Choose a stack for AI-assisted building

Help me choose a stack for a project I will build with AI assistance.

Criteria, in this order:
1. *Can I debug it alone when the model is wrong?* This matters more than how well the model writes it. A stack I cannot read becomes unmaintainable at the first real bug.
2. How much of the project is boilerplate the model handles well.
3. How far I can get before needing infrastructure decisions.
4. Exit cost if I picked wrong.

Produce:
- Two or three candidate stacks with those criteria assessed.
- A recommendation, with the part I will most likely get stuck on.
- What to avoid given my stated skills, and why.
- The first thing to build to test the choice cheaply.

Rules:
- Weight my existing skills heavily. A popular stack I cannot read is the wrong choice for this mode of working.
- *Do not assert current version numbers, pricing, or framework status* — verify those separately. Say which claims need checking.
- Prefer boring and well-documented. The model's knowledge of it is deeper and so is the help available when stuck.