v0.3.4

What's Changed

• feat: pin Splash 1.2.1 and run its recipe under the exact policy by @kartikey-vyas in #34
• feat: add RTX Spark recipes and drop its NPU from detected accelerators by @kartikey-vyas in #35
• feat: send first-party submissions with AGENTPERF_SUBMIT_KEY by @kartikey-vyas in #36
• feat: count Windows shared GPU memory on the RTX Spark by @kartikey-vyas in #40
• feat: convert a safetensors-only llama.cpp draft at deployment and require a minimum llama.cpp build by @kartikey-vyas in #41
• feat: add LFM2.5 24B A2B llama.cpp recipes for each owned device by @kartikey-vyas in #37
• feat: follow NVIDIA's v0.61 DGX Spark recipes and fix the Nemotron vLLM pin by @kartikey-vyas in #38
• feat: add a llama.cpp Qwen3.8 27B Q4 recipe for the DGX Spark by @kartikey-vyas in #39
• chore: bump version to 0.3.4 by @james-shore in #42

Full Changelog: v0.3.3...v0.3.4