Diablo II repeated Load: retired texture wrappers exhaust the auxiliary pool

Evidence: scratch/runs/20261111-diablo2-native-sprite-load-repair-1750/. The genuine four-file XP260 checkpoint first loaded successfully in a fresh native process on the pinned parent (0db20330ccd6761a705eea70c5dd9bfc7704c147, module ce57f0ad80254b974a0d8bcd4e2c9cbf396d2357cbdc7874e6084e980e1e68e1). A native Save & Exit exported byte-identical files; the second Load succeeded. The third Load reproduced the original d3dSprite.cpp:85, Expression: success assertion. Thus checkpoint bytes alone do not cause this failure.

A passive host import observer recorded real guest API arguments, return addresses, and EAX after dispatch; it did not alter arguments, guest memory, or HRESULTs. The failing call was IDirectDrawSurface::QueryInterface(IID_IDirect3DTexture2) returning 0x8007000e (E_OUTOFMEMORY) with a null output. The final failing call returned to runtime address 0x018ad7d2, corresponding to original D2Direct3D.dll 0x100037d2. Disassembly connects this helper's false return to line 85. The immediately preceding CreateSurface succeeded. A read-only snapshot found COM_AUX_NEXT_SHARED=16384, 11,927 auxiliary wrappers whose DX_OBJECTS owner was already freed, and 11,924 free DX object slots. This is auxiliary interface wrapper exhaustion, not failed surface storage or save corruption.

dx_get_wrapper_for_vtbl_locked now reuses an initialized wrapper belonging to a logically freed owner when its append pool is full. It retains the existing lookup-first behavior and holds LOCK_DX during allocation. A zero interface count alone does not permit reuse: the owner's DX_OBJECTS type must be zero, preserving all views of live objects, including separately retained surface interfaces. Reference bookkeeping is cleared for the reclaimed entry. If every owner is live, surface QI still fails without changing a published vtable, output ownership, or reference count.

The Surface4/7 ABI candidate fills the real pool using ordinary guest-x86 COM calls and keeps those views alive. It checks genuine exhaustion, stable live vtable/identity/pixels/reference counts, and a successful new QI after final Release of one filler object. There is no fabricated COM record or cursor fault injection. This regression fails at the reuse assertion on the pre-fix current-main module 5512a65e2ce88cc76b39b3246d35f13919d9df060d3670a9a85cae9c4100f638 and passes with the repair in standard and compatibility modes. The ordinary DirectDraw interface reference-count test and locked canonical build pass. An attempted pinned-module/current-host fixture had mismatched API identities and is retained as an invalid fixture attempt, not regression evidence.

The original game then completed three ordinary Loads in one repaired process using the genuine milestone files and native Save & Exit between Loads. review/repair-world3.png and review/repair-stats3.png show the third Load alive in camp. At that point the cursor was again 16,384, while all 29,752 observed QueryInterface calls and 29,759 CreateSurface calls had succeeded. The same original return site 0x018ad7d2 now returned zero with a valid Texture2 pointer. The fixed canonical module is 1165f15cb29fa4ff4b67621f3ce66f8df68fb50baea929d0b76e9b2f66118135. See the run's RESULT/RECOVERY for subsequent gameplay, separate checkpoints, assistance disclosure, complete evidence audit, and the final stop receipt.