diff --git a/docs/ISSUES.md b/docs/ISSUES.md
index d991f721..4ed75097 100644
--- a/docs/ISSUES.md
+++ b/docs/ISSUES.md
@@ -10554,6 +10554,67 @@ missing is the plugin-API surface.
 
 ---
 
+## #334 — Large static formations lose collision at their boundaries (Neftet)
+
+**Status:** OPEN
+**Severity:** HIGH — walk-through and fall-through on world geometry.
+**Filed:** 2026-08-06, user-reported in live play.
+**Component:** physics / collision / broadphase
+
+**User report, verbatim in substance:** *"In Neftet I miss collision on the
+large stone formations. I can run up to them, but then at a boundary between
+I can run through. I can jump over it but then I just fall through."*
+
+**NOT a regression from the 2026-08-06 collision work.** Confirmed by A/B: the
+user reproduced it on a purpose-built client at `52175aa1` — before AP-22,
+AP-152, AP-156, AD-10 and #276's remainder — and the behaviour was identical.
+It is pre-existing and was simply never filed. AP-156's cell-membership fix
+did **not** resolve it either, which is itself a diagnostic clue (see below).
+
+**The signature:** solid on approach, permeable at a boundary *between* two
+formations, and no floor above it — jumping over lands you inside/through.
+"Solid near the middle, absent at the edges" is the shape to reason from.
+
+### Leading candidate — AP-158 / #333, the broadphase reach filter
+
+`TransitionTypes.cs:3898`-ish measures `currPos - obj.Position` (the part
+**origin**) against `obj.Radius`, admitting a contact only when the geometry
+sits within roughly `movement + 2 m` of that origin. Retail has **no distance
+pre-filter at all**: `CObjCell::find_obj_collisions` @`0x0052b750` dispatches
+unconditionally, its only early-out being `INITIAL_PLACEMENT_INSERT`. The 2 m
+slack is acdream invention; retail's own epsilon in this family is 0.0002 m.
+
+**118 of 477 unique BSP GfxObjs exceed that ~2.5 m budget; 46 exceed 5 m.** A
+"large stone formation" is precisely the class that would: geometry extending
+many metres from its part origin. Near the origin you collide; past the budget
+the broadphase rejects the object before its mesh is ever consulted — solid in
+the middle, permeable at the edges, no floor overhead. That matches the report
+without needing a second mechanism.
+
+This also explains why AP-156 did not help: AP-156 put the object in the right
+**cells**; AP-158 is why it is still rejected *within* those cells. AP-156 was
+a prerequisite, not a cure — the two are sequential, and this issue is the
+observable that proves the second half still bites.
+
+### Second candidate, if the first is disproved
+
+Static-object collision registration at EnvCell seams — the per-cell shadow
+list family (#98's architecture, closed `b3ce505`, and #137's door/wall-opening
+work, closed 2026-07-08). Both are closed, so this would be a new gap rather
+than a regression of either.
+
+### How to settle it
+
+1. Identify the specific Setup/GfxObj of a Neftet formation that reproduces,
+   and measure its BSP root-sphere origin offset. If it exceeds ~2.5 m, AP-158
+   is confirmed as the cause and this issue closes with AP-158's fix.
+2. If the offset is small, the reach filter is exonerated and the second
+   candidate takes over.
+
+**Do not attempt a fix before step 1.** The reach filter's `+ 2f` slack has no
+retail counterpart, so "widen the budget" would be tuning an invented constant
+— the fix is to remove the filter, which needs AP-158's own gate.
+
 ## #32 — Retail edge-slide / cliff-slide / precipice-slide incomplete
 
 **Status:** REMOTE HALF CLOSED — fixed `204d0ae0`, **user-passed 2026-08-04**