Add AI/ML extensions and MCP server for LLM-driven JVM diagnostics

[jbachorik]

Summary

Adds five new modules that bring BTrace into the AI/ML observability space:

• btrace-mcp-server — Model Context Protocol server exposing BTrace as 7 MCP tools (list_jvms, deploy_oneliner, deploy_script, list_probes, send_event, detach_probe, exit_probe) plus 3 diagnostic prompt templates. Enables LLMs like Claude to attach to JVMs and run BTrace probes via natural language.

• btrace-llm-trace — LLM inference observability: token counts (input/output/cache-read/cache-creation), latency (min/avg/max), cost estimation with cache savings, streaming TTFT. Covers Langchain4j, Spring AI, OpenAI Java SDK, Anthropic Java SDK patterns.

• btrace-rag-quality — RAG pipeline observability: vector DB query latency, similarity score tracking (top/low/spread), empty retrieval rates, chunk token budgets, end-to-end pipeline timing (retrieval vs generation). Targets Pinecone, Milvus, Weaviate, Chroma, pgvector, Qdrant.

• btrace-vibe-guard — Runtime behavioral contracts for AI-generated code: latency budgets, call rate limits (sliding-window), range checks, null-safety, boolean assertions, and AI-vs-human code path performance comparison.

• btrace-gpu-bridge — GPU compute and model inference tracing: ONNX Runtime, DJL, TensorFlow Java session tracking, batch throughput, device memory allocation/peak tracking, native FFM call profiling (cuBLAS, cuDNN).

Design principles

All four extensions follow the same architecture:

• Zero allocation on hot paths — fluent builders (llm.call("model").inputTokens(1500).record()) use ThreadLocal pooling, reusing one instance per thread instead of allocating per call
• Lock-free concurrency — all statistics use AtomicLong with CAS loops for min/max, no locks anywhere
• Zero external dependencies — instrument existing client library classes, no additional JARs required
• Standard BTrace extension pattern — @ServiceDescriptor interfaces in src/api/, isolated implementations in src/impl/, auto-discovered via the Gradle extension plugin

Files

Area	Files	Lines
MCP server	12 source files	~2,000
LLM trace extension	5 source + 1 test	~1,100
RAG quality extension	5 source + 1 test	~800
Vibe guard extension	3 source + 1 test	~700
GPU bridge extension	5 source + 1 test	~800
Sample BTrace scripts	4 scripts	~335
Total	41 files	~6,200

Test plan

• Unit tests pass for btrace-llm-trace (20+ tests: duration-only, fluent builder, cache tokens, embeddings, concurrency)
• Unit tests pass for btrace-rag-quality (18 tests: queries, similarity scores, pipelines, empty retrievals, concurrency)
• Unit tests pass for btrace-vibe-guard (18 tests: latency budgets, rate limits, assertions, range checks, AI vs human paths, concurrency)
• Unit tests pass for btrace-gpu-bridge (18 tests: inference, memory alloc/free, native calls, model load, formatters, concurrency)
• MCP server compiles and handles JSON-RPC initialize/tools/list/tools/call lifecycle
• Sample BTrace scripts compile against extension APIs
• Existing BTrace tests unaffected (no changes to core beyond 3 visibility modifiers in Client.java)

https://claude.ai/code/session_012KcpiFxvscxzWWgN5LiRcB

---

This change is