v0.2.0 — Reliability: benchmark framework, packs ecosystem, developer experience

v0.2.0 (Reliability)

Benchmark framework (v0.2)

• Unified runner: python benchmarks/runner.py --domain ecommerce
• Six question types: semantic understanding, entity discovery, relationship reasoning, metric dependency, evidence grounding, safety validation
• SRB score with hard safety gate (multiplicative factor)
• Per-domain datasets: model.yaml / questions.json / expected_context.json / expected_metrics.json / safety_scenarios.json / evaluation.json

Ecosystem & experience

• Five built-in semantic packs: ecommerce, saas, finance, game, healthcare
• Streamable HTTP transport for the MCP server (--http)
• docker compose quick start, Claude Desktop / Cursor config example
• SafetyProvider extension point for external safety engines (JoinLint adapters)
• Killer demo with root-cause evidence (examples/ecommerce/demo.py --pack <name>)

Verification

• 125 tests passing (local + CI), ruff clean
• Benchmark: metric dependency 1.000, safety validation 1.000, SRB 0.581