bench/boundary: a compound-element-type lane (#261)

[lannbot]

Refs #261 (checklist item: "add a compound-element-type lane to bench/boundary").

Every payload shape in the instrument was a flat scalar — list<u8>, u32, stream<u8> — exactly the set #63/#67 moved onto bulk copy paths. Nothing in bench/boundary exercised the per-element interpreted lift/lower loop that compound types actually walk, so #261's ~5 µs/element finding had no regression instrument and no way to be reproduced in-repo. This lands the instrument before the optimization it will measure, so the "before" numbers are honest.

Shapes

export	measures
lift-ops: async func(n: u32) -> list<op>	compound-element lift (cabi/load.ts per-element path)
lower-ops: async func(ops: list<op>) -> u64	compound-element lower (cabi/store.ts per-element path)

op is a 16-case variant over records, with string, option<u16>, list<u8> and a nested variant across the payloads — the width and payload mix mirror the consumer schema #261 reports (17 cases). The width is load-bearing, not decoration: maxCaseAlignment (cabi/layout.ts:80, called from loadVariant) and the embedder facade's variant case resolution (embedder/values.ts:212) are both O(case count) per element. An earlier revision of this lane used a 4-case variant and measured ~3.5x below the reported figure — the instrument, not the box.

Both directions because load.ts and store.ts are separate code with the same defect, and a lift-only lane would leave half of it uninstrumented.

Methodology

Reported as ns/element — the unit that makes these comparable to #261's numbers. The guest caches its Vec<Op> per n in a thread_local!; the warmup builds it and timed calls clone it, so format!-per-element is out of the timed region. The clone residue was measured in isolation (temporary export, cloning without crossing the boundary) at ~15 ns/element, under 0.5% of the figures below — so the lane is measuring the boundary. lower-ops's host array is built once outside the timed loop.

Pre-optimization baseline

compound-element lanes (ns/element; n=10000; jco lane skipped):
shape         polyengine-node-callback      polyengine-node-jspi  polyengine-deno-callback
lift-ops                       3,804.4                   3,909.3                   3,184.6
lower-ops                      3,482.3                   3,646.8                   3,285.3

Within ~1.6x of #261's reported ~5 µs/element on its box — the residual reads as box/config drift, not a methodology defect.

Scope

bench/boundary/** only. No published surface, no version bump, no contract change. just check passes; just version-guard-local passes.