v0.11.2

Release Notes — v0.11.2 (2026-04-11)

This release marks the culmination of the "Architecture Unification & Pipeline Hardening" cycle. It consolidates all improvements made since version 0.11.0 (2026-03-28), transitioning the project into a single, high-precision media workstation.

🚀 Key Highlights (Since 0.11.0)

🏗️ 1. Unified Media Architecture (v0.11.1 Milestone)

The project structure has been completely overhauled to eliminate redundancy and simplify deployment:

• Binary Consolidation: Merged vid_hevc/vid_av1 into a single vid binary and img_hevc/img_av1 into a single img binary. Users can now toggle between strategies using --codec <hevc|av1>.
• Workspace Modernization: Core Logic moved to a managed crates/ workspace (shared_utils, vid, img), ensuring sub-100ms incremental build times and shared dependency management.

🎞️ 2. Video Pipeline Hardening & "Stage 5" Exploration

• Stage 5 Downward Exploration: After finding the optimal CRF, the system now performs a high-precision "boundary push" (0.01 delta) to extract every possible bit of quality until the file size matches the input budget.
• HEVC Ultimate Simplification: Replaced over 360 lines of candidate-ranking logic with a streamlined two-step encode: Search (slow) → Render (slower). This provides identical archival quality with significantly higher predictability.
• Adaptive Quality Gates: Replaced static VMAF/PSNR floors with per-file adaptive baselines. The system now dynamically detects and rejects quality regressions based on the source's native fidelity.

🖼️ 3. JXL (JPEG XL) Next-Gen Exploration

• Adaptive Probing: Introduced a profile-driven exploration model (MicroAdjust to CeilingSweep) that adapts its search strategy based on the image's "oversize pressure."
• Binary Search Convergence: Replaced heuristic "jogging" with an high-precision binary search (0.0001 delta) for sub-0.01 distance values.
• Continuous Quality Sweep: Implemented a JXL-specific "Stage 5 Sidekick" that Steps downward to explore higher visual fidelities until the first size regression.

🧠 4. Intelligence & Decision Flow (Loop Intent v3)

• 7-Layer Loop Intent System: A multi-tiered heuristic+ML tree that accurately distinguishes between "sticker-like" loops and "narrative-like" clips.
• pgvector HNSW Integration: Migrated KNN similarity search from in-memory Euclidean math to PostgreSQL's HNSW vector index, enabling millisecond-level retrieval across thousands of labeled samples.
• Level 4 Inference Logging: The system now logs every internal signal and final decision to a structured forensic table, enabling data-driven tuning of the Active Learning loop.

🛡️ 5. UltraHDR & Media Integrity

• Gainmap Resilience: Implemented a four-strategy "candidate recovery" system for Google UltraHDR JPEGs, supporting malformed MPF segments and XMPF identifiers.
• Baseline-Safe Synthesis: Corrected transfer curve tagging (PQ/ST2084) for JXL HDR synthesis, ensuring proper "glow" and brightness in macOS Preview and web browsers.
• Path Safety Shield: Hardened all external tool calls (FFmpeg, ExifTool, cjxl) with standardized ./ prefixing and shell-escaping to defend against malicious filenames and protocol hijacking.

🔧 Core Maintenance

• 100% Health Milestone: The entire codebase (Rust, Python, Shell) now passes the strict check_all.py quality suite with zero warnings/failures.
• Logic Unification: Standardized all efficiency metrics to strictly less than (<).
• Dependency Stabilization: Reverted the production main branch to stable crates.io sources (fixing API breakages in quick-xml and image).
• macOS App Wrapper: Re-engineered the entry-point script with robust Python discovery and improved AppleScript escaping for drag-and-drop workflows.

---

For a full itemized list of every commit and fix, please refer to the primary CHANGELOG.md.