Port Trinity-Nano-Base (afmoe)

[DarthCeltic]

Summary

• Adds arcee-ai/Trinity-Nano-Base to the most_models_ported track (identity afmoe), introducing the Afmoe execution family to the board — Arcee AI's Trinity model line (fine-grained MoE, 56 layers, 128 experts / 8 active + 1 shared, 131K context via sliding-window + full attention mix, 6.12B total params).
• No converter fix was needed: AfmoeForCausalLM is already a correctly-registered architecture class in this repo's convert_hf_to_gguf.py, so this converted cleanly on the first attempt with the stock, unmodified converter — same clean-conversion category as the earlier smallthinker port.
• Self-converted to Q8_0 GGUF (1057 tensors, 6.52 GB, sha256=a52296d493fbdb810008a638eda4ce7f59717ddd4282ba2750518c7cefd36872), hosted on Hugging Face (darthceltic85/trinity-nano-base-gguf) since it exceeds GitHub's 2 GB release-asset limit.
• License is OpenMDW-1.1, a genuinely permissive open license — confirmed by reading the actual LICENSE file in the source repo rather than assumed from the license: other tag.
• Verified live: model loads cleanly on the CPU backend (arch = afmoe, 128 experts / 8 used, sliding-window + full attention layer mix, fused Gated Delta Net enabled, 4223-node compute graph, 1 split) and real llama-perplexity against WikiText-2 raw gives PPL = 10.5908 +/- 1.92727 — a good, in-range result.

Test plan

• benchmark_config in the claim JSON exactly matches .github/ci/benchmark_config.json's new trinity_nano_base entry
• Source revision is a valid 40-char hex commit SHA
• RECIPE.md exists and honestly documents that no converter fix was required
• All new/modified JSON files are valid JSON
• Model loads and produces a real, in-range perplexity result (not fabricated)