PSVec

cats.Foldable — index-based foldLeft, Eval-deferred foldRight (stack-safe), O(1) size.

cats.Foldable — index-based foldLeft, Eval-deferred foldRight (stack-safe), O(1) size.

cats.Functor — index-based map into one fresh backing array (0/1-element results normalise through unsafeWrap to Empty / Single).

cats.Functor — index-based map into one fresh backing array (0/1-element results normalise through unsafeWrap to Empty / Single).

cats.Traverse — the instance Traversal's .modifyA routes through when the carrier is MultiFocus[PSVec]. Effects run left-to-right in index order; results land in one pre-sized backing array (a TraverseBuf) threaded through G.map2.

That array must not be blindly shared: G may resume a continuation more than once — a multi-shot effect (List, Vector, Chain, LazyList, fs2 Stream, …) once per branch, a product effect (Tuple2K, cats' parallel-composition law) once per component, a re-runnable effect (IO, State, Kleisli, …) once per run. Two rules keep every resumption apart:

• The buffer is allocated INSIDE the first element's continuation, not in a G.pure seed, so every branch of f(fa(0)) and every run of the returned G starts its own buffer (no shared seed — which is also what makes two concurrent runs of one IO safe).
• Later continuations CLAIM their slot (TraverseBuf.put): the first writer at i writes in place, every further writer at i copies the prefix [0, i) — frozen, since slots are only ever written at the claim frontier — into a buffer of its own.

Single-shot effects (Option, Either, Id, Eval, IO, Validated, …) thus pay what a shared array costs (one array plus one small buffer object, no per-element structure), and multi-shot effects pay one prefix copy per extra branch, bounded by the size of the result they build anyway. The claim is a plain field, not a CAS (a CAS measured ~2× slower here): it assumes no single run of G resumes continuations on one intermediate value from two threads at once, which nothing that re-evaluates its inputs (IO, fs2, Kleisli) does.

cats.Traverse — the instance Traversal's .modifyA routes through when the carrier is MultiFocus[PSVec]. Effects run left-to-right in index order; results land in one pre-sized backing array (a TraverseBuf) threaded through G.map2.

That array must not be blindly shared: G may resume a continuation more than once — a multi-shot effect (List, Vector, Chain, LazyList, fs2 Stream, …) once per branch, a product effect (Tuple2K, cats' parallel-composition law) once per component, a re-runnable effect (IO, State, Kleisli, …) once per run. Two rules keep every resumption apart:

• The buffer is allocated INSIDE the first element's continuation, not in a G.pure seed, so every branch of f(fa(0)) and every run of the returned G starts its own buffer (no shared seed — which is also what makes two concurrent runs of one IO safe).
• Later continuations CLAIM their slot (TraverseBuf.put): the first writer at i writes in place, every further writer at i copies the prefix [0, i) — frozen, since slots are only ever written at the claim frontier — into a buffer of its own.

Single-shot effects (Option, Either, Id, Eval, IO, Validated, …) thus pay what a shared array costs (one array plus one small buffer object, no per-element structure), and multi-shot effects pay one prefix copy per extra branch, bounded by the size of the result they build anyway. The claim is a plain field, not a CAS (a CAS measured ~2× slower here): it assumes no single run of G resumes continuations on one intermediate value from two threads at once, which nothing that re-evaluates its inputs (IO, fs2, Kleisli) does.