[Noisemaker111]

Verdict: no saving - the occlusion and distance stages rejected nothing at four of eight cameras, the two shots that differed differed with matched draw counts and moving NPCs, and the grid finds almost no occluders because it trusts only placed solids >= 6 m while Harbor Heat'"'"'s downtown is generated chunks and Brightway Park'"'"'s props are smaller.

Part of epic #2312 A2.8. Closed unmerged; the measured outcome is recorded on the epic. LOD cross-fade is #2316.

What is here:

• render/cullPolicy.ts: a tier'"'"'s instance cull distance from GraphicsQuality + the profile drawDistance, per-game scene override wins, ultra resolves to Infinity, hysteresis on the boundary.
• world/occluderGrid.ts: cell grid built once from authored boxes, queried by sampling the camera ray; a sample counts only when strictly inside a blocker footprint, so it can miss a hidden object but cannot see through a blocking one.
• scatter/InstancedScatter.tsx: foliage batches re-test instances against the tier distance when the camera moves and set count to survivors.
• visibility/CullingProvider.tsx: resolves the policy from the terrain-quality tier and exposes it to batched draws.
• visibility/cullingDriver.ts: builds the grid once from placed solids with radius >= 6 m and hands isOccluded real occluders instead of an empty list.

Verification: bun scripts/verify-fast.ts - 12 files, 19 test files, 11 gates, all passed.

Before/after pair, both sides from a checkout of origin/beta (lane-a28-baseline) and this branch, same pinned camera (--look 0,0 --look-from <dist>,<height>,2.4), default graphics tier on both sides (shoot/drive have no tier flag; #2318 tracks it), debug_report read in the same run as the shot:

shot	draws before -> after	triangles before -> after
brightway-park 60 m	555 -> 555	95,697 -> 100,441
brightway-park 180 m	445 -> 445	72,939 -> 72,939
brightway-park 400 m	313 -> 313	44,150 -> 44,150
brightway-park 45 m, 3 m high	616 -> 616	93,070 -> 93,070
harbor-heat 60 m	663 -> 954	1,740,221 -> 1,972,374
harbor-heat 180 m	563 -> 563	1,345,337 -> 1,345,337
harbor-heat 400 m	797 -> 604	922,317 -> 707,851
harbor-heat 45 m, 3 m high	481 -> 481	1,486,742 -> 1,486,742

Four of eight shots are identical to the draw, so the occlusion and distance stages rejected nothing there. The two Harbor Heat shots that differ differ in both directions and share their draw counts with the noise of a live scene (traffic and NPCs move between runs), so they are not a culling effect.

Why nothing is rejected: the grid trusts only placed solids with radius >= 6 m. Harbor Heat'"'"'s downtown is generated city geometry, not placed objects, and Brightway Park'"'"'s props are all smaller, so both scenes give the grid almost no occluders.

This withdraws the earlier claim in this PR that Brightway Park dropped from ~1,330 to 628 draws (-52.8%): that compared one run against a recorded budget baseline taken with the game'"'"'s own camera, not a pair.

Pixel difference against the baseline shot, mean absolute difference / share of pixels differing by more than 8/255: brightway-park 60 m 1.06 / 3.03%, 180 m 0.04 / 0.05%, 400 m 0.05 / 0.06%, 45 m low 0.58 / 1.19%; harbor-heat 60 m 2.66 / 5.29%, 180 m 2.72 / 4.85%, 400 m 1.82 / 0.20%, 45 m low 4.55 / 13.89%. Separate browser runs of a live sim, so moving entities dominate: Harbor Heat'"'"'s 13.9% sits beside identical draw counts.

Missing: the grid'"'"'s per-frame testMs against the draws it saves. occluderStats() exists on the driver but nothing surfaces it to debug_report, so there is no in-game number.

Recommendation: either extend the grid to generated building chunks (the geometry that actually occludes in these games) and re-run this pair, or drop the occlusion half and keep the scatter distance half, which is correct but saved nothing at these distances because tier 400 m already covered the visible scatter.

[Noisemaker111]

No measured saving: unsupported by the pair in this PR and the outcome is recorded on #2312 (comment), with the tier-flag gap in #2318.