dsh-agent-loop-experiment: Adding a convergence strategy layer to the official agent loop

This is an experimental package that maintains byte-level consistency with the official @deepseek-ai/dsh-agent-loop implementation, stacking convergence strategies on top of public prompts and agent hooks: runtime fingerprinting and no-progress interception, one-time short-reference recovery for large write failures, tiered file change protection, and soft tool requirements. The article provides four export entry points, a convergence configuration table with defaults, commands for installing and rolling back bundle patches, workflows for standalone host integration and offline and online comparative verification, and explains the trigger boundaries of the strategies. It is suitable for those wishing to conduct comparative testing of long-running task convergence behaviors on the DSH Web profile or a standalone host. The plugin runs with the permissions of the current DSH process; users must read the source code before installation to confirm the patch scope and license. It is locked to support DSH 0.2.2-rc.1.