For a start, the tool is called Dracula not PDRACULA. PDRACULA is the rule compiler, not the name of the tool itself.

How the substrate is extracted is entirely dependent on how the rules were written. You can get it to extract the substrate however you want. However, bear in mind that the substrate is often extracted as being conducting (because it is conducting), and then you can only have a single connection to substrate - all connections will be shorted together (which reflects the true connectivity of the substrate). I suspect this is what is happening here.

Often substrate connections are handled using the SCONNECT command in the rule file - which does a soft connect. A soft connect is effectively a one-way connection,  where the connection is only made downwards (to flag problems where you connect through the substrate), and will also flag multiple stamps onto the substrate from different nets (look at your ERC results). However, from the LVS point of view, one of these connections will win, and you'd then see substrate mismatches in the LVS results.

In a typical mixed signal environment, you want to be able to have multiple ground connections, even though the substrate is really shorting them together. Since the metal conducts better than the substrate, the substrate in a particular region would be tied to the local metal ground, and the coupling between the grounds via the substrate would be minimised (although not eliminated entirely).

Some rule decks provide support for this kind of grounding structure by having dummy layers which allow you to isolate one region of substrate from another - often by drawing a thin path around the region you want to fence off (although sometimes by drawing a polygon to represent the isolated area). This is not physical isolation, but something to tell Dracula to cut the substrate into pieces (often it just does a "NOT" of the substrate with this isolation dummy layer before extracting the connectivity). Once this cutting has been done, you can then mark off various regions, and within each region, that local substrate can be connected to one ground net only.

But as I said, exactly what can be done depends entirely on the rule deck. It's not a problem with Dracula, but a problem that all physical verification tools face when trying to balance the fact that physically the substrate is connected, yet you want to be able to connect multiple nets to the substrate (i.e. you really have a short, but you don't want it to be treated as a short in some cases).

Regards,

Andrew.