perf: shared-vertex projection, run-wise colour validation and a bounded stroke walk

[apresmoi]

Three render-path optimisations. Every cell byte-identical, checked against all three fidelity digests.

Bounded stroke walk — applyCellHook 6.19 → 3.35 ms, base-raster −16.8%. The stamper walked 1.1M cells per render, 99.5% of them off-grid: a road that leaves and re-enters the viewport is one segment thousands of cells long. It now iterates only the reachable range. Same samples, same output.

Colour validated once per run, not per cell — spans −45%, atlas −60%. Grids average 17 cells per colour run and the regex ran on every one.

Shared vertices projected once — 259,200 vertex occurrences over 65,341 unique positions on a real terrain tile. Orbit at 283×105 −4%, terrain-only walk −11.3%. Roughly nothing for buildings, whose walls share only vertical edges.

Also: bench/maps-render --walk-look defaulted to 0.15°/frame, which renders synchronously and inflated every walk number that harness produced. Now 0; affected figures re-derived.

Two defects found, reported in docs/design/performance.md, not fixed here — neither is byte-identical:

• The id-map raster is the only polygon loop without the cull runs or the hidden skip, so it claims cells for hidden polygons and blanks layers that paint there.
• OSM row density above 1 is a cliff, 55 → 9.5 fps: all thirteen rows separate into their own passes and the scene is rastered again per render.