docs(register): AP-35 drop false equivalence; AP-16 retarget to per-object/cell 8-light cap — A7 Fix D

AP-35: the "numerically equivalent" claim was false. Residual is now two
parts: (a) per-frame GPU evaluate vs retail's bake-once (architecture/perf
difference only; formula matches), and (b) SelectForObject 8-cap means a
surface reached by >8 point lights is dimmer than retail's uncapped bake.
Cross-references AP-16 for the cap ownership.

AP-16: the old "global nearest-8 viewer-distance into UBO" description was
stale — the UBO point-light path is now vestigial (mesh_modern.vert skips
posAndKind.w!=0 entries; point lights come exclusively from the per-object
SSBO binding 5). Retargeted to the current SelectForObject per-object/cell
8-cap mechanism with correct file:line (LightManager.cs:234), both call
sites (ComputeEntityLightSet + GetCellLightSet), and the retail oracle
distinction (hardware cap 0x0054d480 faithful; bake 0x0059cfe0 not).
Preserved the UBO-directional-only note inline rather than losing it.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

docs/architecture/retail-divergence-register.md

@@ -111,7 +111,7 @@ accepted-divergence entries (#96, #49, #50).
 | AP-13 | `ComputeDamage` is a simplified retail damage formula (no augmentations/ratings) — verified DEAD CODE as of 2026-06-04, M2 scaffolding | `src/AcDream.Core/Combat/CombatModel.cs:184` | Not on the critical path; stubbed from r02 §5 + ACE CombatManager for the future M2 predictive display | If wired into the M2 attack-bar estimate as-is, predicted numbers diverge whenever augs/ratings apply | r02 §5; ACE CombatManager |
 | AP-14 | Encumbrance multiplier is a rough piecewise-linear stand-in (1.0→50%, ~0.7@100%, 0.1@300%) for retail's exact curve | `src/AcDream.Core/Items/ItemInstance.cs:187` | Hand-fit segments capture the curve's shape for scaffolding | Client-side burden-scaled effects (speed prediction) differ from retail at most burden ratios when loaded | r06 §6 (retail encumbered multiplier curve) |
 | AP-15 | WeenieError translation table covers only ~30 common codes (from ACE enum docs, not retail string_table.bin); unknown codes render raw hex | `src/AcDream.Core/Chat/WeenieErrorMessages.cs:26` | Untranslated codes are rare, fall back losslessly, 30-second add when reported | Server messages outside the table show as raw hex instead of the retail sentence | retail string_table.bin; ACE WeenieError*.cs |
-| AP-16 | Global nearest-8 viewer-distance light selection (own r13 design); retail bound D3D lights per object/cell. NO viewer-range candidacy filter — each light's range cutoff is applied per-surface in the shader (the earlier `Range²×1.1` slack filter was removed; it dropped torches the viewer stood outside, the #133 "lighting off" report) | `src/AcDream.Core/Lighting/LightManager.cs:10` | Honors retail's 8-hardware-light constraint while fitting a global-uniform shader; nearest-8 is an allocation-free partial-select (no per-frame list/sort) | With >7 nearby lights, different objects are lit than retail would light (retail's per-object pick can light a far object by ITS nearest lights) | r13 §12.2 (acdream design); retail D3D 8-light constraint |
+| AP-16 | Point/spot lights selected per-object / per-cell as the **8 nearest reaching lights** (sphere-overlap, nearest-first) via `LightManager.SelectForObject`, capped at `MaxLightsPerObject=8`; called from `WbDrawDispatcher.ComputeEntityLightSet` (objects) and `EnvCellRenderer.GetCellLightSet` (cell shells). Retail's bake (`SetStaticLightingVertexColors`) sums ALL reaching static lights per vertex with no count cap. Retail's *hardware* path (`minimize_object_lighting` 0x0054d480) DOES cap at 8 per object, so the cap is faithful to retail's hardware path — not to its bake path. The `LightManager.Tick` UBO path survives for DIRECTIONAL (sun) lights only; `mesh_modern.vert`'s UBO loop skips point/spot entries (`posAndKind.w != 0 → continue`) — point lights reach the shader exclusively via the per-object SSBO (binding 5) | `src/AcDream.Core/Lighting/LightManager.cs:234` (`SelectForObject`); `MaxLightsPerObject` ~line 174; call sites `WbDrawDispatcher.ComputeEntityLightSet` + `EnvCellRenderer.GetCellLightSet` | Matches retail's hardware constraint (8 lights per object/cell); selection is nearest-sphere-overlap which faithfully allocates lights to the surfaces that actually see them | Surfaces reached by >8 point lights are dimmer than retail's uncapped bake — rare (a dungeon room has a handful of torches), but real; see AP-35 for the bake-vs-GPU-evaluate architecture difference | `minimize_object_lighting` 0x0054d480 (retail's 8-light hardware cap); `SetStaticLightingVertexColors` 0x0059cfe0 (retail's bake, no count cap) |
 | AP-17 | Spell metadata from third-party CSV (3,956 rows, bad rows silently skipped), not the portal.dat SpellTable; Family feeds stacking decisions | `src/AcDream.Core/Spells/SpellTable.cs:10` | The dat spell-table port (obfuscated/encrypted aspects) wasn't done; CSV closed #11 fast and unblocked #6 stacking | Any CSV↔dat drift (wrong Family, missing rows) silently produces wrong buff-stacking winners and wrong panel info | portal.dat SpellTable 0x0E00000E |
 | AP-18 | Radar/indicator RGBA hand-tuned from screenshots; dispatch order ports `GetBlipColor` exactly but the real `RGBAColor_Radar*` static data is unrecovered | `src/AcDream.Core/Ui/RadarBlipColors.cs:33` | Color constants live in retail static data not yet extracted; comment invites tightening when recovered | Blip/indicator hues differ subtly from retail color cues | `gmRadarUI::GetBlipColor` 0x004d76f0; RGBAColor_Radar* (unrecovered) |
 | AP-19 | `PortalSideEpsilon` 0.01 (≈1 cm) instead of retail F_EPSILON ≈ 0.0002 — a documented render-root-lag tolerance, NOT a retail constant. DO-NOT-RETRY: T2 (BR-4) tried the retail value; CornerFloodReplay refuted it | `src/AcDream.App/Rendering/PortalVisibilityBuilder.cs:49` | Retail's tight epsilon only works with eye-exact swept curr_cell tracking; our viewer cell lags the eye by up to ~1 cm at pressed corners. Tighten after the #108-membership family + cdstW near-clip pin land | A 1 cm misclassification band at portal planes can flood or cull a portal the eye hasn't crossed — one-frame leaks / grey flashes at knife-edge doorway/corner positions | F_EPSILON @0x007c8c70; `PView::InitCell` 0x005a4b70 |
@@ -131,7 +131,7 @@ accepted-divergence entries (#96, #49, #50).
 | AP-33 | Interior-root look-in cells (**#124** sub-pass) draw their statics + DYNAMICS + emitters WHOLE — no per-part/per-object viewcone check; retail viewconeCheck's each vs the installed view (the **#131** portal closure: a server object in a look-in cell drew nowhere — dynamics-last culls cells absent from the main cone, and post-seal it z-fails anyway) | `src/AcDream.App/Rendering/RetailPViewRenderer.cs` (`DrawBuildingLookIns`) | The main viewcone has no entries for look-in cells; over-include is the safe direction (z-correct, repainted outside apertures by the root's shells); look-in cell counts are small (~1-3 cells) | A few wasted draws on content outside the doorway region (repainted); no under-draw direction remains | `viewconeCheck` 0x0054c250; nested `DrawCells` objects pc:432878 |
 | AP-34 | Landscape-stage alpha deferral is a TWO-PHASE slice split (statics-early / dynamics+particles+weather-late around the **#124** look-ins) + outdoor-root attached scene emitters moved to the post-frame pass, not retail's single deferred alpha flush. Residual: building exteriors' / outside-stage dynamics' own translucent MESH batches still draw within their stage draw call (before later stage content) | `src/AcDream.App/Rendering/RetailPViewRenderer.cs` (`DrawLandscapeThroughOutsideView` late loop) + `GameWindow` post-frame Scene pass | The MDI dispatcher draws translucency inside each Draw call; a faithful FlushAlphaList port needs a global deferred alpha list across all landscape draws — the split covers the user-visible cases (#131 portal swirl, #132 candle flame indoors + outdoors) | Translucent landscape content drawn early and screen-overlapped by content drawn later in the stage gets overpainted (no depth self-protection) — the portal-swirl/candle-flame class re-appears in the residual configurations | `D3DPolyRender::FlushAlphaList` (DrawCells pc:432722) |
 | AP-36 | Dungeon streaming gate triggers on the player's CURRENT cell being a sealed EnvCell (`CurrCell.IsEnv && !SeenOutside`), an approximation of ACE's full landblock `IsDungeon` (all-heights-zero + NumCells>0 + Buildings.Count==0). The retail BEHAVIOR (a dungeon loads no adjacent landblocks) is faithful — only the runtime TRIGGER is the cheap cell predicate instead of classifying the center landblock. **#135 pre-collapse:** at login/teleport the same collapse is triggered EARLY (the instant the streaming center is recentered onto the spawn/dest cell) via `IsSealedDungeonCell` reading the EnvCell **dat** `SeenOutside` flag — because the physics `CurrCell` is null until placement, which waits for hydration; without the early trigger the full 25×25 ocean-grid window loads then unloads (the ~30 s login FPS ramp) | `src/AcDream.App/Rendering/GameWindow.cs:6895` (per-frame predicate) + `:IsSealedDungeonCell` + `:OnLiveEntitySpawnedLocked`/`:OnLivePositionUpdated` (login/teleport pre-collapse hooks) + `src/AcDream.App/Streaming/StreamingController.cs` (collapse/expand/`PreCollapseToDungeon`) | The predicate is already computed for sun/sky gating (playerInsideCell) and exactly matches for sealed dungeons vs windowed building interiors (SeenOutside=true → not gated); no landblock re-classification needed. The dat-flag read is the same `EnvCellFlags.SeenOutside` the hydrated `ObjCell.SeenOutside` is built from (`EnvCell.cs:72`/`PhysicsDataCache.cs:224`), so the pre-collapse decision matches the eventual per-frame gate exactly | A dungeon cell that reports SeenOutside (an entrance cell open to the surface) briefly un-collapses and re-streams the window; a hypothetical windowless building back-room (IsEnv && !SeenOutside but HasBuildings) would wrongly collapse its outdoor neighbors; a sealed-dungeon entrance cell that is itself SeenOutside is simply MISSED by the early trigger and falls back to the existing late collapse (no worse than before #135) | ACE `LandblockManager.GetAdjacentIDs` (dungeons→empty) Landblock.cs:577-582; `IsDungeon` Landblock.cs:1264-1277 |
-| AP-35 | Point/spot lights are now PER-VERTEX Gouraud (`pointContribution` ~line 153 of `mesh_modern.vert`) matching retail's `SetStaticLightingVertexColors` bake path. Half-Lambert wrap (`(1/1.5)·(N·D + 0.5·d)`) AND norm distance attenuation (`distsq>1 ? distsq·d : d`) ARE ported (A7 Fix A, `aa94ced`). Point-light sum clamped to [0,1] on its own accumulator before adding ambient+sun (A7 Fix D D-1, mirrors retail's per-vertex bake clamp). CPU oracle: `src/AcDream.Core/Lighting/LightBake.cs`, locked by `tests/AcDream.Core.Tests/Lighting/LightBakeConformanceTests.cs`. **Residual:** acdream lights in-shader each frame (per-frame GPU evaluate); retail bakes into the vertex buffer ONCE. This is an architecture/performance difference, not a visual one — the per-vertex contribution is numerically equivalent | `src/AcDream.App/Rendering/Shaders/mesh_modern.vert` (`pointContribution` ~line 153; wrap ~line 163; norm ~line 167; point-sum clamp line 210) | Per-vertex Gouraud + wrap + norm + clamp all match retail. Per-frame evaluate vs. bake-once is the only remaining deviation; the pipeline computes the same vertex colours retail bakes. `LightInfoLoader.cs:81` folds static_light_factor 1.3 into Range | A new frame-time consumer that bypasses `accumulateLights` (e.g. a future prepass) would need to replicate the wrap + norm formula; per-frame GPU re-evaluate has higher per-frame cost than retail's one-shot bake for static geometry | `calc_point_light` 0x0059c8b0 (line 0x0059c9a2 ramp; 0x0059c925 wrap); `SetStaticLightingVertexColors` 0x0059cfe0; static_light_factor 0x00820e24 |
+| AP-35 | Point/spot lights are now PER-VERTEX Gouraud (`pointContribution` ~line 153 of `mesh_modern.vert`) matching retail's `SetStaticLightingVertexColors` bake path. Half-Lambert wrap (`(1/1.5)·(N·D + 0.5·d)`) AND norm distance attenuation (`distsq>1 ? distsq·d : d`) ARE ported (A7 Fix A, `aa94ced`). Point-light sum clamped to [0,1] on its own accumulator before adding ambient+sun (A7 Fix D D-1, mirrors retail's per-vertex bake clamp). CPU oracle: `src/AcDream.Core/Lighting/LightBake.cs`, locked by `tests/AcDream.Core.Tests/Lighting/LightBakeConformanceTests.cs`. **Residual (two parts):** (a) acdream lights in-shader each frame (per-frame GPU evaluate); retail bakes into the vertex buffer ONCE — an architecture/performance difference; the wrap + norm + clamp formula is the same, but bake-once is cheaper for static geometry; (b) acdream's `SelectForObject` keeps only the 8 NEAREST reaching point/spot lights per object/cell (`MaxLightsPerObject=8`, see AP-16), whereas retail's bake sums ALL reaching static lights per vertex — a surface reached by >8 point lights is dimmer in acdream than retail's bake result (rare in practice; a room has a handful of torches) | `src/AcDream.App/Rendering/Shaders/mesh_modern.vert` (`pointContribution` ~line 153; wrap ~line 163; norm ~line 167; point-sum clamp line 210) | Per-vertex Gouraud + wrap + norm + clamp all match retail. The two residuals are: (a) per-frame GPU vs bake-once — architecture/perf only; (b) 8-light cap dimming when >8 lights reach one surface — rare. `LightInfoLoader.cs:81` folds static_light_factor 1.3 into Range | (a) A new frame-time consumer bypassing `accumulateLights` would need to replicate the wrap + norm formula; per-frame GPU re-evaluate has higher per-frame cost than bake for static geometry. (b) A densely lit scene (>8 torches reaching one wall) renders dimmer than retail — see AP-16 for the 8-cap ownership | `calc_point_light` 0x0059c8b0 (line 0x0059c9a2 ramp; 0x0059c925 wrap); `SetStaticLightingVertexColors` 0x0059cfe0; static_light_factor 0x00820e24 |
 
 ---