fix(skills): treat slash-command load failures explicitly

[yeyitech]

Summary

• return an explicit SkillInvocationResult from build_skill_invocation_message() instead of a truthy placeholder string
• update CLI, gateway, and webhook callers to handle ok, unknown_command, and load_failed outcomes explicitly
• add regressions for post-scan skill load failures so broken skills no longer inject bogus prompt text into the agent

Closes #11200.

Testing

• PYTHONPATH=/tmp/hermes-pytest-asyncio python -m pytest -o addopts= tests/agent/test_skill_commands.py tests/cli/test_quick_commands.py tests/cli/test_cli_plan_command.py tests/gateway/test_unknown_command.py tests/gateway/test_plan_command.py tests/gateway/test_webhook_integration.py -q

[alt-glitch]

Related to #11408 and #14752 — all three fix the same root cause: build_skill_invocation_message() returns truthy error string on load failure, treated as success. This PR uses an explicit SkillInvocationResult dataclass approach.

[teknium1]

Thanks for the clear reproduction and the explicit-result design.

Automated hermes-sweeper review found that the behavioral bug is already fixed on current main:

• agent/skill_commands.py:509-511 returns None when the post-scan skill payload cannot load, preventing the former truthy failure placeholder from reaching CLI/gateway callers.
• tests/agent/test_skill_commands.py:532-538 covers that failed-load path.
• Commit 4b17c2411ab2518fabe9d87872a03a47d3b8cfcc made this exact correction and shipped in v2026.5.28.

The typed SkillInvocationResult alternative is therefore superseded by the implementation already on main.