I like your “clanking replicator” term. But it simply doesn’t exist. Even if you have all the $$$ on the planet, every bit of science and technology known to mankind, and every person on the planet working for you, you couldn’t design an electro-mechanical-chemical system that replicated itself from raw materials.
Why not? I detailed in another post in this thread how very few humans are involving in smelting metals these days, and pointed out how relatively simple it is to eliminate the last few humans. This is true across the supply chain as capitalism continues its relentless march towards eliminating labor costs. So what's the problem?
Pretty much everything involved in making a robot is mass produced, which means it's largely automated. Discrete electronics? Automated. It's not like there's such a thing as artisanal capacitors (audiophile nonsense to the contrary). Steel? Automated. Aluminum? Automated. Circuit boards? Automated. Motors? Automated 2 centuries ago. The raw materials aren't an issue. The components aren't an issue.
Assembly is the last remaining issue and that can be engineered around. Tesla tried and failed to 100% automate a vehicle assembly line but a big chunk of that failure can be laid at the feet of needing to build a device that's meant to be used by humans. A clanking replicator does not have to be human-friendly. A clanking replicator does not have interior amenities, or give a shit if the body panels are slightly out of alignment. I can't think of any operation involved in the assembly of a robotic arm that can't be done by a robotic arm, as long as it was designed with robotic manipulators in mind. There has been some academic work on the subject, and it was successful. These days I bet you could challenge a YouTube maker to design a robotic arm that could build copies of itself and get a working prototype in a few months. I have a vague recollection there's already been a couple of attempts along those lines, though I don't recall if they were done under academic rubric or by a maker.
A lot of what makes assembly more difficult today is the enormous distances involved, the assumption of the availability of human labor, and the tens of millions of financially independent entities. There are videos on YouTube of mass production of copper wiring in the developing world and it's the most ghetto operation imaginable, with exposed flames and exposed high speed spinning things all over the shop. There's a ton of human labor involved in setting up each production run for each intermediary and each product and absolutely none of those jobs exist in a German factory. And the final stage is spooling a quantity of wire, tying it in three places, and dropping it into a retail box and taping the box shut, all of which is dispensed with in a clanking replicator.
In a clanking replicator, there is never a human-friendly stage, never a retail stage, and only an absolute minimum of storage stages. All of the material handling is robot-friendly, and it stays that way throughout. When the wire is finished it gets fed into a wiring harness assembly machine, not a retail box. Steel rod gets made to length or conveyed from the rolling operation that made it straight to the cutters that cut it to the required lengths. When welding is finished on an assembly, it gets lifted out of a jig that was designed to accommodate the robot manipulator that was designed to fit the jig and the assembly. Everything that can be specific and purpose-built is specific and purpose-built, and that damn well better be most things.
Successful automation is about eliminating variability in material handling. Work pieces move through the process without getting pinched or missed or dropped or crushed because they have no choice. The system is built to guide them with a certain inevitability, so the same action gets repeated hundreds of millions of times without error. That's how the US drinks 141 billion cans of soda every year. That's 141 billion individual work pieces manufactured, filled, and packaged by robots every year and the error rate is extraordinarily low.
Having said that, on the subject of error, you have a point. If something goes wrong and a work piece jams on the aluminum can manufacturing line, it stops itself and waits for human intervention and human initiated restart. In the realms of automated quality control and error correction there's a lot of work to be done, but I think it can be done. It takes more than the dumbest automation to successfully clear a jam but it doesn't require a sapient AI. There's a whooole lot of ground between the two extremes of insensate, brute force automation and the intervention of a creature with over 1000 sensors per square inch of skin and an interpolating, extrapolating, self-aware associative memory reasoning center (a brain). I think there's enough room in that gap to build a successful clanking replicator that could operate indefinitely, correcting its own errors and malfunctions as it goes.
It couldn't design itself. It couldn't improve itself. It won't be as sophisticated as the entirety of the global supply chain. But operating off world, where it doesn't have to worry about property lines or environmental damage or human safety or human sabotage, it could work.