Add a reproducible Strands Evals state-review companion

[noteflowai]

Developers using Strands Evals need a concrete way to connect an improved aggregate to the state rules that changed. Add an optional companion recipe using the native 1.3.0 SDK and four existing saved Docker cases: its equally weighted two-rule mean improves from 0.75 to 0.875, while the notes rule regresses.

• Preserve native reports and align verdicts by case/evaluator identity; reject missing, duplicate or unmatched inventories. Keep the different original EvalArc five-dimension score explicitly separate.
• Add English/Chinese instructions, a tested dependency lock and a focused community discussion draft. The recipe accepts only the fixed published fixtures; it is not an arbitrary report importer.
• Add an independent CI job using the real SDK with network attempts rejected. It checks serialization, forward/reverse regressions, report inventories, committed report equality and output preservation, and gates both deployments.

Local validation: native SDK check passed with zero network attempts; lint, formatting, local documentation links and official HF card validation passed. No model, AgentCore or new agent execution was performed; the Python release remains 0.12.1.