Track what changed without parsing raw release markdown first.

Added

• macOS Apple Silicon OFX installer for `v0.7.3` as a signed `.pkg` plus a notarized `.dmg` wrapper.
• Packaged model inventory metadata for macOS bundles so diagnostics and release validation follow the same contract style used by Windows packaging.

Changed

• macOS release packaging now signs with Developer ID, notarizes the public DMG, and bundles the MLX bridge ladder through `2048px`.
• Matte-only OFX paths can skip foreground materialization and transport when the selected output mode only needs alpha.

Fixed

• The OFX runtime panel keeps the last completed frame timing visible during playback instead of clearing it during in-flight work.
• Cross-platform bundle inspection now detects Windows OFX runtime-server requirements correctly even when the bundle is inspected from macOS.
• Model artifact checks distinguish usable artifacts from missing, invalid, or Git LFS placeholder files.

Added

• No new public features were added in this patch release.

Changed

• Moved `Input Color Space` into the `Key Setup` section in the OFX plugin panel for a more coherent setup flow.
• Aligned the current Windows codebase with the repository lint and formatting gates used by the release workflow.

Fixed

• Preserved the required Windows include order in the OFX runtime server build so the release build stays green under MSVC.

Added

• Windows RTX Lite and Windows RTX Full installers for the supported NVIDIA RTX 30 series and newer path.
• Certified RTX artifact manifests and packaged compiled contexts for the validated Windows RTX model ladder.

Changed

• The Windows RTX quality ladder now exposes only the supported public steps: Auto, Draft (512), High (1024), Ultra (1536), and Maximum (2048).
• Windows RTX packaging now requires a certified artifact manifest and matching model/context hashes before an installer can be produced.
• README and help content now describe the Windows RTX Lite vs Full policy consistently.

Fixed

• Windows RTX installers no longer package stale or manually copied model sets as if they were releasable.
• The Windows RTX Full release now includes the certified 2048 model and compiled context set.
• Bundle validation and doctor reporting now stay aligned with the packaged Windows RTX tracks.

Added

• Advanced Runtime Status in the OFX panel for deeper diagnostics without crowding the main status view.
• Advanced control subgroups for Interior Detail, Matte Cleanup, Processing, and Runtime tuning.

Changed

• The canonical Windows release flow now publishes the validated RTX track by default.
• Runtime panel wording and help content now match the actual OFX behavior more closely.

Fixed

• OFX panel organization is easier to scan and less mixed for advanced users.
• Windows RTX packaging was validated with the full shipped model set through 2048.

Changed

• Default quality mode is now Draft (512) instead of Auto. Auto mode selected resolution dynamically based on hardware, which could cause unexpected first-render delays on lower-end GPUs. Users can still select Auto or any higher quality tier from the dropdown

Added

• Dual-output RGBA multiplexing as the default output: despilled foreground in RGB channels, core matte in Alpha, allowing signal routing with native host tools at zero cost
• Output Mode dropdown with five modes: Processed (RGBA mux), Matte Only, Foreground Only, Source + Matte, and Linear Premultiplied RGBA
• SharedFrameCache: thread-safe LRU ring-buffer (4 slots, shared_mutex) keyed on a FNV-1a hash of frame pixel data, inference parameters, model path, and screen colour — duplicate OFX nodes retrieve pre-computed passes with zero redundant inference

Fixed

• Removed debug stderr traces from the three OFX C entry points (OfxSetHost, OfxGetNumberOfPlugins, OfxGetPlugin) that were writing to the host process in all build configurations

Changed

• High-resolution ONNX models (1536, 2048) now use static batch dimensions, eliminating TensorRT int32 tensor volume overflow during engine compilation
• Simplified model naming: removed redundant `_static_` prefix convention
• Disabled TensorRT subgraph dumping in the model compiler to prevent generating temporary files

Fixed

• TensorRT engine build failures for 1536 and 2048 resolution models caused by dynamic shape propagation
• Git LFS tracking gaps for model files (.pth, .safetensors, .mlxfn)

Added

• Added a dedicated `Frame Times` runtime panel field so per-frame timing stays visible without overloading the status line.
• Added strict Windows TensorRT RTX execution probes for the packaged FP16 artifacts at 512, 768, 1024, 1536, and 2048 in the doctor diagnostics.

Changed

• Changed the default OFX render and prepare timeouts to 60 seconds and 300 seconds, and aligned the runtime server idle/session timeout to 300 seconds.
• Changed Windows packaging and validation to stage the full curated ONNX Runtime DLL set and report expected backend omissions as informational output.

Fixed

• Fixed fixed-quality TensorRT RTX requests so `Ultra (1536)` and `Maximum (2048)` fail explicitly on exact engine compilation failure instead of silently downgrading.
• Fixed repeated TensorRT RTX compile thrash by caching per-session quality compile failures by backend, device, quantization mode, and model bundle.
• Fixed release/test staging so the packaged runtime and automated validation pick up the required backend DLLs consistently for RTX and DirectML variants.

Added

• **macOS Installer Alignment:** The macOS installer now gracefully closes DaVinci Resolve before installation, clears old logs, and reopens Resolve automatically upon completion (matching Windows behavior).
• Strict Windows universal execution probes in `doctor` that verify session creation and a synthetic frame run without CPU fallback.
• New doctor summary fields for execution readiness, preferred backend, and preferred model on Windows universal GPU paths.

Changed

• DirectML now steers OFX users to `FP16 (Full)` instead of allowing the previously unvalidated `INT8 (Compact)` path.
• Windows universal sessions now pin the free `batch_size` dimension to `1` during session creation for DirectML and WinML.

Fixed

• The runtime now distinguishes provider detection from real model execution when evaluating Windows universal GPU readiness.
• The DirectML release package is validated during packaging to confirm the staged runtime actually reports DML support.

Added

• Root-level `help/` directory for user-facing OFX guides, tutorials, support information, and troubleshooting.

Changed

• Kept `docs/` focused on development specification, architecture, engineering guidelines, and release procedures.
• Updated repository references so README and contributor-facing documentation point to the correct user-help location.

Fixed

• OFX **Help & Docs** buttons now open the correct published guides on GitHub.
• Removed stale planning, brainstorm, release-note, and one-off deploy artifacts that no longer belong in the live documentation tree.

Added

• Topic-specific Help & Docs buttons that open practical guides for setup, quality, Alpha Hint, Recover Original Details, tiling, and Resolve workflows.
• Dedicated OFX panel and Resolve tutorial guides in the project documentation.

Changed

• Standardized quality labels in the OFX panel to `Draft (512)`, `Standard (768)`, `High (1024)`, `Ultra (1536)`, and `Maximum (2048)`.
• Moved runtime/version information to the top of the panel for faster diagnosis.

Fixed

• Cleared stale fallback warnings after switching back to a supported quality.
• Cleared CorridorKey user logs during Windows installation to avoid carrying stale diagnostics across installs.
• Replaced non-actionable help rows with direct documentation links that work inside Resolve.

Added

• Regression coverage for fixed-quality fallback candidates and OFX output mode semantics.
• Regression coverage for cached source passthrough erosion footprint behavior.

Changed

• Renamed the original-detail recovery controls to Recover Original Details with clearer Details Edge slider labels.
• Moved runtime diagnostics higher in the OFX panel and expanded tooltips for tiling, color space, and matte guide behavior.

Fixed

• Restored downgrade fallback from fixed quality modes when the exact artifact exists but backend preparation fails.
• Made Processed match the runtime core's linear premultiplied RGBA behavior and kept FG+Matte as the explicit comp-oriented output.
• Removed per-frame heap allocation from the source passthrough hot path.
• Aligned Windows ONNX Runtime packaging/header usage and preserved backend suffixes in packaged artifacts.

Added

• Expanded NVIDIA TensorRT support to include RTX 20 Series (Compute Capability 7.5).
• Implemented robust `device_index` mapping to prevent DirectML from defaulting to integrated graphics (iGPU) in multi-GPU systems.
• Centralized all ONNX Runtime execution provider identifiers into a single constants header, preventing string-matching bugs across platforms.
• Added diagnostic logging on macOS (writes to `~/Library/Logs/CorridorKey/ofx.log`), matching the Windows logging infrastructure.

Changed

• Reverted memory pattern optimizations (`DisableMemPattern`) for DirectML to align with Microsoft's official driver recommendations, preventing memory allocation panics.
• Adjusted ONNX Runtime graph optimization levels (`ORT_ENABLE_EXTENDED` instead of `ORT_ENABLE_ALL`) specifically for DirectML to prevent CPU-centric memory layout reordering that caused `E_INVALIDARG` MatMul crashes.
• Clarified fallback policies to ensure models cleanly fall back to CPU execution for unsupported nodes instead of aborting the session.

Fixed

• Fixed an issue where users with RTX 2080 and other Turing GPUs were incorrectly forced onto the slower DirectML fallback path.
• Fixed a major initialization bug where DirectML would attempt to run heavy AI models on underpowered integrated graphics despite the UI detecting a dedicated GPU.
• Fixed `0x80070057 (The parameter is incorrect)` crashes during FP16 MatMul operations on AMD and Intel hardware.
• Fixed the OFX plugin on macOS incorrectly selecting the CoreML backend instead of MLX on Apple Silicon, which caused High (1024) and above quality modes to fail with missing artifact errors.

Changed

• **Centralized Versioning**: The `CMakeLists.txt` is now the single source of truth for the project version, and `version.hpp` is automatically generated during the build.
• **Robust Build System**: Added `scripts/build.ps1` to automatically manage the MSVC environment, preventing compilation errors in non-standard terminals.

Fixed

• **Cascading Errors**: Added session liveness checks before reusing engines to prevent stale session crashes.
• **Flickering UI**: Replaced infinite re-render loops (`kOfxStatFailed`) with a professional "Source Passthrough" bypass in case of critical processing or engine creation failures.