Array structural split/join allocates and copies (blk_half/blk_node); the paper says both halves come in place and joining is free

What you did

Compared the runtime paper's array cost model with the emitter and runtime routines in the current tree (source reading only; no timing was measured).

What happened

The paper (bend2/docs/BendRT/main.typ, § Arrays): "A match on the tree takes both halves in place, and joining two adjacent halves back is free."

The emitter (bend2/comp.ts at 0b7e2b11): an Array ANode{l, r} constructor emits blk_node(e, l, r) (line 2228), and a match on ANode emits two blk_half(e, sw, hi) calls (line 2647). Those routines allocate and copy: blk_node (line 4068) allocates the merged block, copies both halves word by word, then frees both inputs; blk_half (line 4093) allocates and copies each half, then frees the source on the high call. The compiler's own comment (lines 4007-4016) states it: "A match on ANode is blk_half twice: each half allocated in its class and copied... ANode{l, r} is blk_node: the merged class, l and r copied and freed shallow... A match to the leaves copies O(n log n) words where a view copied none; get, set, swap, size and new open no half."

So for materialized structural split/join (indexed get/set opens no half), the paper's cost model does not match the implementation. Either the paper should say allocate-and-copy, or the views should be implemented.

bend --version

bend 2.0.5

uname -sm

Linux x86_64