Three codehz/ecs perf-suite tests fail: iterator/collection method lost through wrapped types (next/add "is not a function")

[proggeramlug]

Running the codehz/ecs (v0.13.0) comprehensive.perf.test.ts suite with all seven tests enabled (the perf harness normally skips six), three fail under perry while all pass under node 26.5.1:

test	error
should spawn and sync entities efficiently	next is not a function
should handle mixed operations in a realistic game loop	next is not a function
should handle exclusive sparse relation flips efficiently	add is not a function

Established so far

• Not a regression. Bisected by building pre-perf(runtime,codegen): SSO computed keys + megamorphic write stub cache (dynamic-property overwrite 5.3×) #8965 main (bdd395216) with all seven rows enabled: the same three fail with the same errors. This predates the 2026-08-28/29 perf work.
• The obvious minimization does not reproduce. The suspect pattern from archetype/store.ts:166-172 — let s = map.get(k); if (!s) { s = new Set(); map.set(k, s); } s.add(v) — compiled standalone behaves identically to node (sizes, has, all correct). So the trigger is narrower than the raw Map-of-Set idiom.
• The failing tests exercise: entity spawn + world.delete cleanup loops, the mixed-ops game loop, and sparse/exclusive relations (component({ sparse: true, exclusive: true }) + relation(ChildOf, parent)). The passing four never touch sparse relations, so getComponentIdFromRelationId / the sparse-relation store paths are prime suspects for at least the add failure.
• next is not a function is the classic symptom of GetIterator returning a non-iterator (perry's js_get_iterator falls through to returning the value itself when no [Symbol.iterator] resolves), so some receiver shape in these tests resolves no iterator where node finds one.

Repro

# prepared workspace via secret-tests/ecs-suite/run.py::prepare_workspace on ecs-suite/.ecs
# then enable all it( ) tests and compile:
perry compile src/__tests__/perf/comprehensive.perf.test.ts -o ecs.bin
./ecs.bin   # => 7 tests | 4 passed | 3 failed

Suggested next steps

1. Instrument the bun-test shim to print e.stack for the three failures (the shim currently swallows it) — one run pins the exact source lines.
2. Bisect the failing test bodies (they are short) rather than the library idiom.

Worth fixing beyond correctness: the three hidden rows include the sparse-relation benchmark, so perf coverage of that path is currently zero.