docs(lighting): A7 Fix D investigation RESOLVED + implementation spec (#140)

Resolve the Fix D contradiction with decomp (workflow wf_f660eb88 + adversarial verify) + 4 live cdb captures. The D3D-FF model was the WRONG oracle: retail has TWO light systems — STATIC torches BAKE into wall vertices (calc_point_light, triple-clamped: range gate + per-channel min(scale*color,color) + per-vertex [0,1] from black), DYNAMIC lights go D3D hardware. The captured intensity=100 is the purple PORTAL (magenta, dynamic), not a wall torch. Ground truth: 38 static warm torches (orange (1,0.588,0.314)/cream, intensity=100, falloff 3-5) + 2 dynamic. acdream over-brightness = two confirmed bugs: D-1 mesh_modern.vert folds ambient+sun+torches into one UNCLAMPED accumulator (single frag clamp) -> warm blowout; D-2 EnvCellRenderer never binds SSBO 4/5 so the cell shell reads a leaked light set. Spec: D-1 in-shader clamp-split (clamp the torch sum on its own before ambient/sun); D-2 bind the shell's own per-cell light set (mirror WbDrawDispatcher); LightBake.cs is the C# conformance oracle. Adds the 4 reusable cdb capture scripts. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
2026-06-18 17:08:27 +02:00 · 2026-06-18 17:08:27 +02:00 · c407104ab9
commit c407104ab9
parent f384d036a3
7 changed files with 419 additions and 46 deletions
--- a/docs/research/2026-06-18-lighting-a7-fixABC-shipped-fixD-handoff.md
+++ b/docs/research/2026-06-18-lighting-a7-fixABC-shipped-fixD-handoff.md
@ -45,57 +45,83 @@ was wrong (Fix C). Don't re-investigate the purple.

 ---

-## OPEN — Fix D: outdoor OBJECTS too bright near torches
+## RESOLVED — Fix D: outdoor walls too bright near torches (contradiction settled 2026-06-18)

-**Symptom (user, 2026-06-18):** the Holtburg meeting-hall walls blow out warm/bright
-in acdream vs dim in retail. Fix A/B/C did NOT touch this. It's the per-object
-point-light **contribution on objects**.
+**Symptom (user):** Holtburg meeting-hall walls blow out **warm**/bright in acdream
+vs dim in retail. The contradiction ("D3D-FF math says color×100 should blow WHITE,
+yet retail is DIM") is **resolved**: the D3D-FF model was the WRONG ORACLE for these
+walls. Settled by a 5-thread decomp workflow (`wf_f660eb88`) + adversarial verify +
+4 live cdb captures. **⚠ The "DO NOT port the D3D-FF model" warning still stands** —
+not because it'd be too bright, but because it's the wrong path entirely.

-### Grounded (cdb + decomp) — retail's object point-light path
-`Render::config_hardware_light` (0x0059ad30) builds the `D3DLIGHT9`:
- `Diffuse = color × intensity`
- `Attenuation = (0, 1, 0)` ⇒ **1/d** (inverse-LINEAR; acdream's `calc_point_light`
-  is `~1/d²` via norm = distsq·d)
- `Range = falloff × rangeAdjust`, **`rangeAdjust = 1.5`** (0x00820cc4) ⇒ torch Range
-  = 6×1.5 = **9 m** (LARGER than acdream's falloff×1.3 = 7.8 m — range is NOT why
-  we're brighter)
- live `LIGHTINFO` captured: torch `type=0 intensity=100 falloff=6`; a 2nd light
-  `intensity=2.25 falloff=10`
- `d3d_material.Diffuse = (1,1,1)` white (decomp 0x00539774)
+### Render path (Ghidra xrefs — unambiguous, two SEPARATE light systems)
+- **STATIC lights → CPU vertex BAKE.** `RenderDeviceD3D::DrawEnvCell` (0x0059F170) →
+  `D3DPolyRender::SetStaticLightingVertexColors` (0x0059CFE0) → `calc_point_light`
+  (0x0059C8B0, its SOLE caller). Wall torches are STATIC objects → baked into vertex
+  colours. AC town buildings are EnvCell structures, so their walls take this path.
+- **DYNAMIC lights → D3D hardware FF.** `add_dynamic_light` → `insert_light` (0x0054D1B0)
+  → `config_hardware_light` (0x0059AD30); `minimize_envcell_lighting` (0x0054C170)
+  enables ONLY the dynamic subset (class 2) for the cell — statics are NEVER hardware-
+  enabled for the cell. (`minimize_object_lighting` 0x0054D480 enables both, for free
+  GfxObjs.) So `config_hardware_light` — where last session's `intensity=100` was seen —
+  carries DYNAMIC lights for cells, not the wall torches.

-### THE CONTRADICTION (resolve this FIRST next session)
-By `mat(1)×color×100×(N·L)×(1/d)`, a torch 3 m away = `color×33` ⇒ retail's walls
-SHOULD blow to **WHITE** — but they're **DIM**. Material diffuse, range, and
-intensity are all captured and ruled out. So the scaling lives in the building's
-**RENDER PATH**, which is unknown. **⚠ DO NOT port the D3D-FF model — by this math it
-would make objects BRIGHTER (white), the opposite of the fix.**
+### Why retail stays warm-but-DIM (the bake is triple-clamped — `calc_point_light`)
+Per light: `range = falloff×1.3` hard gate; half-Lambert wrap `(1/1.5)(N·D + 0.5·d)`;
+`norm = (distsq>1)? distsq·d : d` (~1/d²); `scale = (1−d/range)·intensity·(wrap/norm)`;
+then the **decisive per-channel cap `result = min(scale·color, color)`** — one light adds
+**at most its own (sub-1.0, warm) colour**, however large intensity is. Caller sums from
+**BLACK** (no ambient/sun in the accumulator) over all static lights, then **clamps the
+sum to [0,1]** per channel before packing `vertex.diffuse`. White needs many in-range
+lights stacking past 1.0; a hall has a handful, each warm-capped.

-### The decisive next capture
-Determine the static building's ACTUAL render path:
- **Hypothesis (a) — MOST LIKELY:** static buildings DON'T use D3D hardware lighting.
-  They use the `D3DPolyRender::SetStaticLightingVertexColors` BAKE (0x0059cfe0 →
-  `calc_point_light`), like EnvCells. The `config_hardware_light` lights I captured
-  were for a DIFFERENT object (player / creature / the purple PORTAL — note the
-  `intensity=100` could be the portal, not the wall torch). If (a) holds, acdream's
-  `calc_point_light` is the RIGHT model and the over-brightness is the **per-channel
-  cap** (`min(scale×col,col)` lets several torches each reach full colour and sum to
-  white) and/or **too many torches selected** per object and/or a missing clamp step.
- **Hypothesis (b):** `D3DRS_LIGHTING` off / lights not `LightEnable`'d for the
-  building draw.
- **How to capture:** break at `SetStaticLightingVertexColors` (0x0059cfe0) and see
-  whether it's called for the building's mesh (confirms the bake path); and/or
-  inspect the render state around the static-object `DrawIndexedPrimitive`
-  (`D3DRS_LIGHTING`, which lights are enabled). Also: at `config_hardware_light`,
-  dump WHICH object/owner the light is being configured for to identify whether the
-  `intensity=100` light is the torch or the portal.
+### Live cdb ground truth (4 captures; scripts in `tools/cdb/a7-fixd-*.cdb`)
+`Render::world_lights` @ **0x008672a0** (LightParms): `num_static_lights` @ +0x104,
+`sorted_static_lights[]` (RenderLight*, info @ RL+0x70) @ +0x3498, `num_dynamic_lights`
+@ +0x3588. Captured standing in Holtburg:
+- **num_static_lights = 38**, **num_dynamic_lights = 2.**
+- **2 DYNAMIC** (`add_dynamic_light`, d3dIdx 1–2): viewer light `intensity=2.25 falloff=10
+  color=(1,1,1)` white; **PORTAL** `intensity=100 falloff=6 color=(0.784,0,0.784)` MAGENTA.
+  → **the `intensity=100` light is the purple PORTAL (dynamic/hardware), NOT a wall torch.**
+- **38 STATIC** wall torches, all `type=0 intensity=100`, **WARM**: orange
+  `(1.0, 0.588, 0.314)` falloff 4, and cream `(0.980, 0.843, 0.612)` falloff 3–5
+  (→ bake range ~3.9–6.5 m). Torches DO carry intensity=100, but the per-channel cap
+  pins each to its warm colour ⇒ retail walls go warm, not white.

-### acdream side — where the fix lands
- acdream runs `calc_point_light` (wrap/norm + per-channel cap) for ALL meshes via
-  `mesh_modern.vert` `pointContribution` (objects AND cells — Fix A).
- If buildings use the bake, the likely fix is in the **cap / sum / count**, not the
-  attenuation model. Files: `src/AcDream.App/Rendering/Shaders/mesh_modern.vert`
-  (`pointContribution` + `accumulateLights`), `src/AcDream.Core/Lighting/LightManager.cs`
-  (`SelectForObject`), `LightBake.cs` (verbatim calc_point_light, still unwired).
+### acdream's actual bug — TWO real causes (both verified in source)
+- **D-1 (math, primary): unclamped accumulator folding ambient+sun+torches.**
+  `mesh_modern.vert` `accumulateLights` starts `lit = uCellAmbient.xyz` (:184), ADDS
+  sun (:196), ADDS each capped torch (:206), returns UNCLAMPED (:208); the ONLY clamp is
+  one `min(lit,1.0)` in `mesh_modern.frag:92` AFTER a lightning bump (:89). The per-light
+  cap (:180) IS faithful. But pouring ambient + sun + up-to-8 intensity-100 WARM torches
+  into ONE bucket and trimming only at the end overflows to warm-white. Retail clamps the
+  torch sum on its OWN (from black); ambient/sun are a separate term.
+- **D-2 (state, compounding): EnvCell shell SSBO binding leak.**
+  `EnvCellRenderer.cs:1225-1230` (RenderModernMDIInternal) binds SSBO 0/1/2/3 only, NEVER
+  4 (`gLights`) or 5 (`instanceLightIdx`) — which the shared `mesh_modern.vert` reads at
+  :204-206. Only `WbDrawDispatcher` binds 4/5. Indoor DrawInside interleaves the two, so a
+  cell shell reads whatever LEAKED light set the last WbDrawDispatcher draw left bound —
+  a different entity's torches, wrong per-instance indices ⇒ wrong/over-bright walls.
+- `LightBake.cs` (verbatim CPU port) exists but is UNWIRED (zero callers); the live path is
+  the in-shader version missing the clamp shape.
+
+### Fix plan (REPORT-ONLY — implement in a separate session, with the no-workaround rule)
+- **D-1:** accumulate point/spot into a LOCAL `pointAcc`, `saturate` it to [0,1] BEFORE
+  adding ambient + sun — mirrors `SetStaticLightingVertexColors` (sum-from-black, clamp the
+  point sum). Keep the per-light `min(scale·baseCol, baseCol)` (vert:180). Files:
+  `mesh_modern.vert` (split accumulator + clamp), `mesh_modern.frag` (reorder/drop the
+  single clamp). Conformance golden: a wall ≤~5 m from an orange `(1,0.588,0.314)` torch
+  bakes warm-but-≤[0,1], NOT white.
+- **D-2:** EnvCell shell must bind binding 4 (global lights) + 5 (per-instance light set)
+  for ITS OWN instances — compute a per-shell set like `WbDrawDispatcher.ComputeEntityLightSet`
+  (LightManager.SelectForObject); option (b) all-(-1) fallback = NO point lights is a STOPGAP
+  (needs approval + a divergence-register row). File: `EnvCellRenderer.cs` RenderModernMDIInternal.
+- **Stale doc to fix in the D-1 commit:** divergence-register `AP-35` still describes the
+  point-light path as per-pixel `mesh_modern.frag:52` with the wrap "NOT ported"; Fix A
+  (`aa94ced`) moved it to per-vertex `mesh_modern.vert:163` WITH the wrap.
+- **Do NOT port the D3D-FF hardware model for the walls** (config_hardware_light's
+  color×intensity / (0,1,0)=1/d / Range=falloff×1.5) — it lights GfxObjs/dynamics, not the
+  baked walls.

 ---

--- a/docs/superpowers/specs/2026-06-18-a7-fixd-torch-overbright-design.md
+++ b/docs/superpowers/specs/2026-06-18-a7-fixd-torch-overbright-design.md
@ -0,0 +1,211 @@
+# A7 Fix D — warm torch over-brightness on indoor walls (#140)
+
+**Date:** 2026-06-18  **Milestone:** M1.5 (Indoor world feels right) → A7 lighting
+**Status:** design approved (user pre-approved 2026-06-18); ready for implementation plan.
+**Investigation source of truth:**
+[`docs/research/2026-06-18-lighting-a7-fixABC-shipped-fixD-handoff.md`](../../research/2026-06-18-lighting-a7-fixABC-shipped-fixD-handoff.md)
+(RESOLVED section) + `claude-memory/reference_retail_ambient_values.md`.
+
+## Problem
+
+The Holtburg meeting-hall walls (and outdoor objects near torches) blow out
+**warm/bright** in acdream vs **dim** in retail. Fix A/B/C (shipped) did not touch this.
+
+The handoff "contradiction" (D3D-FF math `color×100×N·L/d` says walls should go WHITE,
+yet retail is DIM) is **resolved**: the D3D-FF hardware model is the **wrong oracle**
+for these walls. Two SEPARATE retail light systems (Ghidra xrefs, unambiguous):
+
+- **STATIC lights → CPU vertex BAKE**: `DrawEnvCell` (0x0059F170) →
+  `SetStaticLightingVertexColors` (0x0059CFE0) → `calc_point_light` (0x0059C8B0, its
+  SOLE caller). Wall torches are STATIC objects → baked into vertex colours.
+- **DYNAMIC lights → D3D hardware FF**: `add_dynamic_light` → `config_hardware_light`
+  (0x0059AD30); `minimize_envcell_lighting` (0x0054C170) enables ONLY the dynamic
+  subset for a cell. The previously-captured `intensity=100` light is on THIS path.
+
+`calc_point_light` is mathematically **bounded**: range gate `d < falloff×1.3`; the
+decisive **per-channel cap `min(scale·color, color)`** (a torch adds at most its own
+sub-1.0 colour, any intensity); caller sums from **BLACK** then clamps the sum to
+`[0,1]` (no ambient/sun in the bake accumulator). White needs many in-range lights;
+a hall has a handful, each warm-capped.
+
+### Ground truth (live cdb, `tools/cdb/a7-fixd-*.cdb`; `Render::world_lights` @ 0x008672a0)
+
+Holtburg: **38 static + 2 dynamic** lights.
+
+| Light | path | type | intensity | falloff | colour (r,g,b) |
+|---|---|---|---|---|---|
+| viewer light | dynamic / HW | point | 2.25 | 10 | (1, 1, 1) white |
+| **portal** | dynamic / HW | point | **100** | 6 | **(0.784, 0, 0.784) magenta** ← the captured "intensity=100"; NOT a wall torch |
+| 38× wall torch | static / **bake** | point | 100 | 3–5 | **(1.0, 0.588, 0.314) orange** / (0.980, 0.843, 0.612) cream |
+
+Torches carry `intensity=100` too, but the per-channel cap pins each to its warm
+colour ⇒ retail walls go warm, never white.
+
+## Root cause in acdream (both verified in source)
+
+Two independent bugs, both touching the meeting-hall walls; this spec fixes both.
+
+**D-1 (math, primary): unclamped accumulator folding ambient + sun + torches.**
+[`mesh_modern.vert`](../../../src/AcDream.App/Rendering/Shaders/mesh_modern.vert)
+`accumulateLights` starts `lit = uCellAmbient.xyz` (:184), adds sun (:196), adds each
+capped torch (:206), returns UNCLAMPED (:208); the only clamp is `min(lit,1.0)` in
+`mesh_modern.frag:92` after a lightning bump. The per-light cap (vert:180) is faithful.
+But pouring ambient + sun + up-to-8 intensity-100 WARM torches into ONE bucket and
+trimming only at the end overflows to warm-white. Retail clamps the torch sum on its
+OWN (from black); ambient/sun are a separate material-lit term.
+
+**D-2 (state, compounding): EnvCell shell SSBO binding leak.**
+[`EnvCellRenderer.RenderModernMDIInternal`](../../../src/AcDream.App/Rendering/Wb/EnvCellRenderer.cs)
+binds SSBO 0/1/2/3 only, NEVER **4** (`gLights`) or **5** (`instanceLightIdx`) — which
+the shared `mesh_modern.vert` reads unconditionally (:204-206). Only `WbDrawDispatcher`
+binds 4/5. Indoor `DrawInside` interleaves the two, so a cell shell reads whatever
+LEAKED light set the last `WbDrawDispatcher` draw left bound (a different entity's
+torches, wrong per-instance indices) ⇒ wrong/over-bright walls.
+
+`LightBake.cs` (verbatim CPU port of the bake) exists but is UNWIRED (zero callers).
+
+## Design
+
+Decisions (user, 2026-06-18): **D-1 = small in-shader clamp split** (not a CPU bake);
+**D-1 + D-2 land together**, single visual verification.
+
+### D-1 — clamp the torch sum on its own (mirrors `SetStaticLightingVertexColors`)
+
+In `mesh_modern.vert` `accumulateLights`, give point/spot lights their own accumulator,
+saturate it to `[0,1]` BEFORE it joins ambient + sun. The per-light cap and
+`pointContribution` are unchanged; the only new operation is one `min(pointAcc, 1.0)`.
+
+```glsl
+vec3 accumulateLights(vec3 N, vec3 worldPos, int instanceIndex) {
+    // ambient + sun = retail's material-lit term
+    vec3 lit = uCellAmbient.xyz;
+    int activeLights = int(uCellAmbient.w);
+    for (int i = 0; i < 8; ++i) {
+        if (i >= activeLights) break;
+        if (int(uLights[i].posAndKind.w) != 0) continue;   // directional only
+        vec3 Ldir = -uLights[i].dirAndRange.xyz;
+        float ndl = max(0.0, dot(N, Ldir));
+        lit += uLights[i].colorAndIntensity.xyz * uLights[i].colorAndIntensity.w * ndl;
+    }
+    // point/spot torches: their OWN accumulator, clamped to [0,1] (retail baked emissive)
+    vec3 pointAcc = vec3(0.0);
+    int base = instanceIndex * 8;
+    for (int k = 0; k < 8; ++k) {
+        int gi = instanceLightIdx[base + k];
+        if (gi < 0) continue;
+        pointAcc += pointContribution(N, worldPos, gLights[gi]);   // per-light cap unchanged
+    }
+    lit += min(pointAcc, vec3(1.0));   // <-- THE FIX
+    return lit;                        // frag still does final min(lit, 1.0)
+}
+```
+
+Behaviour change is confined to surfaces whose torch sum currently exceeds 1.0 —
+normally-lit surfaces are byte-identical (no regression). Shared by every mesh using
+this shader (outdoor objects AND cell walls), matching the issue's scope.
+`mesh_modern.frag:92`'s final `min(lit, 1.0)` stays as-is (it clamps the total to the
+retail FF pixel clamp). The lightning bump (frag:89) is unaffected.
+
+### D-2 — the EnvCell shell binds its OWN light set
+
+`EnvCellRenderer` must own its lighting like `WbDrawDispatcher` does, instead of reading
+leaked SSBO state. Mirror `WbDrawDispatcher`'s proven pattern
+(`ComputeEntityLightSet`/`AppendCurrentLightSet`/`UploadGlobalLights`):
+
+1. **Wire `LightManager` in** via `Initialize(...)` (alongside `_shader`). Self-contained
+   pass — per `feedback_render_self_contained_gl_state`, EnvCellRenderer already
+   re-uploads its own `uViewProjection`; it now also uploads/binds its own lights.
+2. **Binding 4 (global lights):** upload `LightManager.PointSnapshot` itself, packed
+   identically to `WbDrawDispatcher.UploadGlobalLights` (the `GlobalLight` SSBO layout:
+   `posAndKind`, `dirAndRange`, `colorAndIntensity`, `coneAngleEtc`). Same snapshot →
+   same indices both renderers reference. `BuildPointLightSnapshot` is already called
+   once per frame before rendering. **Extract the packing into a shared helper** so the
+   two renderers cannot drift (a `GlobalLightPacker` in `AcDream.App/Rendering/Wb/` or a
+   static on the snapshot type) — do not copy-paste the struct layout.
+3. **Binding 5 (per-instance light set):** per **cell** (keyed on `allInstances[i].CellId`),
+   compute the set ONCE with `LightManager.SelectForObject(snapshot, cellCenter,
+   cellRadius, set)` (camera-independent; cache per CellId, reuse for all that cell's
+   part-instances — like `WbDrawDispatcher` reuses one set per entity). Write the 8-int
+   set per instance into a new buffer parallel to `_gpuInstanceTransforms` (same shape
+   as `_clipSlotData`); bind at binding 5. On a no-lights frame, fill -1 (shader adds no
+   point light) and still bind a ≥1-element buffer so the SSBO is never unbound.
+4. **Cell centre/radius:** world-space bounding sphere of the cell geometry — reuse the
+   cell's existing visibility bound (the BSP/AABB sphere already computed for culling).
+   The exact field is pinned during planning by reading the cell-storage structs in
+   `EnvCellRenderer` / `EnvCellLandblock`; fallback = centre from the cell-part transform
+   translation, radius from the cell vertex AABB. **This is the one detail to confirm
+   against code in the plan.**
+
+Order independence: D-1 and D-2 are orthogonal (shader math vs buffer binding) and can
+be implemented in either order, but ship together.
+
+## Testing (TDD)
+
+`LightBake.cs` already encodes the correct math: `PointContribution` = per-light capped
+(matches `mesh_modern.vert` pointContribution line-for-line), `ComputeVertexColor` = sum
+reaching point lights → clamp `[0,1]`, skip directional. The new shader `pointAcc` clamp
+mirrors `ComputeVertexColor`'s final clamp exactly.
+
+New conformance test in `tests/AcDream.Core.Tests/` (e.g. `LightBakeConformanceTests`):
+
+- **Golden warm torch, bounded:** an orange `(1, 0.588, 0.314)` `intensity=100`
+  `falloff=4` (Range = 4×1.3 = 5.2 m) torch lighting a wall vertex (facing it) at
+  d = 1, 2, 3, 4, 5 m → result is warm (R ≥ G ≥ B, hue preserved) and **every channel
+  ≤ 1.0** (never white); at d ≥ Range the contribution is 0 (range gate).
+- **No-blowout under stacking:** 8 overlapping `intensity=100` near-white torches summed
+  via `ComputeVertexColor` → each channel clamps to ≤ 1.0 (the `[0,1]` saturate holds).
+- **Hue preserved:** a single orange torch's bounded result keeps B < G < R (warm), not
+  desaturated toward white.
+
+These pin the contract the shader must match. GLSL is not unit-testable in-process
+(standard for this project per the render digest); the shader `pointContribution` +
+`pointAcc` clamp are matched to `LightBake` by **line-for-line review** with the C#
+oracle as the pinned reference (call it out in the implementation commit).
+
+## Bookkeeping — divergence register
+
+- **Correct stale row AP-35** (`docs/architecture/retail-divergence-register.md`): it
+  describes the point-light path as per-pixel `mesh_modern.frag:52` with the half-Lambert
+  wrap "NOT ported". Reality since Fix A (`aa94ced`): per-vertex Gouraud in
+  `mesh_modern.vert:163` WITH the wrap ported. Update the row to match; the D-1 clamp
+  makes the accumulator MORE faithful (no new deviation introduced).
+- **EnvCell shell per-cell 8-light selection** (D-2) inherits Fix B's existing
+  per-object approximation (retail bakes per-VERTEX over the full static list; acdream
+  selects up to 8 per cell-sphere then gates per-vertex in-shader). Confirm Fix B's
+  register row covers EnvCell shells; extend that row if needed — do NOT add a
+  contradicting row.
+
+## Files
+
+- `src/AcDream.App/Rendering/Shaders/mesh_modern.vert` — D-1 clamp split.
+- `src/AcDream.App/Rendering/Shaders/mesh_modern.frag` — verify final clamp stays correct
+  (expected no change).
+- `src/AcDream.App/Rendering/Wb/EnvCellRenderer.cs` — D-2: `LightManager` ref, per-cell
+  light sets, bind SSBO 4 + 5, per-instance light-set buffer.
+- `src/AcDream.App/Rendering/Wb/WbDrawDispatcher.cs` (+ a shared `GlobalLightPacker`) —
+  extract the binding-4 global-lights packing so both renderers share it.
+- `src/AcDream.App/Rendering/GameWindow.cs` — wire `LightManager` into
+  `EnvCellRenderer.Initialize` (minimal).
+- `tests/AcDream.Core.Tests/Lighting/LightBakeConformanceTests.cs` — new.
+- `docs/architecture/retail-divergence-register.md` — AP-35 update.
+
+## Acceptance criteria
+
+- `dotnet build` green; `dotnet test` green including the new conformance test.
+- Conformance test passes on the captured golden torch values (warm, bounded, hue-preserved).
+- Shader `pointContribution` + new `pointAcc` clamp reviewed line-for-line against
+  `LightBake` (cited in the commit).
+- AP-35 corrected; any D-2 register note reconciled with Fix B's row.
+- **Visual (user):** outdoor objects near torches no longer blow out warm-white, and the
+  Holtburg meeting-hall walls render warm-but-dim like retail.
+
+## Out of scope (explicit)
+
+- **Do NOT port the D3D-FF hardware model** (`config_hardware_light`'s
+  `color×intensity`, `(0,1,0)=1/d`, `Range=falloff×1.5`) — it lights GfxObjs/dynamics,
+  not the baked walls. Wrong oracle (handoff warning stands).
+- **Do NOT** wire the CPU vertex bake (`LightBake.cs` as the runtime path) — chosen
+  approach is the in-shader clamp split. `LightBake.cs` stays the test oracle.
+- Sun handling on indoor walls is unchanged (kept in the material-lit term as today);
+  any "should indoor walls receive sun at all" refinement is a separate question.
+- The purple portal is correct — do not touch it.
--- a/tools/cdb/a7-fixd-golden-probe.cdb
+++ b/tools/cdb/a7-fixd-golden-probe.cdb
@ -0,0 +1,15 @@
+$$ A7 Fix D — GOLDEN: dump the nearest static lights (the meeting-hall wall torches)
+$$ + the ambient/sun that acdream folds into its accumulator. Breakpoint-free, instant.
+$$ Render::world_lights @ 0x008672a0; sorted_static_lights[] (RenderLight*) @ +0x3498
+$$ (verified: num_static_lights@+0x104=38, num_dynamic_lights@+0x3588=2).
+$$ Stand near the meeting-hall torches so the nearest sorted lights ARE them.
+.logopen C:\Users\erikn\source\repos\acdream\.claude\worktrees\thirsty-goldberg-51bb9b\a7-fixd-golden-probe.log
+.sympath C:\Users\erikn\source\repos\acdream\refs
+.symopt+ 0x40
+.reload /f acclient.exe
+.echo === ambient_color / sunlight_color / sunlight (what acdream folds into the accumulator) ===
+dt -r1 acclient!Render::world_lights ambient_color sunlight_color sunlight num_static_lights num_dynamic_lights
+.echo === nearest 10 sorted static lights (RenderLight.d3dLightIndex + info: type/intensity/falloff/color) ===
+.for (r $t0=0; @$t0 < 10; r $t0=@$t0+1) { r $t1 = poi(acclient!Render::world_lights + 0x3498 + @$t0*4); .printf "--- sorted_static[%d]  RenderLight=%p ---\n", @$t0, @$t1; dt -r2 acclient!RenderLight @$t1 d3dLightIndex distancesq info }
+.echo === END ===
+qd
--- a/tools/cdb/a7-fixd-golden2-probe.cdb
+++ b/tools/cdb/a7-fixd-golden2-probe.cdb
@ -0,0 +1,17 @@
+$$ A7 Fix D — GOLDEN v2: explicit LIGHTINFO/RGBColor dump of the nearest static
+$$ lights. info @ RenderLight+0x70 (LIGHTINFO); within info: color@+0x50, intensity@+0x5C,
+$$ falloff@+0x60 -> absolute color@RL+0xC0, intensity@RL+0xCC, falloff@RL+0xD0.
+$$ Characterizes the 38-light static set (warm town torches?) + golden for the fix.
+$$ Breakpoint-free, instant, uses current scene.
+.logopen C:\Users\erikn\source\repos\acdream\.claude\worktrees\thirsty-goldberg-51bb9b\a7-fixd-golden2-probe.log
+.sympath C:\Users\erikn\source\repos\acdream\refs
+.symopt+ 0x40
+.reload /f acclient.exe
+.echo === ambient_color (r,g,b) ===
+dt acclient!RGBColor acclient!Render::world_lights+0x0
+.echo === sunlight_color (r,g,b) ===
+dt acclient!RGBColor acclient!Render::world_lights+0xc
+.echo === nearest 8 sorted static lights: type/intensity/falloff + color(r,g,b) + distsq ===
+.for (r $t0=0; @$t0 < 8; r $t0=@$t0+1) { r $t1 = poi(acclient!Render::world_lights + 0x3498 + @$t0*4); .printf "--- sorted_static[%d] RL=%p d3dIdx=%d ---\n", @$t0, @$t1, dwo(@$t1+0x68); dt acclient!LIGHTINFO @$t1+0x70 type intensity falloff; .echo  color(r,g,b):; dt acclient!RGBColor @$t1+0xc0; .echo  distancesq:; dd @$t1+0xd8 L1 }
+.echo === END ===
+qd
--- a/tools/cdb/a7-fixd-lights-v2.cdb
+++ b/tools/cdb/a7-fixd-lights-v2.cdb
@ -0,0 +1,36 @@
+$$
+$$ A7 Fix D (#140) v2 — fills the two gaps v1 left:
+$$   (1) light COLORS (v1's dt did not expand RGBColor); expanded here as a
+$$       typed RGBColor dump + raw dd hex backup (reinterpret IEEE-754 if dt fails).
+$$   (2) the STATIC wall torches (the lights that actually BAKE the walls) — these
+$$       only re-register on a visible-cell-set change, so the player must MOVE
+$$       (walk IN and OUT of the meeting hall, circle past the torches) to trigger
+$$       Render::add_static_light.
+$$
+$$ v1 already proved: intensity=100/falloff=6 light is DYNAMIC (add_dynamic_light,
+$$ d3dIdx=2) = the portal/effect on the hardware path, NOT a baked wall torch.
+$$ viewer light = intensity 2.25 / falloff 10 (dynamic, d3dIdx=1).
+$$
+$$ add_static_light / add_dynamic_light(LIGHTINFO* info, cellID, Frame* offset):
+$$   LIGHTINFO* = poi(@esp+4). color@+0x50 (r/g/b floats), origin@+0x38, intensity@+0x5C, falloff@+0x60.
+$$
+$$ Dynamic logging is limited to the first 8 hits (we already characterised them);
+$$ ALL static hits log. qd when 12 static torches captured OR 1500 total hits (safety).
+
+.logopen C:\Users\erikn\source\repos\acdream\.claude\worktrees\thirsty-goldberg-51bb9b\a7-fixd-lights-v2.log
+.sympath C:\Users\erikn\source\repos\acdream\refs
+.symopt+ 0x40
+.reload /f acclient.exe
+
+r $t0 = 0
+r $t2 = 0
+r $t3 = 0
+
+$$ STATIC wall torches (baked path) — MOVE to trigger. Color (typed + hex) + origin.
+bp acclient!Render::add_static_light "r $t0=@$t0+1; r $t2=@$t2+1; .printf /D \"[STATIC torch] hit#%d\\n\", @$t2; dt acclient!LIGHTINFO poi(@esp+4) type intensity falloff cone_angle; .echo color_typed:; dt acclient!RGBColor poi(@esp+4)+0x50; .echo color_hex(r,g,b):; dd poi(@esp+4)+0x50 L3; .echo origin_hex(x,y,z):; dd poi(@esp+4)+0x38 L3; .if (@$t2 >= 12) { qd } .elsif (@$t0 >= 1500) { qd } .else { gc }"
+
+$$ DYNAMIC lights (portal/viewer) — log first 8 with color, then silent gc.
+bp acclient!Render::add_dynamic_light "r $t0=@$t0+1; r $t3=@$t3+1; .if (@$t3 <= 8) { .printf /D \"[DYNAMIC light] hit#%d\\n\", @$t3; dt acclient!LIGHTINFO poi(@esp+4) type intensity falloff; .echo color_typed:; dt acclient!RGBColor poi(@esp+4)+0x50; .echo color_hex(r,g,b):; dd poi(@esp+4)+0x50 L3 }; .if (@$t0 >= 1500) { qd } .else { gc }"
+
+.printf "v2 armed: STATIC=wall torches (MOVE in/out of hall to trigger), DYNAMIC=portal/viewer; colors expanded. qd at 12 statics or 1500 total.\\n"
+g
--- a/tools/cdb/a7-fixd-lights.cdb
+++ b/tools/cdb/a7-fixd-lights.cdb
@ -0,0 +1,50 @@
+$$
+$$ A7 Fix D (#140) — wall-torch vs portal light OWNERSHIP + the actual LIGHTINFO
+$$ values that feed the EnvCell wall bake. 2026-06-18.
+$$
+$$ Decomp already settled the render path (workflow wf_f660eb88):
+$$   STATIC lights -> CPU per-vertex bake (SetStaticLightingVertexColors ->
+$$     calc_point_light), DOUBLE-clamped (per-light min(scale*color,color) +
+$$     per-vertex [0,1]) -> walls stay DIM even at intensity=100.
+$$   DYNAMIC lights -> D3D hardware FF (minimize_envcell_lighting).
+$$ Render::insert_light copies intensity VERBATIM to BOTH paths, so the only
+$$ open empirical question is: which light carries intensity=100, and what do
+$$ the actual wall-torch LIGHTINFOs look like (intensity/falloff/color)?
+$$
+$$ CLASSIFICATION via config_hardware_light's d3dLightIndex (arg1 @ [esp+4]):
+$$   add_dynamic_light base index = 1   -> dynamic idx in [1..10]  (viewer light / teleport PORTAL)
+$$   add_static_light  base index = 11  -> static  idx in [11..70] (WALL TORCHES, baked)
+$$
+$$ config_hardware_light(d3dIndex, _D3DLIGHT9* out, ulong cellID, LIGHTINFO* info):
+$$   d3dIndex = dwo(@esp+4) ; LIGHTINFO* = poi(@esp+0x10)   (PROVEN last session)
+$$ add_static_light / add_dynamic_light(LIGHTINFO* info, cellID, Frame* offset):
+$$   LIGHTINFO* = poi(@esp+4)
+$$ `dt acclient!LIGHTINFO <ptr> type intensity falloff color` resolves the
+$$   float fields symbolically (PDB types) -> readable values, no hex reinterp.
+$$
+$$ USAGE: with retail in-world standing in/near the Holtburg meeting hall by a
+$$ wall torch, WALK around the hall (and past the teleport portal if present)
+$$ for ~15 s so static torch sets re-register. Auto-detaches (qd) after 600
+$$ total hits, leaving retail running.
+
+.logopen C:\Users\erikn\source\repos\acdream\.claude\worktrees\thirsty-goldberg-51bb9b\a7-fixd-lights-capture.log
+.sympath C:\Users\erikn\source\repos\acdream\refs
+.symopt+ 0x40
+.reload /f acclient.exe
+
+r $t0 = 0
+r $t1 = 0
+r $t2 = 0
+r $t3 = 0
+
+$$ BP1: config_hardware_light — EVERY light (static+dynamic); d3dIdx classifies.
+bp acclient!PrimD3DRender::config_hardware_light "r $t0=@$t0+1; r $t1=@$t1+1; .printf /D \"[CHL] hit#%d d3dIdx=%d (1-10=DYNAMIC portal/viewer, 11+=STATIC torch)\\n\", @$t1, dwo(@esp+4); dt acclient!LIGHTINFO dwo(@esp+0x10) type intensity falloff color; .if (@$t0 >= 600) { qd } .else { gc }"
+
+$$ BP2: add_static_light — every hit is a WALL TORCH (baked path).
+bp acclient!Render::add_static_light "r $t0=@$t0+1; r $t2=@$t2+1; .printf /D \"[STATIC torch] hit#%d\\n\", @$t2; dt acclient!LIGHTINFO dwo(@esp+4) type intensity falloff color; .if (@$t0 >= 600) { qd } .else { gc }"
+
+$$ BP3: add_dynamic_light — viewer light + teleport PORTAL (hardware path).
+bp acclient!Render::add_dynamic_light "r $t0=@$t0+1; r $t3=@$t3+1; .printf /D \"[DYNAMIC light] hit#%d\\n\", @$t3; dt acclient!LIGHTINFO dwo(@esp+4) type intensity falloff color; .if (@$t0 >= 600) { qd } .else { gc }"
+
+.printf "a7-fixd-lights armed: BP1 CHL (classify via d3dIdx), BP2 STATIC=torch, BP3 DYNAMIC=portal/viewer. qd after 600 total hits.\\n"
+g
--- a/tools/cdb/a7-fixd-numstatic-probe.cdb
+++ b/tools/cdb/a7-fixd-numstatic-probe.cdb
@ -0,0 +1,18 @@
+$$ A7 Fix D — instant (breakpoint-free) read of how many STATIC lights the
+$$ current scene bakes with. Confirms whether the meeting hall has static torches
+$$ (-> D-1 summed-torches matters) or near-zero (-> D-2 leaked-SSBO is the cause).
+$$ Stand where the meeting-hall walls are visible. No movement / no breakpoints.
+.logopen C:\Users\erikn\source\repos\acdream\.claude\worktrees\thirsty-goldberg-51bb9b\a7-fixd-numstatic-probe.log
+.sympath C:\Users\erikn\source\repos\acdream\refs
+.symopt+ 0x40
+.reload /f acclient.exe
+.echo === x acclient!*world_lights* ===
+x acclient!*world_lights*
+.echo === x acclient!Render::world_lights ===
+x acclient!Render::world_lights
+.echo === dt typed (num_static_lights / num_dynamic_lights / ambient_color) ===
+dt acclient!Render::world_lights num_static_lights num_dynamic_lights ambient_color sunlight_color
+.echo === dt LightParms at symbol (fallback by explicit type) ===
+dt acclient!LightParms acclient!Render::world_lights num_static_lights num_dynamic_lights
+.echo === END ===
+qd