[solderzzc]

What

Implements the Cosmos3Omni generation branch (*_moe_gen, the second tower of the Mixture-of-Transformers) and wires cosmos3_omni into the VLM factory. Stacked on #107 (already merged), so this is against main.

• SharpAICosmos3OmniGenTower and joint attention: understanding tokens attend causally to understanding tokens only; generation tokens attend bidirectionally to [understanding K/V; generation K/V]; per-tower RoPE (unified 3D mRoPE, interleaved sections).
• Heads: proj_in/proj_out/time_embedder, audio head, domain-aware action linear. The audio and action heads are only built when sound_gen / action_gen is set.
• Wrapper SharpAICosmos3OmniModel (und: Qwen3VL, gen, gen_heads) with generateVelocity(...).
• Sanitizer now maps all 1165 checkpoint keys of the reference layout (including the 216 add_* attention keys); unknown keys throw instead of being dropped. The earlier rejection of _moe_gen is removed.
• Config errors now throw instead of silently falling back to Nano defaults.
• cosmos3_omni is added to the wired VLM keys (7); conformance expectations regenerated.

Tests (air-gapped, nothing downloaded)

66 Cosmos test cases across Configuration, GenTower, Heads, Model, Registration, Sanitize and WeightFiles, plus the VLM wiring and conformance suites. The air-gap test builds the model through the real registry with deterministic synthetic weights and checks that the model's parameter key set equals the sanitized key set. No .github/workflows or Package.swift changes.

What is not verified

• No numerical parity with the PyTorch reference (no reference outputs available); tests are structural (shapes, finite values, understanding output independent of generation input).
• Not loaded with a real Cosmos3-Nano checkpoint.
• The real processor_class is unconfirmed; Qwen3VLProcessor is forced for cosmos3_omni.
• Deepstack injection, generation mRoPE position construction from frame rate / frame count, and image/video features (only precomputed embeddings are accepted) are not implemented.
• Sampler loop, VAE and sound tokenizer are out of scope.

🤖 Generated with Claude Code