v1.0.3: release fixes and GPU providers

Fix OCR word detection across newlines and tabs, correct container health checks for custom ports, and close pooled test connections to make HTTP shutdown tests reliable.

Add explicit ONNX execution-provider and device selection. CPU remains the default; NVIDIA users can run m1chl/nsfw-sherlock:v1.0.3-cuda with the NVIDIA Container Toolkit. AMD uses MIGraphX (ORT_PROVIDER=rocm is an alias) and requires a compatible MIGraphX-enabled ONNX Runtime 1.29.0 installation; no AMD image is bundled.

Release images now require passing lint, tests, vulnerability scanning, and Docker smoke tests. The checked-in lint configuration adapts the Golden configuration with documented project-specific rules. Docker builds that download canonical release models no longer install model-export dependencies.

Validation: zero lint issues with and without OCR; vet and race tests passed; all six model parity checks passed on CPU; the shutdown regression passed 100 repetitions and the full HTTP suite passed ten repetitions. CPU and CUDA images built. GPU inference has not been validated on physical NVIDIA or AMD hardware.

Images are published to Docker Hub and GHCR as v1.0.3 / latest (Linux amd64 and arm64) and v1.0.3-cuda / latest-cuda (Linux amd64).

models-v1

Canonical ONNX classifier exports for the model manifest (opset 17, dynamic batch). These bytes are the ones pinned in models/manifest.json and validated by the parity suite; sha256 matches the sidecars.

Updated dependencies

GO 1.23 and other updates for the smoother ride

Added Labels API

Added API for both web and gRPC which return labels and its values

Minor fixes and improvements

• Updated dependencies
• Minor fixes to codebase

Fixes dependencies

• Fixed second stage build dependencies

First release

First milestone

Detection works well with new model