontend Chunking Budget Alignment: Updated TRANSFORM_INPUT_BUDGET and TRANSFORM_CHUNK_BUDGET to 1,800 tokens (~7,200 chars / ~3–4 full paragraphs), with 2,048 max output token headroom.
• Sub-Microsecond 64-Bit Hashing: Implemented synchronous fnv1a64(str) hashing (< 20 nanoseconds per block) for instant LRU cache key calculations without Promise overhead.
• Context-Aware Predecessor Keys: Paragraph cache keys incorporate the preceding paragraph's suffix (suffix(A, 64)). Deletions, additions, or merges automatically update predecessor keys and invalidate downstream boundary rules without requiring complex cascade-invalidation code.
• Differentiated Proofread Triggers: Live typing auto-checks evaluate only modified/uncached paragraphs; clicking the manual Proofread button or pressing Ctrl/Cmd + Enter bypasses the cache to perform a 100% full-document scan.
• Non-Blocking UI Startup: Backend sidecar startup runs on a background thread so the window and index.html load instantly (< 1ms) without blocking the OS UI thread.
• Instant Window Hiding on Quit: Main window hides (window.hide()) on line 1 of the Quit handler so the desktop window vanishes instantly while process cleanup completes in the background.
• Native Window Theme & Boot Optimization: Configured "theme": "Light" and "backgroundColor": "#f7f7f5" in tauri.conf.json, and set Google Fonts to non-blocking loading in index.html for instant frame-1 boot screen rendering.
• Detached Process Cleanup & Console Window Suppression: Added silent Java process termination on quit (taskkill /F /IM java.exe) and global Windows CREATE_NO_WINDOW / SW_HIDE subprocess overrides to eliminate console window flashes.
• Release Workflow Prioritization: Updated .github/workflows/release.yml to prioritize extracting release notes directly from CHANGELOG.md first.