A distinction there is that they can only become invalid via a
conflict— replaced by another transaction authored by the prior
signers. If no other transaction could be created (e.g. you're a
multisigner and won't sign it again) then there is no such risk.

You need to check transaction's dependencies up to a certain depth to know whether it is safe:

If one of inputs depends on transaction which is signed by parties with unknown trustworthiness, then it isn't safe.

It now introduces chance events ("act of god") into the mix where they
they didn't exist before.

You need to check transaction's dependencies up to a certain depth to know whether it is safe:

If one of inputs depends on transaction time-locked script (or other unrecognized script), then it isn't safe.

Situation is identical, you might need several extra lines of code.

I think it would matter only if we had deterministic, reliable mempool and reorganization behavior. But it's not something we can depend on.