diff --git a/docs/research/2026-06-19-lighting-a7-fixD-round2-torch-reach-CHECKPOINT.md b/docs/research/2026-06-19-lighting-a7-fixD-round2-torch-reach-CHECKPOINT.md
new file mode 100644
index 00000000..94924439
--- /dev/null
+++ b/docs/research/2026-06-19-lighting-a7-fixD-round2-torch-reach-CHECKPOINT.md
@@ -0,0 +1,94 @@
+# A7 Fix D round 2 — REAL cause found (object sun+ambient + torch REACH), CHECKPOINT
+
+**Date:** 2026-06-19  **Branch:** `claude/thirsty-goldberg-51bb9b` (NOT merged — held at the visual gate)
+**Predecessor:** `docs/research/2026-06-18-lighting-a7-fixABC-shipped-fixD-handoff.md`
+**Status:** checkpointed at user request after pinning the root cause. D-1..D-4 are committed +
+correct but **did NOT fix the visible symptom** — they were the wrong subsystem.
+
+## TL;DR — what the visible bug actually is (and is NOT)
+
+The user's symptom (Holtburg meeting-hall facade too bright/warm/washed-out + character backs
+lit) is **NOT** the EnvCell bake, the per-channel clamp, the half-Lambert wrap, or the SSBO leak.
+Those are the D-1..D-4 path. **The visible surfaces are mode-0 OBJECTS**, and the cause is:
+
+1. **Building facade over-bright** = the **torch REACH is too long** (acdream ~7.8 m vs retail
+   ~5.2 m), so each entrance torch floods the whole small facade instead of a tight pool.
+   **CONFIRMED by isolation**: gating object (mode-0) point lights OFF made the building match
+   retail ("looks much better", user 2026-06-19).
+2. **Character backs / slight object over-bright** = the **sun + ambient on objects** (mode 0
+   runs both). Ambient is NOT the culprit (it MATCHES retail exactly — see values). The residual
+   is small for the character (it ~matches retail), so the dominant visible bug is #1 (torches).
+
+## Render-path facts (source-verified, workflow `wf_c4ad8cf8`)
+
+- **Building EXTERIOR** = a flat-mesh `WorldEntity` with `IsBuildingShell=true`, `ParentCellId=null`,
+  built from `BuildingInfo.ModelId` (`LandblockLoader.cs:79-90`), drawn by **WbDrawDispatcher**
+  which hard-sets `uLightingMode=0` (`WbDrawDispatcher.cs:898`). It is **NOT an EnvCell** — so
+  **D-4 (EnvCell walls get no sun) never touched it**.
+- **Characters/creatures/players** = ordinary `WorldEntity` dynamics, also drawn by
+  WbDrawDispatcher at `uLightingMode=0` (plain Lambert + sun). The mode plumbing is CORRECT
+  (mode-0 plain Lambert already zeroes a torch behind a back-face — that part of D-3 works).
+- **EnvCellRenderer** (`uLightingMode=1`, no-sun, wrap) only ever draws **interior** cell shells
+  from the dat EnvCell list — never `info.Buildings`, never characters.
+- Render loop: in-world frames go through `RetailPViewRenderer.DrawInside`; the bare
+  `WbDrawDispatcher.Draw` (GameWindow.cs:8230) is the no-viewer-cell fallback. Both share the
+  ONE `_meshShader` (mesh_modern) program (GameWindow.cs:1845-1857), so `uLightingMode` is one
+  shared uniform; each renderer re-sets it before its draws.
+
+## Ground truth (live cdb retail + acdream probe, SAME-INSTANT)
+
+- **Ambient MATCHES exactly**: acdream `(0.447,0.447,0.495)` == retail `(0.4465,0.4465,0.4951)`.
+  → same sky keyframe → **same time of day; NO time desync** (the earlier "retail 0.3 / acdream
+  purple" was sequential-capture drift + acdream's un-synced spawn frame; ignore it).
+- **retail sun** (`world_lights.sunlight` @ 0x008672a0+0x18) = `(0.573, ~0, 0.445)`, magnitude
+  **0.725**, colour `(0.98,0.84,0.59)` warm. acdream `sun=1` (active, derived from the same sky
+  state via Fix C `|sunVec|=DirBright`). Sun is NOT zero — retail DOES sun-light objects.
+- **retail torches** (golden, a7-fixd-golden2): static, `intensity=100`, `falloff 3/4/5`, warm
+  `(1,0.588,0.314)` orange + `(0.98,0.843,0.612)` cream. `calc_point_light` makes a BRIGHT TIGHT
+  pool (saturates to full warm to ~4 m, gone by ~5.2 m). Faithful in acdream (LightBake.cs).
+- **acdream torches** ([light-detail]): `range=7.8` (Falloff 6×1.3) and `range=6.5` (Falloff 5).
+  acdream `Range = info.Falloff * 1.3f` (`LightInfoLoader.cs:90`) — the 1.3 is correct, NO stray 1.5.
+
+## The OPEN question to resolve FIRST on resume (don't guess)
+
+acdream's orange torch reads **Falloff 6** (range 7.8); retail's orange torch was captured at
+**Falloff 4** (range 5.2). `6 = 4 × 1.5` (smells like rangeAdjust) BUT they **might be different
+torches** (38 static torches, several orange). **Resolve by comparing the SAME torch's Falloff in
+acdream vs retail, matched by world position** (one focused capture): break/dump acdream's torch
+Falloff for a specific Holtburg torch and the retail `world_lights.static_lights[i].info.falloff`
+for the same one. Then:
+- If acdream reads a **too-large Falloff** for the same torch → fix the dat read / conversion
+  (the DatReaderWriter `LightInfo.Falloff` path) so acdream's reach == retail's.
+- If the Falloff matches and reach is genuinely ~7.8 → the building-shell-as-one-object spill is
+  the issue; tighten how building shells receive torches (the per-vertex range gate already
+  localises, so this is unlikely — favour the Falloff hypothesis).
+
+## Proposed fix (after the falloff is confirmed)
+
+Tighten acdream's torch reach to match retail (≈5 m), keep torches ON. Building facade then shows
+a tight warm pool by each flame + dark stone elsewhere (retail-faithful). Files: `LightInfoLoader.cs`
+(the Falloff→Range conversion), possibly the DatReaderWriter light read. Add a divergence-register
+row if any conversion deviates. Re-verify visually (the diagnostic that confirmed the cause:
+object point lights OFF == retail-match).
+
+## State of the committed work (KEEP — all correct, just off-target for the visible bug)
+
+| Commit | What | Verdict |
+|---|---|---|
+| `180b4af` | D-1 clamp point sum on its own | faithful; keep |
+| `39c70f0` | D-2 prep — LightBake conformance test | keep |
+| `cf62793` | D-1 shader clamp | keep |
+| `c62da82` | D-2 EnvCell shell binds own light set (real leak fix) | keep |
+| `b57a53e`/`156dc45` | register AP-35/AP-16 corrections | keep |
+| `0980bea` | D-3 objects plain-Lambert / D-4 EnvCell no-sun | keep; correct but doesn't touch the building (it's an object) |
+
+`tools/cdb/a7-fixd-*.cdb` capture scripts are committed. **Diagnostic shader hack reverted**
+(working tree clean). Branch NOT merged — finish the torch-reach fix, visual-verify, then merge.
+
+## DO-NOT-RETRY (cost a lot this session)
+
+- Don't re-tune the EnvCell bake / per-channel clamp / wrap / SSBO binding for the building — the
+  building is a mode-0 OBJECT, none of that path lights it.
+- Don't chase a time-of-day / ambient desync — ambient + time MATCH retail exactly (0.446).
+- Don't "remove the sun" globally — retail DOES sun-light objects (sun 0.725).
+- The visible building bug is the **torch REACH** (confirmed by isolation); start there.