Swap the second teacher from GPT 5.6 Sol to Moonshot Kimi K3

[ai-hpc]

Replaces the OpenAI GPT 5.6 Sol teacher with Moonshot Kimi K3 (moonshotai/kimi-k3 on OpenRouter, kimi-k3 on yunwu). The pinned teacher pair is now Claude Fable 5 + Kimi K3.

Why Kimi K3 (research)

On the published GPU-kernel-optimization benchmark, Kimi K3 performed competitively with Fable 5 and substantially outperformed GPT 5.6 Sol (and Opus 4.8 / GPT 5.5) at profiling, rewriting, and benchmarking CUDA/Triton kernels — it even ships its own Triton-like compiler, MiniTriton, and produced a kernel ~14.8× faster than optimized PyTorch on H200. For a Triton-kernel dataset generator, that kernel-optimization strength is exactly the axis that matters.

Honest caveat: K3 trails Sol/Fable on general coding, and the 14.8× figure is a single vendor-run benchmark not yet independently reproduced. The swap is justified specifically for the kernel-optimization workload here, not as a blanket "best coder" claim.

Sources: Tom's Hardware, Windows News (14× kernels), OpenRouter model page, Kimi K3 tech blog.

What changed (provenance-correct)

• Real moonshot provider (not Kimi masquerading as openai), so the manifest's pinned allowed_teachers records the true teacher. MOONSHOT_TEACHER_MODEL = kimi-k3.
• policy.py / gateways.py: model maps, SUPPORTED_PROVIDERS = {anthropic, moonshot}, yunwu/OpenRouter slugs. Kimi has no -sol-style variant, so the logical/routed slug distinction collapses to one slug; dated OpenRouter build suffixes are still stripped.
• Env YUNWU_MODEL_OPENAI → YUNWU_MODEL_MOONSHOT; sparkproof-yunwu-probe pins kimi-k3.
• training_cot: recover_openai_winner_cot → recover_moonshot_winner_cot. Kimi returns plaintext reasoning, so the Fable CoT-recovery path is now a rare fallback (docstrings/metadata de-Sol'd; sol_winner_provider → winner_provider).
• Default providers ["anthropic","openai"] → ["anthropic","moonshot"] across the CLIs and multi_candidate.
• Docs: README, CONTRIBUTING, .env.example updated. Historical CHANGELOG entries left intact as an accurate record of past releases.

⚠️ Breaking / coordination

• gpt-5.6 / gpt-5.6-sol are no longer accepted teacher models — existing Sol bundles will fail verification against the new policy and must be regenerated with the new pair.
• Cross-repo: SparkDistill's teacher/providers.py must be updated to the moonshot provider key to stay in sync (the policy comment references it).

Tests

355 pass, ruff clean. New test_kimi_k3_replaces_gpt_sol_as_second_teacher locks the swap (Kimi pinned, openai/gpt-5.6-sol rejected). ~24 files re-pointed to the new provider/model.