[sauravsingla]

Completes the post-0.7 proof/release phase without changing frozen scored research artifacts.

Key changes:

• plug-and-play AgentWeaveApplication.from_mcp() and .from_mcps() factories
• multi-MCP composition with collision-safe model aliases and canonical-key executor dispatch
• real official MCP SDK end-to-end tests, including two MCP servers exposing the same native tool name
• runtime red-team coverage for malformed JSON/schema, argument escalation, and hidden high-risk tool selection
• deterministic routing scale study through 100,000 tools with wall-clock latency and Python peak-memory evidence
• separate real-provider Issue Benchmark LLM-dynamic tool selection vs pre-inference routing and hybrid delegation #38 harness preserving all four strategies plus hybrid ablations and reporting provider tokens/wall-clock latency
• credentialed real-provider workflow that refuses to fabricate a live result when credentials are absent
• built-wheel installation smoke matrix for base/MCP/LangGraph/AutoGen extras
• stronger release workflow: core + upstream integration + real MCP E2E + wheel smoke before immutable GitHub release/PyPI publishing
• 0.7 quickstart, MCP composition docs, live-provider protocol, and road-to-1.0 evidence/API gates

The frozen controlled Issue #38 JSON and prior scored benchmark artifacts are intentionally untouched.

Merge only when the exact PR head is green across core CI, integration compatibility, runtime security/scale proof, package installation smoke, and existing repository proof workflows.