All of them are wrong.
What we learn from are information-rich events, as Donald Reinertsen describes in his fantastic book Principles of Product Development Flow.
“Either excessive or insufficient probability of failure reduces the efficiency with which we generate information. […] Avoid oversimplifications, like 'eliminate failures' or 'celebrate failures.' There is an optimum failure rate.”
People learn most when they cannot predict the outcome of their behaviors. When the chance of failure or success is about 50%, Reinertsen says. (And there’s no way around it. This insight is based on solid information theory.)
Failure and success are orthogonal to learning. What you learn from are the experiments and tests that you run. When you make mistakes, or simply follow good practices, you will probably learn very little. No matter whether the result is a failure or a success.
It makes no sense to celebrate a success you have achieved despite having made stupid mistakes. It also makes no sense to celebrate failure you ended up in despite having followed good practices. In either case celebrations would just reinforce the wrong things.
You should focus neither on failures nor on successes. What you should focus on are experiments that maximize learning.
Celebrate learning, not success or failure.