diff --git a/docs/ISSUES.md b/docs/ISSUES.md
index ae486384..ad437873 100644
--- a/docs/ISSUES.md
+++ b/docs/ISSUES.md
@@ -4225,6 +4225,50 @@ selector.
 
 ---
 
+## #125 — GL InvalidOperation during staged texture upload: failed uploads are STICKY (never retried) + uncaught crash in GenerateMipmaps
+
+**Status:** OPEN — HIGH (the likely #119 missing-stairs mechanism + a process crash)
+**Filed:** 2026-06-11 (in-tower WB_DIAG launch, `tower-wbdiag3.log` — preserved in the worktree root)
+**Component:** render — WB staged texture pipeline (ObjectMeshManager / ManagedGLTextureArray)
+
+**Evidence (one launch, character spawned inside the #119 tower):**
+1. `[wb-error] Error uploading mesh data for 0x0100321D` — GL
+   `InvalidOperation` thrown in `ManagedGLTextureArray..ctor:70`
+   (TextureAtlasManager ctor → CreateTextureArrayInternal), caught by
+   `UploadMeshData`'s try/catch → returns null. **The drop is STICKY:**
+   `_preparationTasks.TryRemove` runs BEFORE the upload
+   (ObjectMeshManager.cs:685), so a failed upload is never re-prepared —
+   that mesh is permanently invisible for the session (only a one-line
+   [wb-error] marks it).
+2. Same session, `Tick()` → `GenerateMipmaps()` →
+   `ManagedGLTextureArray.ProcessDirtyUpdatesInternal:283` threw the SAME
+   GL InvalidOperation **uncaught** → process death (exit 82). Both on the
+   render thread (Tick/OnRender) — not a thread-affinity bug.
+
+**Why this matters for #119:** the missing tower stairs are per-cell
+Setup statics whose parts are individually uploaded; an intermittent GL
+error burst during atlas creation/flush kills whichever uploads are in
+flight — "partially invisible", varying with load order, hitting the
+late-loading AAB3 interior statics consistently. The dat + extraction +
+registration + dispatcher are all exonerated by read/test
+(Issue119TowerDumpTests; the [up-null] pair was a separate, legitimate
+no-draw class).
+
+**Open questions (next session):** (a) what makes the GL context error
+out — a stale error queued by an earlier unchecked call being
+mis-attributed to WB's diligent glGetError checks (classic GL
+attribution trap; suspects: the #117 stencil punch state, the new #118 /
+#121 passes, or a pre-existing per-frame state leak), vs. a genuine
+invalid texture-array creation state; (b) whether upload failures should
+re-enqueue instead of dropping (retail has no such failure mode — the
+sticky drop is OUR invention and must go regardless); (c) the uncaught
+GenerateMipmaps path needs the same handling either way.
+**Repro lever:** the test character's save spawns INSIDE the tower —
+every launch loads the exact content; `ACDREAM_WB_DIAG=1` prints the
+meshMissing counters.
+
+---
+
 # Recently closed
 
 ## #113 — Phantom staircase: REOPENED 2026-06-11, folded into the HOLISTIC BUILDING-RENDER PORT
diff --git a/src/AcDream.App/Rendering/Wb/WbDrawDispatcher.cs b/src/AcDream.App/Rendering/Wb/WbDrawDispatcher.cs
index 2f5a16f9..f7214e9e 100644
--- a/src/AcDream.App/Rendering/Wb/WbDrawDispatcher.cs
+++ b/src/AcDream.App/Rendering/Wb/WbDrawDispatcher.cs
@@ -1583,7 +1583,12 @@ public sealed unsafe class WbDrawDispatcher : IDisposable
         int nz = 0;
         foreach (var v in copy) if (v > 0) nz++;
         if (nz == 0) return 0;
-        return copy[copy.Length - nz / 2];
+        // Sorted ascending: zero-padding front, samples at the back. (nz - 1) / 2
+        // from the end keeps the offset >= 0 for all nz >= 1 — the original
+        // nz / 2 form indexed copy[copy.Length] (crash) on the first diag flush
+        // when exactly 1 sample was recorded. Same fix as GameWindow's
+        // TerrainDiagMedianMicros twin.
+        return copy[copy.Length - 1 - (nz - 1) / 2];
     }
 
     private static long Percentile95Micros(long[] samples)