Aranna DLL startup outcome, 2026-11-05

The one current original session completed DLL initialization normally and reached the genuine EULA. After ordinary Accept, its first actual failure was the C:\Resources\Logic.dsres critical-error dialog, last error 0x00000017 (23). No menu or gameplay was reached. Checksum instrumentation remains deferred; the dialog alone does not establish a checksum calculation or ReadFile divergence.

Evidence: scratch/runs/20261105-aranna-loader-outcome/. Task: NEW-GAME-ARANNA-DEMO-20261010. Diagnostic repair published on main: f78181348d42a1daf2fa2d149c83f401433017e5.

What changed and why

The preceding ordered-read investigation failed before EULA: DLL initializer did not return: module=0x9c0000. Its browser forwarding discarded the attached outcome, fault/cause, and owning register context. Mode 1 and a checksum breakpoint were armed before startup. That session's lost outcome cannot be reconstructed; observer interference remains unexcluded. The historical ordinary original run reached EULA and then error 23, but was not the ordered-read session's outcome. The Worker read controls found no read/seek divergence and do not establish original checksum behavior.

lib/dll-loader.js now retains halt/yield alongside the already captured pre-restoration EIP, ESP, return value, and fault. lib/guest-worker.js forwards the existing loader log and explicitly serializes error name/message/stack, outcome, fault/cause, and nested rollback errors. lib/guest-thread-host.js reattaches those details and the owning Worker slot/command stage before rejecting the caller. This repairs diagnostic loss; it does not change guest execution, callback scheduling, fault mode, or frame restoration. The three callback-census hashes were updated after reviewing those lifetime invariants; existing pending ownership classifications remain unchanged (lifetime-review.json).

Validation before the original

The new test/test-dll-startup-rollback.js controls exercise real generated PE DllMain code through production GuestThreadHost/Worker transport: successful TRUE return, NULL-jump incomplete frame, UD2 trap, and FALSE attach followed by an incomplete detach. The added regression failed on unchanged production code because the existing success trace was lost (baseline-transport.log). After the repair both standard and compatibility modules passed, including retained original/rollback causes, final owning EIP/ESP, halt/yield, and caller register restoration. Existing rollback tests also passed.

Actual Chrome Worker controls repeated all four cases with the same passive hooks planned for the original, defaultViewport:null, and no debug setters or breakpoints (browser-controls.json, generated-browser/). Incomplete outcome: EIP0, ESP0x74fffec, halt2, yield0. Trap outcome: EIP0x504000, ESP0x74fffec, halt0, yield0, retained RuntimeError unreachable as fault/cause. The restored caller ESP0x74ffffc is distinct and is not the failing frame.

Additional passing checks: test/test-dllmain-completion.js, test/test-worker-bootstrap-retirement.js, callback-census reconciliation tests, and flock /tmp/wine-build.lock bash tools/build.sh (build-reviewed.log). The first full build correctly rejected stale census hashes (build.log); the reviewed hash updates fixed that validation failure without bypassing it.

The one original observation

Runtime gate: legacyADD1955 FINAL18:41:45, AoE3repairedDXSETUP2005 FINAL18:55:54, and Conanresidency1825 FINAL19:17:29 were observed in order. Independent release verification passed 385 absent PIDs, 152 refused ports, 41 absent profiles, four free locks, zero heavy processes, and disk headroom (release-gate.json). Runtime acquired19:18:52UTC; the original ran19:22:29–19:24:15UTC. There was one original diagnostic, zero original retests, and no installer/media copy.

All 28 canonical inputs (281799760 bytes) were rehashed in place against the recorded prior manifest (media-verification.json), using scratch/game-media/aranna-demo-20261105/extracted/. Ordinary threads/GPU startup used Chrome151 on Linux x86_64, SwiftShader, defaultViewport:null, viewport 1000x800. Passive transport/module retention made no guest export calls; no fault-mode setter or breakpoint command was observed. Default execution mode was preserved; there is no independent numeric mode readback claim.

Owning Worker0 observations (observations.jsonl):

Both actual original Workers received the same standard module (2173782 bytes): 78a9e8d7563c94b1e002a688c5e65d4269946ab8dd9eb68aa3edf271f4cb1c24. The tested compatibility module (2176428 bytes): 5663c49c8620789c681af2133839a886ba5d5ea877c988289c7a0a682bf29b66. browser-modules.json, retained tested modules, and tested-modules.json record identity; runtime-source.json records exact runtime commit f78181348 and runtime-relevant.binarydiff is empty. The precommit SHA and relevant binary patch, owned new-file patches, captures, and hash manifest are retained. Git is the source archive; no source-tree archive or repeated media was made.

Causal limit and next boundary

Current initialization success establishes the normal startup boundary on the current main and permits future resource investigation. It does not prove why the earlier instrumented startup failed: that outcome was lost and other main changes intervened. No CPU/ReadFile divergence or causal guest fix is claimed. Next, distinguish the resource validation/read failure after successful startup, using existing tracing first; introduce checksum observation only after proving it leaves initialization and EULA progress unchanged. There is no reason to repeat this unchanged original diagnostic.

Final independent cleanup at19:26:03UTC passed six absent owned PIDs (including watcher), four refused ports, two absent profiles, four free locks, no heavy processes, and 7.11GB free (final-cleanup.json). Generated-browser cleanup also passed independently. No boxes were assigned: box ID/state/expiry N/A; no sandbox lifecycle action, subworkers, or deployment. Runtime/file ownership is handed back with the final board release, within45minutes of acquisition.