Diablo II native SaveExit worker wait, 2026-11-11

Run: scratch/runs/20261111-diablo2-native-saveexit-worker-wait-1910/. The runtime stall is unresolved. No speculative guest, save, scheduler or sound change was made. This note records the narrowed causal boundary and the exact comparison that replaces the parent's invalid mixed fixture.

Identity and comparison

The source/host closure at a70573e009ca17e46b74de8b3e548a70ba7442cd produces the published repair's standard module 1165f15cb29fa4ff4b67621f3ce66f8df68fb50baea929d0b76e9b2f66118135. The locked build passed. Relevant tracked runtime sources match repair 7c985edcc8eb8c23e7f2b64ba8b0d6f2eb177bad; the closure manifest also pins untracked lib/wininit-boot.js, recorded as a patch rather than a source archive.

Rebuilding with only src/09a8-handlers-directx.wat replaced by its Git 3dc8bb1103f960b476021e4e5da18f11f452e6d3 version reproduces the exact baseline standard module 5512a65e2ce88cc76b39b3246d35f13919d9df060d3670a9a85cae9c4100f638. The unchanged surface ABI test, consuming each exact standard module through bootRenderHarness({prebuiltModule}), fails on baseline at “final Release makes wrapper space reusable” and passes on repair. Source-compiled surface ABI and wrapper-pool sizing tests also pass. This validates wrapper reuse, not a SaveExit repair. No native before/after SaveExit claim uses the mixed fixture.

Causal boundary

The parent's stalled main EIP 0x0143b310 relocates to Storm 0x6ffc9310: an infinite WaitForSingleObject on a job-completion event. 0x0e000f58 is in the host's generation-tagged synchronization handle domain; it is not a worker-thread join handle. The originating Storm path creates a manual-reset, initially nonsignalled event. In the new run, passive import traces show main creating/waiting on these events and worker1 signalling them from 0x01437c5e (Storm+0x15c5e). Full sequences are retained in successful-event-chains.json.gz, with original-byte hashes in own-closed-compression.json.

The parent's worker1 exit had EIP zero, previous decoded blocks 0x01438da1 / 0x01438d90, EAX/ECX zero, ESI 0x0923016c, EDI 0x0923018c, and ESP 0x7effffa0. Disassembly identifies the block as Storm's streamed-audio IDirectSoundBuffer::Lock call: it loads [ESI+0x28] into EAX, its vtable into ECX, pushes eight arguments and calls [ECX+0x2c]. The stack depth and registers fit a null buffer call before a COM handler is entered, rather than evidence of a wrong Lock epilogue. This is a leading hypothesis from the captured registers and unedited DLL bytes, not proof of the pre-call values: an API call can change EAX, and a bad return or interpreter corruption is not excluded. The parent did not capture the pre-dispatch object memory or any writer clearing its buffer field.

The worker also services file jobs. Its loss can strand a later completion wait, explaining why saved files can coexist with a frozen options screen. It does not establish which write invalidated the sound-buffer pointer, or whether the wrapper repair indirectly affects timing. Artificially signalling the event would falsely report an unfinished job complete and is not a repair.

Enter/LeaveCriticalSection and sound-buffer calls were observed during the third native cycle; primary sound-buffer pointers, not texture aux views, were used by the streamed Lock path. The DirectSound Lock ABI consumes its normal 36-byte frame. The aux repair only reclaims wrappers whose owner type is zero under LOCK_DX; its new writes target the dedicated aux/refcount arenas. These inspections do not exclude a lifetime or synchronization race elsewhere.

Passive pre-detach and post-slice worker observers preserve the stack, object, buffer and sync ring on a recurrence. Their instrumentation can perturb timing. The second observer setup installed successfully but failed while writing an already-existing evidence filename; a separate read-only check confirmed both observers armed on all three guest workers. It was not replayed.

Native checkpoint and limits

The complete inherited XP296/life46 four-file set and the separate older XP260/life66 set remain byte-identical. Akara healed the XP296 character through native input. Four SaveExits completed normally; five same-process Loads reached camp. Native character panels after the fourth and fifth Loads verify XP296/life66. Additional sound/critical-section logging was removed before the fourth SaveExit; passive sync and exit observers remained enabled. The new complete healed296-save/ set has a 753-byte D2S with SHA-256 e51c394a482f65de379192b28bf5de45bea1798ac37537f56a42c1e0eec17fd9. The later complete field296-save/ set has the same D2S and a 7,830-byte samehero.ma0 (SHA-256 04a751c98f3a497536db8784217729e62a2ebde7aa033ccbed911665d3da6e9c). Native Blood Moor exploration added map coverage; no further XP or Den entry was achieved. The fifth Load verified this latest checkpoint. Each native full image was inspected after its delivery finished and before a separate input call. Three stale-scene inputs were refused before dispatch with zero effects; fresh captures were reviewed before separately named inputs. A final capture request omitted its required label and was rejected by the client validator before publishing any command. A new capture succeeded. The parent's image-review batching deviations remain explicit in its review-corrections.json; this run does not retroactively qualify them.

Final native route, checkpoint, recovery and box-stop results are recorded in the run's RESULT.md and RECOVERY.md. The next causal experiment needs the writer/lifetime transition of the streamed job's buffer field before the null call, alongside the exact completion handle and section owner. A passing repeat alone must not be called a causal SaveExit fix.