feat(langchain): add LC-103 human-review approval gate

[Mantitah]

Fork of Trustabl’s rule repository, extended with LC-103, a LangChain v1 detection rule for dangerous create_agent configurations that lack visible human-review prerequisites.

[jhumel-code]

Thanks @Mantitah. The rule is schema clean: every required field is set, there are no extra keys, confidence is in range, language is explicit, and langchain_agent is a valid applies_to for agent scope. All four predicates already exist on main at schema 14, so no version bump is needed, and I like that the explanation is honest that middleware presence does not prove HumanInTheLoopMiddleware is actually wired.

My concern is overlap. LC-101 in langchain/agent_safety.yaml already fires at langchain_agent on exactly these three hosted classes with no further condition, so every LC-103 hit is also an LC-101 hit and a create_agent wiring PythonREPLTool would emit two high findings for one call site. On top of that, a checkpointer passed in when a parent graph is compiled is invisible to us, so the missing-checkpointer arm will fire on correct code. Please drop LC-103 to medium (LC-101 stays the capability call, LC-103 the gate refinement), or narrow the match so it is not a strict subset of LC-101.

Two notes that are not on you. LC-103 is also claimed by #103 and #108 for an unrelated Requests built-in rule, so whoever lands second renumbers. And a shipped rule needs a rationale doc in trustabl-rulebook; we can write that or pair on it.

The engine half at trustabl/trustabl#186 mirrors this byte for byte, which is right. Happy to take this once severity is settled.