Why are selfish mining attacks (γ=0) with poor propagation still profitable?

Selfish mining papers from 2013 by Ittay Eyal and Emin Gün Sirer (0) introduce variable γ.

We show the ratio of honest miners who chose to mint in pool blocks and the ratio of other (1-γ) of non-pool miners mines in other branches.

The basic idea is that the pool of attackers withholds blocks until they learn about competing blocks from honest miners. Since honest nodes prefer the first block, this variable describes the ability of an attacker to compete for blocks in front of honest blocks.

Specifically, this refers to the proportion of Minor node Attack blocks are first shown. As explained in the Simulation section:

As in real life, we assume that block propagation times are negligible compared to mining times. For two branches of the same length, artificially divide non-pool miners, and their ratio of γ mining mines into the mines on the branches of the pool and on the other branches.

The simulation generates Figure 2.

I’m confused about the most pessimistic scenario γ=0. That is, whenever a pool of attackers tries to race in front of an honest block where they fail miserably. Intuitively, I would expect such a pool to lose money consistently, but why does the red line not stay below the grey line?

Is there an additional assumption in the paper that I am missing?

(0) https://arxiv.org/abs/1311.0243