It fits execution well: the problem genuinely is sequential, the reward, implementation shortfall against a benchmark, is well defined, and a simulator can be built from real order book data with reasonably honest dynamics.
It fits alpha discovery badly. RL needs enormous numbers of episodes, and market history gives you very few genuinely independent ones. Training on a simulator that does not react to your orders teaches the agent to exploit the simulator, which is a sophisticated form of overfitting.
If you experiment with it, hold out entire market regimes rather than random slices, cap position sizes inside the environment, and compare against a simple benchmark policy. Most published improvements vanish against a well-tuned twap-algorithm.
Related: execution-algorithm, overfitting, markov-property, arrival-price-benchmark