v4.1.0: video reaches the projector, fast-forward on the branch tree

Video reaches the projector

The multimodal prefill takes a video as one input: the sampled frames of a clip with their timestamps and optional notes, laid out as llama.cpp's video helper lays them out, on liblloyal v3.1.0. Frames fuse in pairs on Qwen-VL projectors and the tokens match the server's video path. Images and audio are unchanged beside it, and all three kinds share the one typed input marshaller.

Fast-forward on the branch tree

_branchFastForward(parent, child) exposes liblloyal's fast_forward: a child's state becomes its parent's, the parent keeps its handle, and no cells move. The SDK uses it to write a warm multimodal turn on a fork and fast-forward the trunk to it, so a turn that fails for room costs only the fork.

A commit the decode refused intact now restores the branch's revision along with its sampler, grammar and metrics, so a later fast-forward is not refused for a change that never landed; a poisoning failure keeps its bump. Duplicate handles in one commit are refused before any token is accepted.

Packages

Thirteen platform packages, each built and tested on its own runner: macOS arm64 and x64; Linux arm64 and x64, with CUDA and Vulkan variants; Windows arm64 and x64, with CUDA and Vulkan variants. The Linux x64 CUDA flavor also ships as a signed download manifest for CUDA 12.9. Every package passed the L4 GPU gate before publishing.

The SDK peer range now names the one SDK major this binding implements, >=5.0.0-0 <6.0.0.

Ships with lloyal-ai 1.19.0, alongside liblloyal v3.1.0 and hdk v6.1.0-rig.

What's Changed

• feat(multimodal): video input by @lloyal-research in #60

Full Changelog: v4.0.0...v4.1.0