I have, in the past, run a model to investigate this with the best information I had on hand. It was a simulation study and as such the results weren't 100% but I had a rough idea of what was going on.
Going on simple models and inputs based on what I could find, I found that usually you'd need to pull w/ about 7-10 minutes to match the win expectancy that you would get by not pulling (that is to say, the "pull the goalie" strategy) is optimal when down by a single goal.
I could build the code again... but I really don't want to. The problem with the sim study was that you never get exact result. That's actually a harder inference problem than setting up the original simulation
Do note, that you're changing what is likely an equitable rate in the 2.2-3.0 scoring rate for both sides (assumed equal) to something like 3 to 4 vs. 15 to 30. I forget what NHL appropriate numbers are and I changed up the inputs to see what happened.
---
That being said, models are models, approximations of reality, no matter how sophisticated.
If teams opted for longer lengths without the goalie the teams would get better at this situation. Which way that pulls? I don't know, but I'd favor the winning team if I were a betting man. I'd also likely agree with the shift issue and exhaustion... on the other hand, with current icing rules and that you have six men on the ice you are more likely to trap the opposing team. Even then, how sure are we that the lower lines vs lower lines would be better or worse?
No sure thing here, but I think I'd hedge earlier if I were in control of the coach robot.