Virgil Lemma foundations

[Snider]

@coderabbitai summary

Summary by CodeRabbit

• New Features

  • Qwen 2/3 and Qwen 3.6 model support; new adapter with buffered and streaming generation.
  • Block‑prefix cache service and memvid bundle index for faster prefix restores.
  • Agentic memory: wake/sleep workflows, state bundles and memvid integration; session‑state artifact export.

• Improvements

  • Device‑aware memory planner; expanded chunked generation, prompt‑cache warm/restore and KV snapshot flows.
  • Build/toolchain updated (C++23) and macOS deployment target raised.

• Documentation

  • Extensive new/updated docs: architecture, runtime, inference, memory, MoE, training and benchmarks.