[deaneeth]

Summary

Phase 8: the remaining levels of the visualizer.

• L1 softmax cell: for a chosen row, the engine's running max, rescale and sum at each tile completion, next to the float softmax of the same INT8 inputs.
• L2 tile dataflow: live values of the MM1 and MM2 arrays and the tile's S, Z, P, m, α, l and O, kept from the last busy clock.
• L4 naive vs flash: two engine instances on the same inputs. HBM bytes, cycles, tiles and K/V loads are measured; footprint and occupancy are labelled formulas. Error grids against float attention, and an N sweep (4 to 32) of HBM bytes on seeded inputs.
• Engine: new readouts tileDone and getDatapath(). No change to the datapath logic.

Verification

• pnpm test:e2e (Playwright, production build): 12 passed. All four views at 375 px, the light and dark colour schemes, the NAIVE message.
• pnpm test (Vitest): 15 passed.
• node wasm/test.mjs: PASS, including one completion per FLASH tile and a non-decreasing row max within each group.
• SEEDS=1 bash sim/run_e2e.sh: 3/3 passed (the full 200-seed sweep passed on main before this branch).
• bash sim/lint.sh, pnpm lint, pnpm typecheck, pnpm build: clean.

Notes

• The light-mode screenshots have not been reviewed by eye; the tests only check that the page renders with a background.
• L2 shows the last busy clock's wires; between tiles the wires are idle.
• The N sweep uses runToDone rather than a separate runFast.