Memory budgets for derived state (V15) #321

Part of epic #305.

Problem. Per-request memory is metered, but no budget covers standing-query result sets, plan caches, HNSW graphs or arrangements; max_graph_memory_bytes bounds stored facts only (src/storage_engine/program_commit.rs:84-87; src/execution/memory.rs; /metrics memory = tuples x 64, rest/handlers/admin.rs:124-125).

Target. Per-KG accounting of arrangement bytes (trace batch sizes) and view rows, charged against max_graph_memory_bytes extended to derived state (or a new max_view_memory_bytes); +rule refused at registration when the initial build would exceed the budget ("view too large", with the measured size); a KG whose views grow past the budget enters views_over_budget (reads keep serving at the last frontier, writes refuse with resource_exhausted, same contract as #275); metrics per KG/view.

Acceptance. Tests: registration refusal, over-budget state and recovery after deletes; metric values match map_batches sums; the lab pack at 1M edges reports about the micro-benchmark's 344 MB.

Size M. Dependencies V2. Order slot 6.