I need to provide a 1 pulse per second 'heartbeat' signal with a very good accuracy over an extended period of time (± approx. 200 μs for several hours).
If 200µs applies to 1s period that would be 200ppm which is easy. So I assume instead you mean 200µs total drift over several hours. Say over 4 hours, that rounds to 10ppb, For this you need OCXO with a specification of "time drift over N hours < 200µs".
I don't understand about the VCO part. If you want that kind of stability with a VCO then you need the same kind of stability on the control voltage.
Instead you could measure the OCXO frequency with the required accuracy, then simply use a divider/counter to output your 1Hz signal. In this case the OCXO frequency doesn't matter, only stability matters, so maybe you can use a cheaper OCXO.
If the OCXO is around 10MHz, to get the required 200µs drift over several hours (say 4 hours) you'll need to measure its frequency with an accuracy of 10ppb so about 0.1Hz which is... maybe a little bit complicated.
Then you need a fractional counter, trivial in a FPGA.
There is an external synchronisation pulse, which I get once, and sets the time reference (2 pulses, 1 sec apart from an atomic standard), I then need to maintain that 1 second reference from then on (/vital edit). After these two pulses, there is then no connection to the frequency standard
Now that's a can of worms, because if you want to measure than 1 second pulse, seeing it only once, with an accuracy of 10ppb... that's 10ns... let's hope there's no noise on the edges and they're sharp enough. In fact the rise time is probably going to eat your error budget. And you're going have to implement a time to digital converter, or maybe just use a PLL to multiply the 10MHz OCXO to 500MHz and then count the periods of that clock during the 1 second pulse.
Personally since this project feels a bit expensive, I would make the FPGA board output the 10MHz OCXO frequency, and feed that to an off the shelf frequency meter synchronized to the atomic clock. Then the frequency meter tells the FPGA (via GPIB) what its OCXO frequency is, and then it's calibrated.