\section{Preliminaries}\label{sec:prelim} \subsection{Conventions}\label{sec:conventions} \paragraph{Reduction to a simple graph.} Site percolation on a graph with loops or multiple edges is site percolation on its underlying simple graph $G_s$, which has the same vertex set and in which $x\sim y$ if and only if $x\ne y$ are joined by at least one edge. The graph $G_s$ is infinite, connected and locally finite; $\Aut(G)$ acts on $V$ through a subgroup of $\Aut(G_s)$, so $G_s$ is quasi-transitive; and vertex boundaries, hence $h_V$, are unchanged. All site quantities ($p_c$, $p_u$, $\ptwo$, $T_p$ and the label coupling) coincide for $G$ and $G_s$. Unless stated otherwise, \emph{graphs are simple from now on}, $d<\infty$ bounds the degrees, and unimodularity refers to $\Aut(G_s)$. In contrast with~\cite[Lemma~2.1]{OAINU}, which subdivides multiple edges for bond percolation, simplification adds no new sites. \paragraph{Site notation.} A configuration is $\omega\in\{0,1\}^V$, with $1$ meaning open, and $\Pp_p$ is the product Bernoulli($p$) law. Domains are vertex sets $D\subseteq V$ with their induced graphs. The cluster $C_x^D$ is the set of vertices joined to $x$ by an open path inside $D$; it is \emph{empty when $x$ is closed or $x\notin D$}. Put $C_x=C_x^V$, $s_x=|C_x|$ and $T_p^D(x,y)=\Pp_p(y\in C_x^D)$, so $T_p(x,x)=p$. For $A\subseteq V$, $\partial A$ is the outer vertex boundary, $\Nb[A]=A\cup\partial A$, $N(z)$ is the set of neighbours of $z$, and $\Nb[z]=\Nb[\{z\}]= N(z)\cup\{z\}$. Inside a domain $D$, for $W\subseteq D$, put $\partial_DW=\partial W\cap D$ and $\Nb_D[W]=W\cup\partial_DW$. Every vertex of $\partial C_x$ is closed, and adjacent open vertices share a cluster. $\Adj$ is the adjacency kernel. For a closed vertex $v$ put \[ \widehat C_v=\{v\}\cup\bigcup_{a\sim v}C_a, \] the cluster of $v$ after opening it. If $C^{(1)},\dots,C^{(m)}$, $m\le d$, are the distinct open clusters adjacent to a closed $v$, with sizes $n_i$, put $N_v=\sum_in_i=|\widehat C_v|-1$ and $Q_v=\sum_{i\ne j}n_in_j$, a sum over ordered pairs. We use the constants \begin{equation}\label{eq:constants} \kappa_p=\Bigl(\frac{1+d}{p}\Bigr)^2,\qquad C_{\mathrm{del}}(p)=\frac1p+d,\qquad c_{\mathrm{nb}}(p)=1+\frac dp+\frac{d^2}{p^2}, \end{equation} all bounded on $p\ge p_c/2$, and two consequences of Harris's inequality. (H1) If $z\sim a$ with $a,z\in D$, then $p\,T_p^D(v,z)\le T_p^D(v,a)$, since $\{v\lr z\text{ in }D\}\cap\{a\text{ open}\} \subseteq\{v\lr a\text{ in }D\}$; iterating, $p^lT_p^D(v,z)\le T_p^D(v,a)$ when $z$ and $a$ are joined by a path of length $l$ in $D$. (H2) $\delta_{wy}\le T_p(w,y)/p$, since $T_p(y,y)=p$. Consequently, entrywise, \begin{equation}\label{eq:H} I+\Adj\,T_p\le\frac{1+d}{p}\,T_p,\qquad I+T_p\Adj\le\frac{1+d}{p}\,T_p. \end{equation} Kernel facts from~\cite[Section~2]{OAINU}: if $|A(x,y)|\le B(x,y)$ with $B$ nonnegative and bounded, then $\norm A\le\norm B$; if all row and column sums of $B\ge0$ are at most $b$, then $\norm B\le b$. \subsection{Types, mass transport, trace and walk}\label{sec:types} In the unimodular case we use the type averages of~\cite[Section~2.3]{OAINU}. Let $\Gamma=\Aut(G)$ be unimodular, with vertex orbits $O_1,\dots,O_{h_0}$ and representatives $o_1,\dots,o_{h_0}$. With a fixed Haar measure on $\Gamma$, put \[ \sigma_i=\frac{\operatorname{vol}(\Gamma_{o_i})^{-1}}{\sum_j\operatorname{vol}(\Gamma_{o_j})^{-1}}, \qquad[F(o)]=\sum_{i=1}^{h_0}\sigma_iF(o_i),\qquad c_\sigma=(\min_i\sigma_i)^{-1}. \] A bracket $[F(o)]$ around a function of the root $o$ always denotes this finite type average; brackets after $\E_p$ are ordinary expectation brackets. Operator norms, adjoints and positivity refer to counting-measure $\ell^2(V)$. Let $P$ be the symmetric lazy walk, $P(x,y)=1/(2d)$ for $x\sim y$ and $P(x,x)=1-\deg(x)/(2d)$. We use the following statements of the unrefereed preprint~\cite{OAINU} (Section~\ref{sec:earlier}) as stated; they concern only the graph, invariant operators, or finite product measures. \begin{lemma}[{\cite[Lemma~2.7]{OAINU}}]\label{lem:MTP} For every invariant nonnegative function $m$ on ordered vertex pairs, $\bigl[\sum_ym(o,y)\bigr]=\bigl[\sum_xm(x,o)\bigr]$. The identity also applies to the expectation of a covariant random mass and to absolutely summable signed or complex masses. In particular, an invariant stochastic kernel on a single orbit has column sums one. \end{lemma} \begin{lemma}[{\cite[Lemma~2.8]{OAINU}}]\label{lem:trace} For a bounded invariant operator $S$, $\tr S=[S(o,o)]$ is a positive normalized trace: $\tr I=1$ and $\tr(AB)=\tr(BA)$ for bounded invariant $A,B$. If $A$ is positive semidefinite, invariant and bounded, and $B=B^*$ is invariant and bounded, then $|\tr(BA)|\le\norm B\tr A$ and $A(x,x)\le c_\sigma\tr A$. \end{lemma} \begin{lemma}[{\cite[Lemma~2.9]{OAINU}}]\label{lem:walk} $P$ is positive semidefinite and $\rho:=\norm P<1$. Its projection onto the vertex types is an irreducible lazy finite Markov chain, and there are $c_*>0$ and $l_0<\infty$ with $\sum_{y\in O_j}P^l(x,y)\ge c_*$ for every $x\in V$, $j\le h_0$ and $l\ge l_0$. \end{lemma} The bound $\rho<1$ is where nonamenability enters the unimodular argument; Section~\ref{sec:discussion} lists its uses. \subsection{Site tools}\label{sec:tools} The following site forms of the preliminary lemmas of~\cite[Section~2]{OAINU} are proved in Appendix~\ref{app:tools}. They are routine adaptations, except that the Russo identity acquires a remainder term. Lemma~\ref{lem:crit-finite} holds on every graph of Theorem~\ref{thm:main}; the others are stated where they are used. Put $\chi_x(p)=\E_ps_x$ and $\chi_{\max}(p)=\sup_x\chi_x(p)$, and, when $\Aut(G)$ is unimodular, $\chi(p)=[\chi_o(p)]$. \begin{lemma}[Critical finiteness]\label{lem:crit-finite} Let $G$ be as in Theorem~\ref{thm:main}, simple. Then $00$ forces exponential growth. \begin{lemma}[Association, disjoint witnesses, tree bound]\label{lem:tools} Increasing nonnegative functions of the site bits are positively correlated, and increasing events with disjoint finite open witnesses satisfy the van den Berg--Kesten inequality. For all $x,y,z$ and every domain $D$, \begin{equation}\label{eq:tree} \Pp_p(x\lr y,\ x\lr z\text{ in }D)\le\kappa_p\sum_wT_p^D(x,w)T_p^D(w,y)T_p^D(w,z). \end{equation} In particular $\E_ps_x^2\le\kappa_p\chi_{\max}(p)^3$. \end{lemma} \begin{lemma}[{Threshold order; site form of~\cite[Lemma~2.4]{OAINU}}]\label{lem:threshold-order} For every infinite, connected, locally finite, quasi-transitive graph, $\ptwo\le p_u$ for site percolation. \end{lemma} \begin{lemma}[{Subcritical finiteness and type comparison; site form of~\cite[Lemma~2.10]{OAINU}}] \label{lem:sharpness} Let $\Aut(G)$ be unimodular. For $p0$ with $\chi_x(p)\ge c\chi(p)$ for all $x$ and $p_c/2\le p