Hackathon/langchain hitl approval

[dezsow]

Mirrors LC-112 into the engine rule fixture and adds focused fire/silent regression tests. Coverage includes missing approval prerequisites, middleware without a checkpointer, an empty middleware list, the structurally safe case, and agents without privileged execution tools.
This does not change LangChain discovery, predicate semantics, or evaluator behavior; it exercises existing machinery against the new rule.

[jhumel-code]

Thanks for splitting this cleanly across the two repos, @dezsow. The fixture addition matches the YAML in trustabl/agent-reliability-rules#110 byte for byte, so rules-sync stays happy, and you kept to existing machinery: agent_class, agent_uses_hosted_tool_class, agent_kwarg_list_empty and agent_kwarg_missing all exist on main at schema 14, so no version bump is warranted. The case selection is good too, particularly the empty-middleware-with-checkpointer-present case, which is the one that actually exercises the any branch rather than passing by accident.

One blocking item. TestPolicyRules_AllRulesCovered builds its coverage set only from policyRuleCases, policyAgentRuleCases, policyRepoRuleCases, policySubagentRuleCases and policySkillRuleCases in internal/rules/policies_test.go. A table in a standalone lc112_policy_test.go is invisible to it, so once LC-112 is in the fixture the guard will report it as having no coverage at all and fail. Please add a fire and a silent policyAgentCase for LC-112 to policyAgentRuleCases (keep the dedicated file as extra depth if you like), with Language set on every AgentDef so TestFixtureAgentsHaveLanguage stays green.

I also left a severity question on the rules PR that would change this YAML, so the two want to land together. LC-112 will need a rationale doc in trustabl-rulebook before it ships; that part is ours and we are happy to pair on it.

Happy to take this once the coverage entries are in policies_test.go and the rules-side question is settled.