Map/Set indexed reads return NaN after array pointer hardening

[proggeramlug]

Regression

At exact current main ef6e111a4fb85dcc9fd81f69c30bf57a0bf40a61, indexed reads from live Map/Set receivers routed through array-like runtime fallbacks return NaN. Six pre-existing tests fail individually and deterministically:

• array::collection_tag_tests::a_live_map_receiver_still_reports_its_size_through_the_registry: NaN, expected map key 1.0
• array::collection_tag_tests::a_live_set_receiver_still_reads_its_elements_through_the_registry: NaN, expected set value 5.0
• value::dyn_index::collection_tag_tests::collection_receivers_still_use_their_authoritative_registries: NaN, expected map key 1.0
• typed_feedback::tests::typed_feedback_boxed_fallback_preserves_fractional_keys_for_array_like_receivers: NaN, expected set value 20.0
• typed_feedback::tests::runtime_dynamic_index_fallbacks_preserve_fractional_keys_for_array_like_receivers: NaN, expected set value 20.0
• typed_feedback::tests::polymorphic_index_fallbacks_preserve_fractional_keys_for_array_like_receivers: NaN, expected set value 20.0

For example:

cargo test -p perry-runtime --lib \
  array::collection_tag_tests::a_live_map_receiver_still_reports_its_size_through_the_registry \
  -- --exact

Causal proof

This is an ordering regression introduced by #8041 (971d6ffb6b10c04724cd28d9ee6da2d750c84420), not collection-registry lifetime or GC rooting:

1. fix(gc): install array growth forwarding for low-address arenas #8041 correctly hardened clean_arr_ptr to return null for every tracked object whose GcHeader::obj_type != GC_TYPE_ARRAY.
2. js_array_get_f64 calls clean_arr_ptr(arr) before its existing GC_TYPE_MAP / GC_TYPE_SET branches.
3. A genuine live Map or Set is therefore rejected at the clean funnel and returns NaN; its collection tag and authoritative registry can never be consulted.
4. js_array_length is the positive control: it checks the Map/Set tag plus authoritative registry before clean_arr_ptr, so the same live receivers still report the correct sizes. The array collection tests reach their expected length assertions and fail only on the subsequent indexed read.
5. js_dyn_index_get first confirms the receiver through the authoritative collection registry, then delegates to js_array_get_f64, which reproduces the same early rejection.

This also explains the three typed-feedback failures, since their get fallbacks converge on that dynamic/array accessor. Integer indexed reads are broken; the fractional-key undefined/no-mutation assertions remain the required control.

Required fix

Route authoritative Map/Set receivers before strict array cleaning in js_array_get_f64, matching the proven js_array_length ordering, or use an equivalent dedicated collection accessor.

Keep #8041's strict rejection unchanged for genuine non-array/non-collection receivers. The fix should retain:

• receiver-tag gating before registry probes;
• the authoritative Map/Set registry check (never trust the tag alone);
• no collection registry probes for genuine arrays;
• fractional Map/Set keys returning undefined without mutation;
• all six exact tests passing locally.