feat: recipe system — hardware-aware model recommendations

[raullenchai]

Summary

• Add rapid-mlx recipe <model> CLI command: auto-detects Apple Silicon hardware, computes max context length, estimated tok/s, and generates the optimal rapid-mlx serve command
• Enhance rapid-mlx models with hardware-aware table showing fits/OOM/tight status per model
• 40 Apple Silicon hardware profiles (M1→M4, all memory tiers) with auto-detection via sysctl
• 16 model recipes from existing benchmark data (Qwen3.5, Qwen3-Coder, GLM, Phi-4, Mistral, Gemma, etc.)
• Bandwidth-scaling estimation: estimated_tps = measured_tps × (target_bandwidth / reference_bandwidth)
• JSON output mode for agent consumption (--json)
• --run flag to start server directly with recommended config

Usage

rapid-mlx models                          # auto-detect hardware, show all models
rapid-mlx models --hardware m4-max-64     # specific hardware
rapid-mlx recipe qwen3.5-35b             # detailed recipe
rapid-mlx recipe qwen3.5-35b --json      # for agents
rapid-mlx recipe qwen3.5-35b --run       # start server directly

Test plan

• 37 new tests (hardware profiles, model recipes, engine computation, CLI integration)
• All 37 pass
• 1279 existing tests pass (1 pre-existing failure in test_reasoning_parsers unrelated)
• Lint clean (ruff check + format)
• Tested models/recipe commands with auto-detect, explicit hardware, OOM cases, JSON output

🤖 Generated with Claude Code