This is a very good question! There are two answers.
First off, you can absolutely do what you are describing. There are systems like Metamath that assume nothing beyond first-order logic and the axioms of set theory, and apply these verbose rewrites. See https://en.wikipedia.org/wiki/Metamath for more. However, Lean (and many past theorem provers, too) have 1) preferred to represent mathematical objects as types rather than sets, and 2) built a large amount of interactive tooling rather than the blind rewriting of Metamath. The first choice is essentially of elegance -- the Curry-Howard correspondence indeed lets you write propositions as types and proofs as terms inhabiting those types, which is a connection of deep beauty. Ontologically, there is also an argument to be made that representing a mathematical object as a type is more "correct". We don't really think of ex. two as "{{},{{}}}" or "{{{}}}". We think of it as "Two", or maybe "Zero +1 +1". This is a win for types. Mathematical objects are not grocery bags.
Second, the historical approach to theorem proving has been interactive theorem proving. Theorem provers were supposed to be a second brain for students and computer scientists and mathematicians... keeping track of assumptions, suggesting possible tactics to take... aiding the human in the process of proof (and subsequently understanding), rather than replacing them. The path of uncritical adoption of AI that Lean has taken, sadly, much replaces the human. There is no real pragmatic reason for Lean to retain its interactive (human) elements any more (besides inertia) as AI scales up. It might as well be Metamath, indeed.
First off, you can absolutely do what you are describing. There are systems like Metamath that assume nothing beyond first-order logic and the axioms of set theory, and apply these verbose rewrites. See https://en.wikipedia.org/wiki/Metamath for more. However, Lean (and many past theorem provers, too) have 1) preferred to represent mathematical objects as types rather than sets, and 2) built a large amount of interactive tooling rather than the blind rewriting of Metamath. The first choice is essentially of elegance -- the Curry-Howard correspondence indeed lets you write propositions as types and proofs as terms inhabiting those types, which is a connection of deep beauty. Ontologically, there is also an argument to be made that representing a mathematical object as a type is more "correct". We don't really think of ex. two as "{{},{{}}}" or "{{{}}}". We think of it as "Two", or maybe "Zero +1 +1". This is a win for types. Mathematical objects are not grocery bags.
Second, the historical approach to theorem proving has been interactive theorem proving. Theorem provers were supposed to be a second brain for students and computer scientists and mathematicians... keeping track of assumptions, suggesting possible tactics to take... aiding the human in the process of proof (and subsequently understanding), rather than replacing them. The path of uncritical adoption of AI that Lean has taken, sadly, much replaces the human. There is no real pragmatic reason for Lean to retain its interactive (human) elements any more (besides inertia) as AI scales up. It might as well be Metamath, indeed.