Proposal: split global vs shared memory words (@/! and S@/S!)

[tetsuo-cpp]

Problem

Shared memory pointers and global pointers are both pushed onto !forth.stack as raw i64 values. By the time @ / ! are lowered, pointer provenance (address space) is lost.

This causes ambiguous lowering (llvm.inttoptr + generic llvm.load/store) and can lead to incorrect PTX memory access behavior for shared memory kernels (observed as CUDA illegal memory access in tiled matmul test).

Proposal

Introduce explicit shared-memory words and keep existing words global by default:

• @ / ! -> global memory load/store
• S@ / S! -> shared memory load/store

This makes address-space intent explicit at the operation level, even when stack values are untyped runtime i64 addresses.

Why this approach (short term)

• Minimal disruption to current untyped stack design
• Avoids full pointer tagging/type-provenance redesign
• Allows generating address-space-correct lowering for shared memory now

Scope

1. Add new words to parser/translator: S@, S!
2. Add dialect ops for shared load/store (or equivalent explicit markers)
3. Lower S@ / S! via workgroup/shared address-space pointers
4. Keep @ / ! semantics as global memory access
5. Update docs and kernel examples/tests (including tiled matmul) to use S@ / S! for shared buffers

Non-goals

• No pointer tagging in this issue
• No redesign of !forth.stack value representation

Acceptance criteria

• Tiled shared-memory matmul test passes with S@ / S!
• Generated PTX/shared path uses shared-address-space accesses (e.g. ld.shared/st.shared or equivalent address-space-correct form)
• Existing global-memory tests using @ / ! continue to pass

Follow-ups

Longer-term, consider pointer provenance/tagging if we want truly generic @ / ! over mixed pointer kinds at runtime.