(I've deleted the previous answer which was wrong.)

Now I have, without proof, by heuristics a direct formula for the entries in the weight matrix if we consider its entries as
$$ W=\small \begin{bmatrix}
0 & . & . & . & . & . \\
1/2 & 1/2 & . & . & . & . \\
3/4 & 3/2 & 3/4 & . & . & . \\
7/8 & 17/8 & 17/8 & 7/8 & . & . \\
15/16 & 5/2 & 25/8 & 5/2 & 15/16 & . \\
31/32 & 87/32 & 61/16 & 61/16 & 87/32 & 31/32
\end{bmatrix}$$ as it was also shown by Rob Johnson.

It was easier to find the recurrences and then the direct formula when I included the unit-weight per person, such that I have:
$$ X= \small \begin{bmatrix}
1 & . & . & . & . & . \\
3/2 & 3/2 & . & . & . & . \\
7/4 & 5/2 & 7/4 & . & . & . \\
15/8 & 25/8 & 25/8 & 15/8 & . & . \\
31/16 & 7/2 & 33/8 & 7/2 & 31/16 & . \\
63/32 & 119/32 & 77/16 & 77/16 & 119/32 & 63/32
\end{bmatrix}
$$ or, expanded to make it better visible, rowscaled by powers of 2:
$$ Y =\small \begin{bmatrix}
1 & . & . & . & . & . \\
3 & 3 & . & . & . & . \\
7 & 10 & 7 & . & . & . \\
15 & 25 & 25 & 15 & . & . \\
31 & 56 & 66 & 56 & 31 & . \\
63 & 119 & 154 & 154 & 119 & 63
\end{bmatrix} $$

Let $r$ denote the rowindex, $c$ denote the column-index, both beginning at zero, $c \le r$ then we have
$$ w_{r,c} = 1+2c - {\sum_{k=0}^c \binom{r+2}{k}\cdot(1+c-k)\over 2^r}$$

In terms of the $X$-matrix this was:
$$ x_{r,c} = 2\cdot(1+c) - {\sum_{k=0}^c \binom{r+2}{k}\cdot(1+c-k)\over 2^r}$$
and in terms of the $Y$-matrix this was
$$ y_{r,c} = 2^{r+1}\cdot(1+c) - \sum_{k=0}^c \binom{r+2}{k}\cdot(1+c-k)$$

*[added]* I found it interesting to consider a more general aspect. The idea is to assume an infinitely wide pyramid of "zero-humans" or "ghosts" which have weight $o$ (which for the final evaluation is then to set to zero). This is the infinitely wide grey shaded area in the following picture. The structure for this area is very simple and homogenuous: in each row $r$ the weight in one entry is $r \cdot o$.

Now one of the top-most ghosts "gets a soul", becomes human and thus gravitation gives him a pressing weight "w". All ghosts which are touched by his feet become then human too, and their weight changes to "w" as well, and this iterates downwards.

The picture shows, how this distributes to the bottom:

Note, how the coefficients at a *w* and at the *o* in the same cell add to make the equilibrium-coefficient of the same row, if we would set $w=o$.

It is in general the description of a certain disturbance inserted into a equilibrium, and I'm curious, whether this can even be made continuous (like the generalization of the binomial-coefficients using the Gamma-function)

*[added2]* With iterations of this type it is often useful to express the problem in a matrix-notation and convert iteration into matrix-powers. Then sometimes one can find a closed form in terms of fixed matrices and the iteration-parameter is then only in the exponent of a matrix. (Clearly that does not finally solve the problem of the finite sums, sometimes recursive, in the original, non-matrix, representation but moves it into the internal dotproducts for the matrix-powers. But for the reader of the formula this might be convenient).

First let's define a rowvector $U=[1,0,0,...]$ of some dimension *n* which represents the weight of the person on the top where *n* is the final width of the pyramide.

Then one iteration should have the effect, that $U_0$ changes to $V_1=[1/2,1/2,0,0,...]$ which is the received weight for the lower layer and then $U_1$ must also have the unit-weights for the persons to become $U_1=[3/2,3/2,0,0,0...]$ .

A matrix, which performs $U_0 \to V_1$ is $M= \frac12 \small\begin{bmatrix}1&1&0&0&0&0&...\\0&1&1&0&0&0&...\\0&0&1&1&0&0&... \\ \vdots & & & & & &\ddots \end{bmatrix}$ such that $$U_0 \cdot M = V_1 = \left[ \frac 12,\frac12, 0,0,0,... \right] $$.

After that we need a matrix, which transfers two ones for the unit weights into $V_1$ and we can do this simply by addition of the unscaled $M$ such that
$$ U_0 \cdot 2M = [1,1,0,0,...]$$ and we have
$$ U_0 \cdot M + U_0 \cdot 2M = U_1 = [3/2,3/2,0,0,...]$$

But now, if the number of persons increase by the iterations, the second summand is not sufficient; we need so many unit-weigths as we have persons in one iteration. So we need to find a matrix, whose k't power gives k leading ones in the top row. A simple candidate is the matrix of the upper unit-subdiagonal, let's call it $L$:
$$ L = \small \begin{bmatrix}0&1&0&0&0&0&...\\0&0&1&0&0&0&...\\0&0&0&1&0&0&... \\ \vdots & & & & & &\ddots \end{bmatrix} $$
Then the k'th power of shifts the diagonal to the right to the k'th position, and the sum of $L^0 + L^1 + L^2 + ... + L^K$ gives $k+1$ ones.
Then for the first iteration we can write
$$ U_0 \cdot M + U_0 \cdot (L^0 + L^1) = U_1 $$ or
$$ U_0 \cdot (M + (L^0 + L^1)) = U_1 $$

We can then simply proceed by
$$ U_1 \cdot (M + (L^0 + L^1 + L^2)) = U_2 $$ which is also
$$ U_0 \cdot (M^2 + (L^0 + L^1)M+ (L^0 + L^1 + L^2)) = U_2 $$
and it is obvious how this iterates.

But we see now, that we have partial geometric series depending on the iteration-parameter, and they have a closed form, namely the cyclotomic polynomial of the k'th order.

For the cyclotomic polynomial we need to write $$ L^0+L^1+... + L^k = (I - L^{k+1}) \cdot (I - L)^{-1} $$ Let's denote that inverse as $R$ . We find that it is just the upper triangular unit-matrix (as expected by the series representation!) $$R = \small \begin{bmatrix}1&1&1&1&1&1&...\\0&1&1&1&1&1&...\\0&0&1&1&1&1&... \\ \vdots & & & & & &\ddots \end{bmatrix} $$
Then we have
$$ U_0 \cdot ((I-L^1)R M^k + (I-L^2)R M^{k-1}+ (I-L^3)R M^{k-2} +...+(I-L^{k+1})R) = U_k $$ (where I smoothed the formula at the leading $M^k$

Because it is also $M=I + L$ the matrices $M$ and $L$ and $R$ commute all and we can rotate the coefficients for any convenience. We can isolate a geometric sum of powers of $M$:

$$ U_0 \cdot ( R( M^k + M^{k-1}+ M^{k-2} +...I) \\
-R(L^1 M^k + L^2 M^{k-1}+ L^3 M^{k-2} +...+L^{k+1})) = U_k $$
Now the powers of $M$ are also partial geometric series, and we need the reciprocal, let's denote it as $S$: $$ S= (I - M)^-1 =\small \begin{matrix} 2&2&2&... \\ 0&2&2& ...\\ 0&0&2&...\\ ... \end{matrix}$$ which happens to be just $S=2R$ and is also commuting with every matrix in the game so far.

Then we have
$$ U_0 \cdot \left( R( I - M^{k+1})S \\
-R(L^1 M^k + L^2 M^{k-1}+ L^3 M^{k-2} +...+L^{k+1}) \right) = U_k \\ $$

and the last parenthese of variable length can be reduced similarly because $M$ is invertible and with a matrix $W=L \cdot M^{-1}$ and $T=(I -W)^{-1} $we have
$$ U_0 \cdot \left( R( I - M^{k+1})S -R L M^k (I + W+ W^2 +...+W^k) \right) \\
= U_0 \cdot \left( R( I - M^{k+1})S -R L M^k (I -W^{k+1}) T) \right) \\
\qquad \qquad = U_k \\ $$

This formula has a fixed length for any iteration parameter. Only the variable matrix
powers contain that parameter. (Possibly from this one can invent fractional iterates, but that's not the focus here...)