Context

Phase 1 of #573 migrated Map<K, V> to the heap-wrap-as-ADT scheme: every Map value is now a pointer to a GC-managed wrapper ADT, and Phase 2c of $gc_collect fires host_decref_handle(MAP, handle) to evict unreachable entries from _map_store.

This issue tracks the parallel migration for Set<T>, which still uses raw i32 handles into _set_store and therefore still leaks per execute().

Scope

Mechanical mirror of the Map migration:

1. vera/wasm/calls_containers.py — wrap on Set-returning ops (set_new, set_add, set_remove); unwrap on Set-consuming ops (set_contains, set_size, set_to_array).
2. vera/wasm/helpers.py — remove "Set" from _HOST_HANDLE_TYPES (wrappers are real heap pointers and need rooting).
3. vera/codegen/api.py::host_decref_handle — add kind == 2 branch popping _set_store.
4. vera/browser/runtime.mjs::host_decref_handle — same kind === 2 branch popping setStore.
5. Define _WRAP_KIND_SET = 2 and _SET_HANDLE_TAG = 0xFEEDC002 constants alongside the Map equivalents (already reserved in the phase-1 PR).
6. Regression tests mirroring TestHostHandleReclamation573::test_map_chain_reclaims_transients for Set.
7. Update the post-Active reclamation of host-store handles (deferred from #346) #573 expectation in TestOpaqueHandleParamRooting347::test_set_param_not_shadow_pushed (Set params will be shadow-pushed post-migration).

No call-site changes needed for JSON/HTML — they don't use Set internally (only Map).

Acceptance

• 5 000-element array_fold over set_add chain: _set_store size at exit < 100 (mirroring the Map test's bound).
• All existing Set tests still pass.
• Browser parity tests pass.
• Set<T> is removed from _HOST_HANDLE_TYPES in vera/wasm/helpers.py.

Out of scope

• Decimal migration (separate follow-up).