regression(main): wolf-ecs-shaped nested subclass loops lost their versioned fast clones (packed-guard sites 3 → 0) #9106

Symptom

On main f3f405271c (and 744df26d8e), the issue_8690_loop_versioned_arraylike case read_only_loops_have_preheader_proofs_and_fallback_free_fast_blocks fails its shape assertion: scan()'s IR contains 0 js_packed_arraylike_loop_guard( sites where 3 are expected — the outer fast/slow copies no longer own preheader-versioned inner loops. issue_8773_closure_capture_packed_loops fails 1 case the same way. Program output stays correct; the flagship nested-subclass-loop shape silently runs at generic-loop speed.

Window

Both suites were green on macOS at 8b40634abd (the #9071 replay) and fail at f3f405271c. The window contains the packed/versioned-loop landings #9070 (versioned length-bound hoist), #9084 (read admissions), #9091 (unboxed accumulators), plus #9075/#9080/#9085. The shape gates were doing their job — this was caught only because an unrelated branch ran the full battery.

Repro

cargo test -p perry --test issue_8690_loop_versioned_arraylike

(macOS; needs the default-target libperry_runtime.a — on a fresh box cargo build -p perry-runtime-static first, and note the suite fails in ~0.5s with an archive-mismatch harness error otherwise, which is NOT this regression.)

Notes

The fixture is the wolf-ecs Query extends Array / Archetype extends Array nested scan — the exact shape the ECS campaign optimized to 2.3×/1.5× of node, so this likely walks back part of that standing until fixed. Independent of PR #9105 (verified by running both suites on detached main in the same worktree).

Reported by the ECS/ops perf campaign, session https://claude.ai/code/session_019WVcWKmYsUBnnFB7nBgbBJ; the loop lane's owner (85-successor / whoever holds #9084/#9091) is the natural assignee — coordination on the cc-socks channel.