Koru Standard Library: Grid — a fixed, positionally-addressed table A grid is the OTHER kind of table, and the distinction is one question: CAN A ROW MOVE? std/store — yes. `take` swap-removes, so position is not identity, so a row is named by a generational HANDLE and every access pays a brand + bounds + generation check. The store earns that machinery by supporting removal. std/grid — no. Every cell exists from declaration and none is ever inserted, removed or relocated, so POSITION IS IDENTITY. A grid pays nothing: no handles, no generation word, no sparse indirection, no freelist, no `len`. A grid is therefore the degenerate, fastest table in the language, and it gets there by forbidding something rather than by optimising. USAGE: import std/grid std/grid:new(cells, size: 4096) { count: 0[i64], sx: 0.0[f32] } std/grid:stored { cells[h].count: cells[h].count + 1 } SIZE, not capacity. `capacity: N` means "up to N, currently `len`". A grid is EXACTLY N cells, all live from the moment it is declared. Different contract, different word, so the two declarations cannot be misread. OUT OF RANGE TRAPS (ruled 2026-08-03). An index outside `0..size` is a panic naming the grid, matching the store's stale-handle floor: fail loud, never silently address a neighbour. Wrapping is the PROGRAM's decision and belongs in the source as an explicit `% size` where a reader can see it — a spatial hash wants exactly that and should say so. The check is a floor, not a tax to be paid forever: an index the compiler can prove in range needs none, and a `! sweep` arm over a grid is precisely that case — the loop constructs the index from `0..size`, so validating it would be validating a fact established one line above. That is the same floor-then-prove-away arc the store's handle check follows. NOT REACTIVE, deliberately, and the seam is left open. Nothing about a grid forbids watches or interceptors — reactivity in Koru is spliced at comptime and costs nothing when nobody listens. But every reactive surface has to answer what it FIRES, and a whole-grid clear firing four thousand times is a question no workload has asked yet. It gets designed when one does.