...one of the MTOW you promote TDI is, if not wrong, then sub-optimal

One of the points I make in PBP is that "sub-optimal" is not a easy thing to judge as soon as you have more than one metric to optimize. PBP suggests at least four: maintainability, robustness, performance, and conciseness (and argues that that is the approximate order of importance of the four).

The recommendation to use anonymous hashes when passing one-off argument lists trades a little less performance for a little more conciseness, some improved robustness, and (if used consistently throughout the code) better maintainability.

Of course, you're perfectly correct to factor out the hash memory reallocations into a single declaration if you're going to repeat the same call hundreds (or millions) of times, but as a default habit, just building an anonymous hashref at the point you need it is cleaner, less error prone, and more concise. And the performance cost on a single call at modern processor speeds isn't worth the hassle of coordinating between separate hash declarations and uses.

All of which is why inlined hashes were the default recommendation in PBP.

PS: You also asked what worse mistakes I made in PBP. One of them was not finding a way to adequately get across that the advice in the book is meant to be a starting point, a default zero-state, an initiator of conscious thought, rather than an end-point of discussion. Gods know, I tried to convey that (devoting most of the first chapter to the idea), but I stupidly didn't put any code in that chapter, so most readers just seem to skip straight over it. :-(