feat(workbench): support durable NoKV restore

[wchwawa]

Summary

• preserve MCP structuredContent across stdio and HTTP transports
• promote structured JSON tool errors to stable top-level code, message, retryable, and details fields while preserving legacy plain-text behavior
• update the NoKV workbench skill to v0.4 with durable restore retry, terminal-state, and manifest validation rules
• strengthen deployment preflight to validate the raw restore schema and capability-gated tool advertisement

Design boundary

LingTai does not keep a second restore state machine or orchestrate a multi-step copy. After MCP reconnect it reuses the saved numeric snapshot ID and resends the same strict request. No Agent startup path is hard-coded to NoKV, and the change does not auto-revive a destination workbench or couple restore to CPR internals.

The transport decoder is protocol-generic; NoKV-specific policy remains in the intrinsic workbench skill and deployment preflight.

Validation

• 179 focused MCP, capability, skill, reconnect, config, discovery, LICC, and skill-manual tests passed
• Ruff checks passed for all changed Python files
• companion NoKV full --require-all live E2E passed with real RustFS, NoKV serve, Agent registration/reconnect, and two independent MCP clients

Companion NoKV PR: NoKV-Lab/NoKV#407
Prerequisite Holt PR: NoKV-Lab/holt#29