feat: Splash inference engine, CoreAI/Muse-Glimmer recipe, and model registry pins

[unfoundbox]

Summary

• Register Inco AI Splash runtime (fused Metal kernels, 8-bit KV-cache decode, DFlash 2 drafter, 123 ms 32K context reopen) in spacepilot/registry/runtimes/splash.yaml
• Add Qwen3.8 27B Splash variant in spacepilot/registry/models/qwen3-8.yaml with Inco AI cited benchmarks (74 tok/s short, 54 tok/s at 32K)
• Merge PR feat(registry): register muse-glimmer-coreai recipe with K=4 DFlash drafter #131 (feat/recipe-muse-glimmer-coreai) for Meta Muse Glimmer 30B CoreAI recipe with K=4 DFlash speculative drafter and ATEM XML streaming
• Merge PR feat(registry): add MiniCPM5-2B with flown M1 Max measurements for BF16 and MLX 4-bit #130 (feat/registry-minicpm5-2b) for MiniCPM5-2B
• Fix unpinned variants in auk.yaml and nemotron-3.yaml (resolving test_every_shipped_variant_is_pinned)
• Regenerate web/registry.json with all 40 models and 69 pinned variants (zero unpinned warnings)

Receipts

• tests/test_model_revisions.py: 20/20 PASS
• tests/test_muse_glimmer_recipe.py: 4/4 PASS
• tests/test_runtimes.py: 14/14 PASS
• tests/test_registry.py & tests/test_export_registry.py: 21/21 PASS

---

^{Need help on this PR? Tag @codesmith with what you need. Autofix is disabled.}