Set/Map forEach with mid-iteration delete visits holes and skips entries (gap test red on main) #9072

test-files/test_gap_set_map_foreach_fused_receiver.ts fails on clean main — cases 5–7 (delete during Set.prototype.forEach / Map.prototype.forEach) return wrong elements, and test-parity/gap_snapshot.json expects the test to pass, so any full gap run is red on it.

Node 26.5.1 (pinned oracle):

[1,2] 0
["a:1","b:2","c:3"] 0
["p","q","r"] 2

Perry (both main @ 86c910471 and a current feature branch, perry-dev builds, PERRY_NO_AUTO_OPTIMIZE=1 — byte-identical on both arms, so pre-existing):

[1,null] 1
["a:1","NaN:undefined","b:2"] 1
["p","q"] 2

The shape is exactly what the test's own header warns about: delete during forEach used to compact the backing vector and was fixed for that; the current failure smells like the tombstone-hole delete work (#9020 Map pattern / recent Set-Map hole handling) changed delete-during-iteration bookkeeping — [1,null] (a hole surfaced to the callback), NaN:undefined (a hole key/value pair visited), and a stale size of 1 after deleting both entries. Case 7 additionally skips "r" after deleting an earlier entry.

Found while running the full gap suite for #9066 on 2026-08-29; A/B'd from the merge-base before attributing (both arms identical, so it is not from that PR).