Heterogeneous targets · P1-3: longer Scope ladder + lateral shared_with spaces

P1-3 — Longer Scope ladder and lateral spaces

Scope is Device / Sm / Cta / Thread — pre-Hopper. scope: cluster is a parse error today. B200 needs clusters + DSMEM (a CTA addressing a peer CTA's SMEM); the TPU's VMemShared (cross-subcore) has the same lateral shape. within: is strict containment with capacity checks — neither is nested.

Proposed — extend the ladder + add a lateral facet distinct from within:

// B200: thread block cluster + distributed shared memory
Memory DSMEM {
  within: Memory::L2, capacity: 3648 KB, managed: explicit,
  scope: cluster,                    // NEW level, between sm and device
  shared_with: peers(16)             // NEW: lateral, 16 CTAs address each other
}
// TPU: cross-subcore shared VMEM
Memory VMemShared {
  within: Memory::HBM, capacity: 16 MB, managed: explicit,
  scope: subcore,                    // NEW level, below sm
  shared_with: peers(2)
}

Scope becomes Device | Cluster | Sm | Cta | Subcore | Thread, ordered broadest→narrowest so the existing narrowing check keeps working unchanged.

Diagnostic

E60xx: 'DSMEM' is shared with 16 peers but 'SMEM' (scope: sm) is not;
  a peer-shared space cannot be narrower than the scope it is shared across

Where it lands

src/syntax/decl.rs (Scope, shared_with on MemoryDecl), src/parser/decl.rs (scope match arms), src/hir/memory.rs (narrowing + peer coherence).

From the Heterogeneous Target Gap Analysis (§9). Status tracked there and in the P0 decision log. Sibling to the landed P0 work (P0-1 fa056b4e, P0-3 33ce0b64, P0-4 5e10637a, P1-4a-core 34b31902).

Read §9 of docs/discussions/heterogeneous_target_gap_analysis.md and the P0 decision log first. Then inspect Scope and shared_with in src/syntax/decl.rs, scope parsing in src/parser/decl.rs, and narrowing and peer coherence in src/hir/memory.rs. Done means the new scope levels parse and preserve ordering, lateral peer sharing is distinct from within:, and the stated coherence diagnostic is enforced.