Architectural follow-up to the mirror fix landing in the PR that closes #695 and #705.

Context

The mirror fix closes the immediate bugs correctness-wise:

• The conservative GC scan now reaches heap-pointer values via shadow stack → wrapper → WASM-resident bucket array → val_ptr.
• The Python-side _map_store and _set_store remain the source of truth for actual map/set contents.
• The bucket array is a write-only mirror, populated by host_attach_bucket (CLI) and imports.vera.attach_bucket_to_wrapper (browser).

The architectural debt: data lives in two places. Drift between the Python store and the WASM bucket is possible if a future change writes to one but not the other. Code paths multiply. Browser parity required reimplementing the population logic in JavaScript.

Goal: move to bucket-as-truth

Delete _map_store and _set_store. Make the WASM bucket array the sole source of truth. Host imports take wrapper_ptr (not opaque handle) and read/write the bucket directly. The wrapper IS the map / set value.

Three places this needs to land

1. CLI Map host imports (vera/codegen/api.py):

   • host_map_new, host_map_size, _define_map_insert, _define_map_get, _define_map_contains, _define_map_remove, _define_map_keys, _define_map_values — all 8 currently take handle and read _map_store[handle]. Move: take wrapper_ptr, read bucket via wrapper_ptr + 8 (offset to bucket_ptr), use _dict_from_bucket to decode and _build_map_wrapper to return a new wrapper.
   • Delete _map_store, _map_alloc, host_attach_bucket (Map branch).
   • Codegen update (vera/wasm/calls_containers.py): drop _emit_unwrap_handle and _emit_wrap_handle for Map call sites — replace the post-call wrap with a simpler shadow-root sequence that pushes the returned wrapper_ptr onto the shadow stack.

2. CLI Set host imports (vera/codegen/api.py): same pattern as Map. host_set_new, host_set_size, _define_set_add, _define_set_contains, _define_set_remove, _define_set_to_array.

3. Browser runtime (vera/browser/runtime.mjs): JS parallel. Delete mapStore and setStore JS Maps; rewrite all imports.vera.map_* and imports.vera.set_* to use the WASM bucket layout. Equivalent encode/decode helpers in JS.

Bucket layout (preparatory work, not in the mirror PR)

The mirror PR ships with a 12-byte slot layout (key_word_0, key_word_1, val_word) and no bucket header. For the move, the layout needs to grow:

• Slot size: 20 bytes (occupancy flag at +0, key low/high at +4/+8, val low/high at +12/+16). The occupancy flag lets non-string keys distinguish empty vs live without relying on key_word_0 == 0 (which fails for Int 0 keys). Val word pair lets string values store (ptr, len) inline without an extra heap allocation.
• Bucket header: 8 bytes (capacity at +0, count at +4). Lets map_size return in O(1) via header.count instead of scanning slots.

Decimal exempt

Decimal is value-typed (PyDecimal in Python, BigInt in JS) — no heap pointers in the store entry, so the #695 class of bug cannot apply. The bucket_ptr field on Decimal wrappers stays 0 forever; host_attach_bucket short-circuits when kind == 3. Phase D leaves Decimal alone.

Performance challenge (what blocks the move)

A first-pass implementation hit O(N²) in pure-Python decode/encode: per-slot reads through caller["memory"].data_ptr(store)[addr:addr+4] are ~5μs each, so a 10000-element map_insert chain hangs (tens of minutes vs the old _map_store dict-copy at low-microsecond constants).

The fix is to batch the whole bucket region in one wasmtime memory access plus one struct.unpack / struct.pack call instead of per-i32. The encoders for string keys / string values still need per-string _alloc_string calls, but the slot writes can be batched at the end. Mechanical work, but ~1–2 hours of careful instrumentation to confirm the 10000-chain test (TestHostHandleReclamation573::test_map_chain_reclaims_transients) runs in reasonable time.

Acceptance

• _map_store and _set_store deleted.
• JSON parser path (_alloc_map_wrapper for Map<String, Json>) routes through the new bucket builder.
• All existing tests pass, including the 10000-element reclamation chain.
• Browser tests pass.
• No mapStore / setStore in runtime.mjs.
• Architecturally, the WASM bucket array is the single source of truth.

References

• Map<K, T_heap> values stored in _map_store are invisible to conservative GC scan — post-walk reachability of JObject children #695 — original Map<K, T_heap> reachability bug (closed by mirror PR).
• Set<T_heap> values stored in _set_store are invisible to conservative GC scan — sibling to #695 #705 — Set<T_heap> sibling bug (closed by mirror PR).
• Mirror PR — links here after PR is opened.