Add Integartion Tests

[ajit-zer07]

Summary

• Add a separate integration_tests/ Rust crate with 55 gRPC integration tests across 3 tiers (protocol, Rig agent tools, real LLM E2E)
• Add WatchSignals streaming RPC for ambient Signal broadcast
• Add comprehensive test and runtime documentation

What this PR does

Integration test suite (integration_tests/)

A standalone Rust crate that tests the runtime through the real gRPC boundary. It is not part of the main Cargo build — cargo build --release ignores it entirely.

Tier 1 — Protocol (47 tests): Scripted gRPC tests covering all 6 modes (decision, proposal, task, handoff, quorum, multi-round), plus:

• Error paths (invalid version, empty mode, unauthorized sender, oversized payload)
• Mode-specific reject paths (non-initiator commit, premature commitment, wrong phase)
• RFC cross-cutting features (signals, version binding, dedup invariants, cancel auth, discovery, TTL)

Tier 2 — Rig Agent Tools (5 tests): Each MACP operation implemented as a Rig Tool trait, validated through ToolSet::call() — the same interface an
LLM agent uses.

Tier 3 — E2E with Real LLM (3 tests, #[ignore]): Real OpenAI GPT-4o-mini agents coordinating through the runtime:

• Decision with signals: orchestrator (code) proposes → 3 specialist LLMs evaluate in parallel → signals on ambient plane → orchestrator commits
• Decision: simplified version without signals
• Task delegation: planner (code) creates task → worker (LLM) completes → planner commits

WatchSignals RPC

New server-streaming RPC that broadcasts ambient Signal envelopes to all subscribers. Signals are non-binding messages on the ambient plane (per RFC-MACP-0001 §5.1) — they
correlate with sessions via SignalPayload.correlation_session_id but never enter session history.

Changes: proto/macp/v1/core.proto, src/runtime.rs (signal broadcast channel), src/server.rs (WatchSignals handler)

Documentation

• New docs/testing.md — comprehensive integration test guide
• Updated README.md — added Testing section
• Rewritten CLAUDE.md — added architecture diagram, mode authority table, gRPC surface table, env vars table, signal/dedup rules, integration test section
• Updated docs/protocol.md — added WatchSignals to RPC list

Test plan

• cargo fmt --all -- --check passes
• cargo clippy --all-targets -- -D warnings passes
• cargo build succeeds
• cargo test --all-targets passes (unit + Rust integration tests)
• cd integration_tests && MACP_TEST_BINARY=../target/debug/macp-runtime cargo test --test tier1 -- --test-threads=1 — 47 protocol tests pass
• cd integration_tests && MACP_TEST_BINARY=../target/debug/macp-runtime cargo test --test tier2 -- --test-threads=1 — 5 Rig tool tests pass
• cd integration_tests && OPENAI_API_KEY=... MACP_TEST_BINARY=../target/debug/macp-runtime cargo test --test tier3 -- --ignored --test-threads=1 — 3 E2E tests pass
  with real OpenAI
• cargo build --release produces only macp-runtime binary (no test artifacts)