This is a hack. It is provided as is. There are no guarantees. I don't even guarantee that it has anything to do with forest, multiline nodes or TeX. At this point, in case it is not clear, I should state that I guarantee nothing whatsoever. If the code dials the CIA, MI5 and the KGB and offers them your cat in exchange for a tube of Smarties, don't say I didn't warn you.
Caveat emptor...
Here's the idea:
If we put a tabular
into a node in a tree which does not use align
to create multiline nodes, we get just the same results as we do using align
. However, tabular
itself doesn't add vertical space, so we can't set that space to 0 as it is already zero.
If we put a minipage
into a node in the same tree, the spacing is the same as putting the contents of the minipage
in the node alone. (Assuming the node has only one line.)
So, if we could create multiline nodes as minipage
s rather than tabular
s, then it would work as desired. (Whether what is desired is desirable is left for you to determine, gentle reader.)
This is not as straightforward as it sounds because a minipage
, unlike a tabular
requires a width and we don't want to have to set a width for all nodes in the tree.
Surely, there are easier ways to do this and doubtless there are far more efficient ones. Especially a TeX guru would likely do this far better.
At this point, the gentle reader is urged to review again the statement above concerning guarantees and the danger to the cats of all readers, however gentle.

We start by creating a new style squat multiline
for trees of this kind:
\forestset{
squat multiline/.style={
We need a couple of custom forest options:
declare dimen={thiswidth}{0pt},
declare autowrapped toks={thiscontent}{},
Now we start manipulating the nodes. We do this in cycles to ensure that each set of keys are processed for each node before the next set of keys is processed for each node.
for tree={
delay={
This is round one:
thiscontent/.wrap pgfmath arg={##1}{content()},
In effect, this saves the content of the node. It copies it, basically. We delay this else the content won't be set when we try to copy it.
},
Now for the next round:
delay+={
content/.wrap value={\begin{tabular}{@{}c@{}}##1\\\end{tabular}},
This wraps the content in a tabular
environment like that used by align=center
. We are going to use this to get the width for the minipage
.
typeset node,
We typeset the node now, earlier than usual, because we need to get the width next.
},
The final round:
before typesetting nodes={
We set thiswidth
using the width of the node as it currently is.
thiswidth/.wrap pgfmath arg={##1}{int(abs(max_x()-min_x()))},
Now we reset the content of the node in a minipage
of width thiswidth
. We use the copy of the contents from thiscontent
because otherwise we'll get the tabular
wrapper as well.
content/.wrap pgfmath arg={\begin{minipage}[t]{\dimexpr\forestove{thiswidth}}\centering ##1\end{minipage}}{thiscontent()},
We've already got one lot of inner xsep
because it was included when we typeset the node originally. So we set it to 0 to avoid getting double the usual amount:
inner xsep=0pt,
We need to tell forest to typeset the node again:
typeset node,
},
Just the usual anchors from the original example:
parent anchor=south,
child anchor=north,
},
}
}
Now we can use the new style. The first example compares the default spacing with that used by squat multiline
when the content of the nodes are identical:
\begin{forest}
for tree={draw}
[, phantom
[A, for tree={parent anchor=south, child anchor=north}
[B]
[C
[D]
[E]
]
]
[A, squat multiline
[B]
[C
[D]
[E]
]
]
]
\end{forest}

Then the example with larger nodes and a roof:
\begin{forest}
for tree={draw},
squat multiline
[A,
[B\\long node]
[C
[D\\long multiline, triangle]
]
]
\end{forest}

The style does not appear be responsible for the slight misalignment of the roof here as I get the same issue when I compare the default style with drawn nodes.
Comparison:

In any case, I assume that the combination of a roof and draw
is unlikely to be desirable, and the misalignment is only noticeable because the nodes are drawn for purposes of illustration.
Complete code:
\documentclass[multi,tikz,border=10pt]{standalone}
\usepackage{forest}
\begin{document}
\forestset{
squat multiline/.style={
declare dimen={thiswidth}{0pt},
declare autowrapped toks={thiscontent}{},
for tree={
delay={
thiscontent/.wrap pgfmath arg={##1}{content()},
},
delay+={
content/.wrap value={\begin{tabular}{@{}c@{}}##1\\\end{tabular}},
typeset node,
},
before typesetting nodes={
thiswidth/.wrap pgfmath arg={##1}{int(abs(max_x()-min_x()))},
content/.wrap pgfmath arg={\begin{minipage}[t]{\dimexpr\forestove{thiswidth}}\centering ##1\end{minipage}}{thiscontent()},
inner xsep=0pt,
typeset node,
},
parent anchor=south,
child anchor=north,
},
}
}
\begin{forest}
for tree={draw}
[, phantom
[A, for tree={parent anchor=south, child anchor=north}
[B]
[C
[D]
[E]
]
]
[A, squat multiline
[B]
[C
[D]
[E]
]
]
]
\end{forest}
\begin{forest}
for tree={draw},
squat multiline
[A,
[B\\long node]
[C
[D\\long multiline, triangle]
]
]
\end{forest}
\end{document}
tabular
environment by looking at the standard LaTeX definitions and playing with that. But there's nothing in forest which is just using TikZ which is just using the standard environment, I believe (but could be wrong). Note that one advantage of this is that your nodes will align better. Can you not compensate by simply reducinginner ysep
for the tree?inner ysep
. I edited my post to show why it didn't work well for me.Cy
for example).inner ysep
or not. See page 54 of the manual for discussion about why this happens and why manual intervention is better than 'fixing' the package.\mbox{}
inside a node in a tree without multiline nodes, you get exactly the same spacing. It is just the effect of thetabular
environment and the way it aligns things, for example.