Why is Spectre ignoring time tolerance or time step conditions from Verilog-AMS models?

Dear community

I am using the following Verilog-AMS model to generate sharp pulses:

module my_source(MINUS, PLUS);

inout MINUS;
electrical MINUS;
inout PLUS;
electrical PLUS;

parameter real delay = 1.0u;
parameter real period = 10.0u;
parameter real amplitude = 8.0;

real out;

analog begin

@(initial_step) begin
out = 0;
end

@(timer(delay,period)) begin
out = amplitude;
end

@(timer(delay+2n,period)) begin
out = 0;
end

I(PLUS, MINUS) <+ transition(out*25e-9,0,1n);

end
endmodule

This signal is further amplified by some analog circuit. The problem is that the observed amplitude at the end of the signal chain varies over time unless I set a very strict reltol of 1e-6 or force a small maximum step size in the transient simulation, e.g. 50ps. This slows down the simulation a lot. To prevent that, I want to force the simulator to temporarily limit its step size. This can be done with the $bound_step function, but for whatever reason, Spectre seems to ignore that. So in my code I add

@(timer(delay,period)) begin

$bound_step(50p);
out = amplitude;

end

and then reset it after 300ns

@(timer(delay+300n,period)) begin

$bound_step(1m);

end

According to the log, the step size is not limited to 50ps and the amplitudes still vary over time. However, if a call $bound_step outside of the timer function, then Spectre does indeed use a step size of 50ps, but this is then applied for the complete simulation and thus very slow. So why is Spectre ignoring the call of $bound_step inside the timer function? Are there other ways to achieve the desired outcome?

Thanks for any suggestion.