fix: avoid shell execution when publishing OpenClaw skills

[n6p975v4v4-cyber]

Summary

• invoke clawhub with an argument vector instead of a shell command
• preserve normal publishing and --dry-run passthrough behavior
• add a regression test for repository-derived skill names containing shell syntax

Security impact

The publishing helper previously constructed a shell command from generated skill directory names. Shell interpretation could execute content embedded in an untrusted name when a maintainer ran the release helper. Direct argument-vector execution keeps those names as literal clawhub arguments.

Validation

• focused security regression
• OpenClaw publishing subsystem tests
• static syntax and patch-integrity checks
• rule-copy and version-consistency checks
• CI-equivalent root, Pi-extension, and MCP test suites
• post-validation mock trial for normal and --dry-run publishing