The Atlantic article linked above covers this in more detail, and is an excellent read all-around.
The old system will be based on detecting trains in "blocks", sections of track insulated from each other where the train's wheels make the circuit. I was about to say this would use relays, but Wikipedia says the first all-relay system was installed in 1929. So it's at least partly mechanical (levers moving in the way of other levers). The logic implemented in this will prevent the signalman from sending trains on a collision path, but only if drivers obey all red lights.
The new system will either be something similar, except implemented in electronics and stopping the train if the driver doesn't. But for a metro system, in order to get higher capacity, it will probably do away with the "blocks" altogether, instead ensuring that every train has a safe distance ahead of it. Maybe they will also automate the trains, so the train drives itself.
Either way, they'll probably upgrade the trains to support the new system, then install the new system in whatever sections allow it, then remove the old system once the new system is working correctly. I'd expect there's a huge amount of sunk cost once the project starts (design, modifications to trains, staff training etc) so it doesn't make any sense to stop halfway and keep the costs of both systems.
The old mechanical systems aren't really unsafe. They're just inefficient, full of moving parts so they break all the time, and require the transit agencies to run their own workshops to maintain them because they're so obsolete.
The newer computerised systems tend to have no moving parts and can pack more trains onto the tracks.
Perhaps the intent was to say something like "you can't reap marginal benefits from the replacement of a single signal since the whole system must be modernized to allow the advanced signalling afforded by new technology," or something.
Its possible to imagine numbers where the cost of being dual capable is so high, the cheapest solution is some kind of forklift upgrade of the whole system. Which is kinda what they're doing.
A good hardware analogy is you have a company thats a windoze PC shop. Maybe even all the hardware comes from the same Dell contract, same OS image on every box, its pretty cheap. Surely OSX would improve any end user experience, but imagine the cost of instantly doubling the helpdesk workload to now handle two kinds of hardware and two OS images and two licensing agreements and generally two of everything.
Or a good software analogy is given a somewhat reliable mysql cluster, dropping postgresql on it and trying to use both at the same time is likely to result in some heartache.
Nothing, of course, despite the naysayers. You wouldn't, initially, try to replace - just augment; record and report the inputs and be a test-bed for the eventual digital replacement. In this case, because the originals are physical switches, it'd be more like a modern-machined duplicate of the original with a micro-controller stuck on the side.
The data it would send be immediately helpful for arrival timers and such.
Presumably you can't use anything currently manufactured.
So, you'd have to commission a unique, backward compatible design. And test it. And qualify it.
And no company is going to do that without you basically paying them the same up-front amount as if you just contracted to replace all the signals.
Does anyone have any idea why this is? What prevents new signals from being backwards compatible with the old system?
And if this were true, wouldn't all of the signals need to be the same hundred year old type? How would line extensions ever be able to be done in that case? Is the system segmented in some sense? If it is segmented, what prevents further segmentation to allow incremental updates of the signals?