Because if you throw away old mechanics you lose part of the game. It's better to use the seed of the old terrible mechanic to build the new workable mechanic.
Improving a mechanic rather than scrapping it and trying something completely different is sensible enough, if it can be done. But, refinement, alone, can't accomplish what it could in combination with innovation. I guess it comes down to a risk/reward evaluation. If you aren't open to change, you don't have much chance for improvement, but you avoid inadvertently screwing things up worse than they already are. Of course, even if you allow nothing you control to change, things can still change around you.
Much like how they used "hit dice" from older editions in 5e, mixing the good idea of healing surges with the nostalgia.
Not the best example, I think. On the one hand, the use of the Hit Dice terminology for a completely different mechanic is such a transparent and cosmetic allusion to the classic game that it seems almost condescending. OTOH, it's not an old mechanic being improved, it's the accomplishments of a more recent mechanic being rolled back (healing surges solved the old 'heal bot' problem, while Hit Dice do not).
A better example might be level. In classic D&D, characters had a level, dungeons had levels, and spells had levels, and those levels didn't exactly correspond 1:1, even characters with the same experience points earned in the same adventures could be of different levels, monsters didn't have levels but some could drain levels or perhaps cast as a if they had a certain class/level. In 3e, characters with the same experience points /were/ the same character level, but not necessarily the same class or caster levels, and monsters had a CR that corresponded to level, but spells were still on a different level scale, and it also added level adjustments. In 4e, characters, monsters, spells, and magic items all had levels that corresponded neatly. There you have the old mechanic (or term) retained, but, incrementally, over time, made more consistent and streamlined. Another good example would be resolution: going from a mix of d6 (surprise, initiative), % (some skills, con checks, random tables, other oddball mechanics), and d20 (roll high or roll low), to d20 (roll high) for surprise, initiative, attacks, saves, skills & checks. Retaining % system shock or d6 surprise rolls wouldn't have made the game any better, and dropping them didn't make it any less D&D.
You could have a much tighter game with three classes (warrior, skill user, spell-caster) combined in different ways for the other classes. Resolution uses 2d6 adding a bell curve for average results with a 7 meaning success, <6 failure, 2 a fumble, and 12 a critical success. But that wouldn't be D&D.
It wouldn't be familiar as D&D to us, since we're accustomed to associating D&D with many classes, a uniform randomization mechanic, and other oddities. It would be part of what defined D&D for someone who began playing D&D that way, though.
One change isn't much and is fine. But it's never just one change. If it's acceptable to drop elements bits and pieces of the game just fall away and are lost. After all, it's easier to make something new than fix older elements. Slowly, subtly, the spirit and feel of the game is lost.
So it's avoiding the slippery slope? That's prettymuch the classic traditionalist (change-adverse) argument.
You might say the feel of the game is 'lost,' because it's no longer exactly the same, but it could well be evolving into something better.
Feel ties directly to mechanics. There's a tonne of D&D IP in the Delve board games (Ravenloft, Ashardalon, Drizzt) but they don't feel like playing true D&D.
They're board games.
