This is a good question. if you look at applyStrategy you'll see that each symbol in the loop is run in isolation -- independently. You may want to check out applyStrategy.rebalancing which does a nested loop of the form:
for(i in 2:length(pindex)){
#the proper endpoints for each symbol will vary, so we need to get
#them separately, and subset each one
for (symbol in symbols){
#sret<-ret[[portfolio]]
This means it loops over a section of timestamps, and then for each symbol, which is what you want when you want some interactions between symbols (applyStrategy simply does a for symbols, then an inner by timestamp loop, so you'll never get interactions going on) in terms of equity.
When I first starting using quantstrat I initially had the same frustration. My solution was to modify applyStrategy.rebalancing do become a (slower) double loop, for each time stamp, then an inner loop across each symbol.
Yes, this means you can't directly compute portfolio PL accurately in quantstrat. So things like opening positions that are position of current portfolio equity can't be done directly. (But you can modify the code to do it if you want).
Why does quantstrat behave this way by default? The authors will give you good reasons. In short, my view is that (after some brief discussions with the authors), if a signal has predictive power, and gives you an edge in a strategy, it'll work regardless of how you combine it with other symbols later. quantstrat is about identifying whether signals are good or not in relation to mktdata you pass to it.
Logically, if a signal is good on a per symbol level, then it will likely do ok on a portfolio level too (if not better, with smoother portfolio PL). quantstrat's current approach will give you a reasonable approximation to what portfolio PL will look like, but not in a true "compounding return" sense. To do that, you'd want to scale your positions according to the current portfolio PL (which isn't possible in applyStrategy as noted above). This simplification of running a strategy per symbol only also makes simulations much, much faster. Note that you can introduce still interactions with other symbols in applyStrategy by adding additional columns to the symbol data that relate to other symbols though (e.g. in pairs trading, etc).
At the end of the day backtest results are always simplifications of trading in reality, so there isn't a big motivation to get "super" accurate backtest results that project profit/trading revenue very accurately.
