This position has been well discussed with respect to SF on various engine forums, but mostly on talkchess.com
As indicated by Linnemann, the issue is that the variation with Rxf5 and Qxh6 includes some quiet moves, and SF reduces that line heavily.
In general, this sort of problem is unavoidable (although some specific cases might get remedied fairly cheaply).
Tossing away lines that will likely be irrelevant is a crucial part of modern engines' strength.
For the vast majority of positions, if you're down massive amounts of material and have to make some quite moves, then you're just losing.
For some unusual positions, like this one, that happens not to be the case.
As a developer, you're then faced with a choice: do I change the general behavior of my engine so that it pays more attention to these almost certainly losing lines (which will weaken its general game play), or do I continue to reduce these lines heavily, which leads to stronger play in general but worse performance in these unusual positions.
Most developers, but especially so the SF team, only really care about playing strength.
They're not going to push through a patch that helps in a handful of unusual positions at the expense of general playing strength.
That's why there are forks of SF, like SF Matefinder, that are designed to cope with this sort of position in analysis (but are weaker in general play than the main SF; incidentally, it does find it much faster than the main SF, getting the right move at iteration 18, and the right eval at iteration 19).
To further illustrate, it's actually not all that rare for really old engines that are much, much weaker than SF to find solutions to some weird problems much faster than SF does.
It's just because they don't reduce likely irrelevant lines as much (or at all), so once the tactic is in their horizon, they see it.
For SF, that line looks hopeless, so it gets heavily, heavily reduced (so that that line is probably not even searched to depth 10 until the general search reaches iteration 25+).
Of course, that same behavior is a huge part of the reason SF is much, much stronger than those engines in general play.
In general, it's been found that when engines are modified to do especially well on tactical suites or unusual positions like studies, that version is weaker than the original, for all these reasons.
To wrap this all up, it's probably not going to get "fixed" any time soon.
"Correcting" behavior in .0001% of all positions is probably not a good idea if it leads to worse performance in the other 99.9999% of positions.
It's definitely a bit annoying to find positions like this, but there will always be such positions.
The best we can hope for is to find clever tricks that eliminate as many of them as possible without undue overhead on all the "normal" positions.
As indicated by Linnemann, the issue is that the variation with Rxf5 and Qxh6 includes some quiet moves, and SF reduces that line heavily.
In general, this sort of problem is unavoidable (although some specific cases might get remedied fairly cheaply).
Tossing away lines that will likely be irrelevant is a crucial part of modern engines' strength.
For the vast majority of positions, if you're down massive amounts of material and have to make some quite moves, then you're just losing.
For some unusual positions, like this one, that happens not to be the case.
As a developer, you're then faced with a choice: do I change the general behavior of my engine so that it pays more attention to these almost certainly losing lines (which will weaken its general game play), or do I continue to reduce these lines heavily, which leads to stronger play in general but worse performance in these unusual positions.
Most developers, but especially so the SF team, only really care about playing strength.
They're not going to push through a patch that helps in a handful of unusual positions at the expense of general playing strength.
That's why there are forks of SF, like SF Matefinder, that are designed to cope with this sort of position in analysis (but are weaker in general play than the main SF; incidentally, it does find it much faster than the main SF, getting the right move at iteration 18, and the right eval at iteration 19).
To further illustrate, it's actually not all that rare for really old engines that are much, much weaker than SF to find solutions to some weird problems much faster than SF does.
It's just because they don't reduce likely irrelevant lines as much (or at all), so once the tactic is in their horizon, they see it.
For SF, that line looks hopeless, so it gets heavily, heavily reduced (so that that line is probably not even searched to depth 10 until the general search reaches iteration 25+).
Of course, that same behavior is a huge part of the reason SF is much, much stronger than those engines in general play.
In general, it's been found that when engines are modified to do especially well on tactical suites or unusual positions like studies, that version is weaker than the original, for all these reasons.
To wrap this all up, it's probably not going to get "fixed" any time soon.
"Correcting" behavior in .0001% of all positions is probably not a good idea if it leads to worse performance in the other 99.9999% of positions.
It's definitely a bit annoying to find positions like this, but there will always be such positions.
The best we can hope for is to find clever tricks that eliminate as many of them as possible without undue overhead on all the "normal" positions.
