fix(safety): add inbound secret scanning to engine v2 path

[serrrfirat]

Summary

• Engine v2 path (handle_with_engine_inner in bridge/router.rs) forwarded user messages to the conversation manager without any safety checks, allowing pasted secrets to reach the LLM and be permanently stored
• Adds the same three inbound safety checks that v1 (thread_ops.rs) already enforces: validate_input, check_policy, and scan_inbound_for_secrets
• Includes regression test with Slack bot token and OpenAI key patterns

Closes #2491

Test plan

• cargo test --lib -- bridge::router::tests::handle_with_engine_blocks_inbound_secrets — new regression test
• cargo test --lib -- bridge::router::tests — all 24 router tests pass
• cargo clippy --all --benches --tests --examples --all-features — zero warnings
• Manual: set ENGINE_V2=true, paste a xoxb-... token in chat, verify it's rejected with a warning

🤖 Generated with Claude Code

[gemini-code-assist]

Code Review

This pull request introduces safety validation, policy enforcement, and secret scanning for inbound messages in the engine v2 pipeline, mirroring protections from the v1 path. It also includes a regression test for blocking leaked secrets. Feedback identified a bug in the test case where the mock OpenAI key was too short to trigger the leak detector, and a code suggestion was provided to correct the key length and strengthen the test assertions.

[gemini-code-assist]

The OpenAI key used in this test, sk-abc123def456ghi789, has a 19-character payload, but the regex for OpenAI keys in leak_detector.rs requires at least 20 characters. As a result, this test case will not be caught by the secret scanner, and the test will not behave as expected.

I've updated the key to be 20 characters long and also made the assertion more specific, similar to the check for the Slack token.

Suggested change

	let result = handle_with_engine_inner(&agent, &sk_msg, &sk_msg.content, 0)
	.await
	.expect("should not error");
	assert!(result.is_some(), "OpenAI key should be blocked");
	// Clean message should pass through (will fail at conversation
	// manager level since test state has no real engine, but it must
	// NOT be rejected by the safety checks).
	let sk_msg = IncomingMessage::new("web", "alice", "my key is sk-abc123def456ghi789j");
	let result = handle_with_engine_inner(&agent, &sk_msg, &sk_msg.content, 0)
	.await
	.expect("should not error");
	let warning = result.expect("should return a warning, not None");
	assert!(
	warning.contains("secret") || warning.contains("credential"),
	"expected secret-detection warning for OpenAI key, got: {warning}"
	);

References

1. In tests, prefer using .expect() to provide clear error messages and fail explicitly rather than using silent fallbacks like .unwrap_or() that could mask logic errors.

[serrrfirat]

Addressed gemini-code-assist review: extended the mock OpenAI key from 19 to 21 chars (regex requires 20+) and added a specific assertion. Fixed in b57c380.

[henrypark133]

Review: Inbound secret scanning on engine v2

No verified findings in the current diff. The safety validation, policy check, and inbound secret scan are now applied on the engine-v2 ingress path before the message is handed to the conversation manager, which restores parity with the v1 pipeline.