H3 Add Latent Guide (aligned latent rows) [experimental]

EXPERIMENTAL. Hands the model a whole video LATENT as an ALIGNED GUIDE — H3's keyframe condition rows, which pack right after the text on the target's own spatial grid and time axis, arrive nearly clean (noise_aug 0.999) and are never denoised. Core's Add Guide node takes IMAGES and VAE-encodes them; this one takes the latent you already have, so a latent-space init (H3 Temporal Insert's lerp, a bank, a previous pass) never makes a round trip through pixels. Why bother: measured on the trainer's ruler (the aligned-guide A/B/C study, 2026-08-28), the same lerp movie read as aligned guide rows scores 0.41-0.43 of the lerp baseline with NO adapter at all, where a LoRA trained on the in-x_t arrangement gets 0.46: the base model can read a correctly-positioned guide and does not need to be taught to. Keeping the lerp in x_t as well buys nothing (it measures as the in-x_t arm again), so the intended pairing is H3 Temporal Insert with init_mode=noise into the sampler and a SECOND one with init_mode=lerp into this node, on a FULL schedule. What that buys DECODED, and what it costs (one clip, same-seed rerun band measured 2026-09-01): the guide keeps the inserted span at the source's speed, as the Motion Adapter recipe does, and invents less in the frame; it also costs about 12% of the source's detail (0.88 of source sharpness, rerun band 0.003), and on playback the guide arms were judged WORSE than the adapter recipe for exactly that reason. Read this node as a 'protect what is already right' instrument, not as a sharper de-rope. token_idx is the LATENT-TOKEN offset of the guide's first token inside the target grid (0 = a full-length guide starting at the top). Tokens cover (1, 4, 4, 4, 4) pixel frames per 5, and this node converts to the pixel-frame anchor core wants for you. Clock-correct by construction: under H3 True Clock or H3 DyRoPE the guide's rows are retimed to the exact t of the target rows they duplicate, and a guide that cannot be mapped RAISES instead of silently sitting on the stock grid (which, on a dense clip, is hundreds of rotary units away from what it is supposed to align with). Costs one target-video-worth of sequence length per full guide — roughly a doubling of the video rows, which is superlinear in step time. A short guide over the span you care about costs proportionally less; that is what token_idx is for. Do NOT pair with the published Motion Adapter: it was trained on the in-x_t arrangement and measured WORSE with a guide than the base model is (0.87 vs 0.41 of the lerp baseline).