You've perfectly described my biggest nightmare, and also the reason I'm here! I'm well aware the codebase is likely a "footgun" factory. That's why the beta is completely sandboxed (no real funds), and I'm hoping to crowdsource the "gun safety" inspection from experienced devs like you. Consider this a public bug bounty invitation.
There's a famous quote from Harold Dodge: "You can not inspect quality into a product." It has to be designed in from the start.
That's a fantastic quote, and in a traditional development world, I would agree 100%. Quality absolutely has to be designed in.
However, what I've discovered through this process is a new paradigm that feels different. When you're working in a tight feedback loop with an AI, the "design -> build" cycle shrinks from weeks to minutes. You can "design in" a core principle, ask the AI to implement it, and then immediately ask it to "inspect" its own code for flaws against that principle.
It's less like building a bridge and more like sculpting. You start with a block of marble (the AI's vast knowledge), and you iteratively chip away, constantly inspecting and refining.
I agree that you can't "add" quality at the end. But I'm finding that with AI, the "beginning" is never more than 5 minutes away. It's a continuous process of "design-inspect-refine". I'm still exploring the limits of this approach, and your point is a crucial part of that exploration.