VideoHash v0.4.0

Streaming batch API, scheduling that keeps every drive busy, and lower memory use.

Added

• generateHashesStream(for:maxConcurrentTasks:) yields each file's result as it finishes, with its original index. Cancellation stops the producer.
• Signpost intervals for each hashing stage (subsystem com.videohash), one signpost ID per video.
• Tools/classify-volumes.swift: a standalone, read-only tool that classifies drives as internal, ext-fast, ext-med or ext-slow, with an optional 5-second read test.
• hash-benchmark has a new stream mode.

Changed

• Volume-balanced scheduling (batch and stream): each freed slot takes the next file from the volume with the fewest files in progress. Before, a fixed interleaved order could leave a slow drive holding every slot while a fast drive sat idle.
• Slow-path files get an extra slot (stream): when a file falls back to frame-by-frame extraction, the batch starts one more file until that one finishes, at most doubling maxConcurrentTasks.
• ffmpeg frame extraction runs its input groups one at a time, up to 10 inputs per process, decoding with VideoToolbox (-hwaccel videotoolbox).
• Lower memory use: validated assets are reused, decoded frames are streamed into the sprite, and ffmpeg exit is awaited asynchronously. Peak memory is about 49% lower for two concurrent videos.
• Cancellation propagates through extraction, and temporary frames are cleaned up.

Hash values

• The memory work keeps every recorded hash identical. The reference parity tests (PHashReferenceParityTests) also pass with VideoToolbox decoding.

Tests

• New tests cover the stream API, volume-balanced and slow-path scheduling, unsupported colour conversions, extraction lifecycle and fallback.