test(langchain): add LC-103 fixture coverage

[jiaxinaspenlin-dotcom]

Summary

Adds the engine-side fixture and test coverage for LC-103, which detects LangChain agents that wire langchain_community's Requests* built-in HTTP tools.

What changed

• Mirrored LC-103 into:
  • testdata/rules-fixture/langchain/agent_safety.yaml
• Added three agent-rule test cases in:
  • internal/rules/policies_test.go

Coverage includes:

• FIRE: ReactAgent with RequestsGetTool
• FIRE: AgentExecutor with RequestsPostTool
• SILENT: ReactAgent with TavilySearchResults

Every AgentDef sets Language.

No analyzer changes were required because the existing LangChain hosted-tool discovery already recognizes all five Requests* classes.

Validation

• go vet ./... — pass
• go test -race ./internal/rules/... — pass
• go test -race ./... — 26 packages passed, 0 failures
• go build -o /tmp/trustabl ./cmd/trustabl — pass
• check-rules-sync.sh — in sync with production

The fixture is byte-identical to the corresponding rule definition in trustabl-rules.

Paired changes

• feat(langchain): add LC-103, agent wires a raw HTTP Requests built-in tool trustabl-rules#103 — LC-103 rule definition
• docs(langchain): add LC-103 rationale for Requests built-in tools trustabl-rulebook#79 — LC-103 rationale documentation

All coordinated changes use the shared branch:

feat/langchain-requests-builtin-tool

[jhumel-code]

Thanks @jiaxinaspenlin-dotcom, this is a clean engine half. I compared the testdata/rules-fixture/langchain/agent_safety.yaml hunk against the one in trustabl/trustabl-rules#103 and it matches line for line, so rules-sync stays green provided the two land together. The three cases follow the LC-101 pattern directly above them, set Language, and the two fire cases exercise different applies_to kinds (ReactAgent resolves to langchain_agent, AgentExecutor to langchain_agent_executor), which is more than the minimum the coverage guard asks for.

Two nits, neither blocking. applies_to also lists langchain_state_graph and nothing covers it, so mirroring LC-101's StateGraph case would close that out. The silent case uses TavilySearchResults, which is not in LangChainHostedToolClasses and so is never emitted as a HostedToolRef by a real scan; it does exercise the class filter, but an agent with no hosted refs would be the shape users actually hit.

Coordination note: LC-103 is also claimed by #183 and #186, so whichever lands first, the others will need to renumber and rebase. That is not a mark against this one.