Keep duplicate column naming linear

src/core.jl

@@ -4666,18 +4666,24 @@ materialize(v::Vector) = v
 # A duplicate takes the smallest `name_k` not used by ANY
 # name — original or already assigned — so `a,a,a_1` becomes `a,a_2,a_1`
 # (renames never collide with an original that appears later).
+#
+# `taken` only grows, so every suffix below the last one assigned to a name is
+# still taken: the search resumes there, which keeps N copies of one name linear.
 function makeunique!(names::Vector{Symbol})
     taken = Set(names)
+    length(taken) == length(names) && return names
     seen = Set{Symbol}()
+    nextk = Dict{Symbol, Int}()
     for i in eachindex(names)
         nm = names[i]
         if nm in seen
-            k = 1
+            k = get(nextk, nm, 1)
             newnm = Symbol(nm, :_, k)
             while newnm in taken
                 k += 1
                 newnm = Symbol(nm, :_, k)
             end
+            nextk[nm] = k + 1
             push!(taken, newnm)
             names[i] = newnm
             push!(seen, newnm)

test/api.jl

@@ -276,6 +276,17 @@ end
     @test_throws ArgumentError A.File(IOBuffer("a,b\nx,y\n1,2\n"); header=[3, 1])
 end
 
+@testset "duplicate column names" begin
+    @test Base.names(A.File(IOBuffer("a,a,a\n"))) == [:a, :a_1, :a_2]
+    @test Base.names(A.File(IOBuffer("a,a_1,a\n"))) == [:a, :a_1, :a_2]
+    @test Base.names(A.File(IOBuffer("a,a,a_1\n"))) == [:a, :a_2, :a_1]
+    # Each copy resumes the suffix search, so many copies of one name stay
+    # linear; a search that restarts at 1 for every copy allocates gigabytes.
+    n = 4_000
+    @test A.makeunique!(fill(:x, n)) == [:x; [Symbol(:x_, k) for k in 1:(n - 1)]]
+    @test (@allocated A.makeunique!(fill(:x, n))) < 100_000_000
+end
+
 @testset "row windowing and source modes (raw-row semantics)" begin
     sourceparity("a,b\n1,2\n3,4\n5,6\n"; kw=(; limit=2))
     sourceparity("a,b\n1,2\n3,4\n5,6\n"; kw=(; footerskip=2))