Step 46: Warm-up — JIT pre-compilation, CUDA kernel pre-loading, configurable WarmupOptions #92

Summary

JIT pre-compilation pass at startup. CUDA kernel pre-loading. Configurable WarmupOptions. Readiness probe gates on warm-up completion.

At server startup, run configurable dummy inference passes after model loading but before marking IsReady=true. This triggers .NET Tier-0→Tier-1 JIT promotion (with Dynamic PGO profile data) and exercises CUDA/cuBLAS kernel pipelines, eliminating the cold-start latency penalty on the first real request.

Acceptance Criteria

• WarmupOptions sealed record in DotLLM.Engine with Enabled, DummyPrompt, MaxTokens, Iterations properties
• WarmupRunner static class that executes warm-up passes using TextGenerator
• Server integration: warm-up runs in ServerStartup.LoadModel() before IsReady=true
• Hot-swap: warm-up runs automatically on model swap via LoadModel()
• CLI flags: --no-warmup (disable), --warmup-iterations N on serve command
• Per-iteration timing logged with [dotllm] prefix
• /ready probe returns 503 during warm-up, 200 after completion
• docs/WARMUP.md documentation covering JIT, CUDA, Dynamic PGO, configuration
• CLAUDE.md Documentation Index updated with WARMUP.md
• docs/ROADMAP.md step marked complete
• Unit tests for WarmupOptions
• Integration test for WarmupRunner with SmolLM-135M

References

• Roadmap: Step 46 (Phase 6 — Improved Serving)
• Depends on: Step 34 (ASP.NET server) ✅
• Design spec: docs/SERVER.md (Warm-up section)