Aranna natural animation completion, 2026-11-07

No emulator mismatch was demonstrated and no emulator implementation was changed. Read-only disassembly corrects the previous diagnostic's interpretation of animator byte +0x7e: it behaves as a consumed pose-dirty flag, cleared by the pose-update exit, rather than an animation enable that must remain set. Its sampled zero does not explain missing completion. A generated native-i386/WASM control found no difference in the nearby x87 comparison/branch patterns or the indirect C++ callback convention. Natural title completion remains unexplained; world gameplay remains unverified. Earlier menu/Choose Hero evidence is accepted.

Evidence and reproducible diagnostic inputs are in scratch/runs/20261107-aranna-animation-completion/. This follows the completed EULA/title diagnostic; it does not repeat its corrected EULA route or infer a clock defect from clock-read ratios. No original was launched, no box was created, and no guest memory, configuration, assets, state, or completion signal was changed.

What the retained evidence actually proves

The predecessor's resolved snapshots retain animator 167032084, frontend receiver 150089148, clip fields +0xc4=13, +0xd4=13, +0xd8=0, +0xdc=1, pending-clip byte +0xe4=0, and binding [0x4c0236,0,150089148,0x4f3cab] at animator +0xf4. Frontend state is 4 with neither an immediate nor delayed pending request. The advancing guest clock does not measure the float dt actually passed to this animator.

The retained observations contain actual frontend execution at 0x4bccac and branch-edge observations at 0x4bccbf/0x4bccca during completed transaction 298, at 05:15:49.436Z. The transaction's outer wrapper returns at 0x43b4af. retained-title-executions.json preserves these records and their arguments. These are positive observations of the input/frontend path, not instruction traces inside the animator. In particular, event 74 reports branch edges; its fall/target fields must not be presented as a callback-entry probe.

The observer transforms retained by the preceding runs probe frontend/state admission and wrapper paths. They do not probe 0x69b615, 0x69deb8, 0x6a4bd7, 0x69b048, 0x4f3cab, or 0x4c0236. Thus a missing callback record cannot prove that the callback was never invoked. A binding still present in two snapshots cannot establish its invocation history either.

Personally inspecting the retained images also corrects a predecessor reporting error: original/response-11.png is teal desktop, but original/response-12.png shows the title over stone doors again. Response 12 SHA-256 is 330d41b9883271405b69e7d4fc45b67e301133c7c55a511e26d897ec5688151a. This does not establish why presentation changed, or establish menu/gameplay. The original files remain in the predecessor run; no duplicate media/capture archive was made.

Static lifecycle, distinguished from observed execution

Addresses below come from the existing demo executable scratch/game-media/aranna-demo-20261105/extracted/DSLOA.exe, SHA-256 fced0961231248c09c10d86ff51935569db5f83ed3b2ba2739f0501dbc318589. Bounded read-only disassemblies are retained with this run. These establish code paths, not that each path executed in the retained original.

Boundary Behavior established by disassembly
0x4bf1bb Tests/sets the frontend's one-shot +0xe4, calls 0x4be043, then constructs and installs the completion binding through 0x69b012. Repeated title input need not restart this path.
0x4be043 Requests clip 13 through 0x69b244, clears the existing binding through 0x69b027, and calls 0x69b615 with zero dt before the new binding is installed by its caller. This ordering alone does not prove an early-completion race.
0x69b244 / 0x69b3d1 Records requested clip/flags and a pending byte. Clip 13 has special request handling. Pending work is accepted only after readiness checks; 0x69b615 clears pending only on success.
0x69b615 Processes a pending clip, checks the resource/clip table and readiness, passes its dt through 0x6aaa55, and can call the separate update function 0x69deb8. It does not use +0x7e as a general time-advance gate.
0x69deb8 Tests other gates (+0x78, +0x7c, override/dirty/resource conditions), updates pose state, and sets +0x7e on changed paths.
0x69e7d0 / 0x69f1bf The pose body can skip when +0x7e=0. The common exit unconditionally executes and byte [ebx+0x7e],0; children can inherit the flag immediately before it is cleared. Zero is therefore an ordinary possible post-update value.
0x6a6a11 / 0x6a4bd7 Resource/phase routines advance time and produce status masks, including endpoint handling. They contain x87 compare/status/parity branches. Their invocation and inputs for this title animator were not observed.
0x69b048 Dispatches the binding at +0xf4 through its adapter pointer, passing binding address and payload. The static caller connecting natural clip completion to this dispatcher remains unidentified.
0x4f3cab / 0x4c0236 Adapter computes receiver plus adjustment and calls the member pointer. The installed frontend member requests state 5 through 0x521f84 with zero delay.

The resource manager pointer at animator +0xe8 is 147347508 in the retained snapshot. Its pointee, active phase/list, clip readiness result, status masks, and actual dt history were not retained. A nonpending clip request with clip number 13 does not supply those missing values. Neither a Win32 timer fault nor an x87 branch fault follows from these snapshots.

Generated CPU controls

generate-fixture.py authors a small assembly fixture from independent expected CPU semantics. No guest bytes are copied into it. The same linked code/data run as a native i386 Linux ELF and as a generated PE loaded by the emulator; check-fixture.cjs checks each against explicit expected condition bits and results. The native run is an independent CPU oracle, not a native Windows or original-game reference.

The 126 cases cover less/equal/greater/unordered, negative, signed-zero, and small finite values, with TEST AH masks 0x05, 0x41, and 0x44. They check x87 C0/C2/C3, parity/zero, actual JP/JNP/JNE branch results, and stack TOP after pop. Both intervening FSTP and contiguous FNSTSW AX; TEST AH; Jcc shapes are used. The generated indirect member-binding control independently checks nonzero receiver adjustment, payload, exactly one call, and restoration of ESP through the cdecl adapter/thiscall RET 4 combination.

All cases pass on native i386 and on both tested WASM binaries, with set_x87_pipeline4_fusion and set_x87_affine_fusion set to 0 and to 1 before decoding. These are toggle configurations, not a measured count of fused-handler entries. Results retain exact module/fixture/ELF hashes. The original's actual retained owning module is SHA-256 e7d6617cecdf3275db05454763c07cad7e6eb557afc2a14ca3413d960328fd31; its source/binary provenance remains the predecessor's pinned closure and observer-transform ancestry, not the current checkout's build. This run records the current checkout SHA, runtime file hashes, relevant dirty/new runtime patches, and generated input patches without archiving any source tree.

The focused existing test/test-x87-compare-eflags.js and test/test-x87-dc-fcom-alias.js also pass. The locked build and its final result are recorded in locked-build.log and validation.json. These controls exclude the tested instruction/calling-convention mismatches; they are not an old-implementation-failing regression and do not prove whole-emulator or game correctness. No repair is claimed.

Precise remaining evidence

A causal explanation needs an unmodified original's bounded owning-CPU history for this animator, starting before the one-shot title request: entry/return of 0x69b615 with actual dt, pending/readiness decisions, the resolved manager's phase/rate/status evolution, and the natural caller/event leading to 0x69b048. The binding installation and any dispatch must be ordered against those events, with adapter/member entry and resulting state request recorded. This separates no update, resource not ready, phase not reaching its endpoint, missing script dispatch, and incorrect CPU/callback execution. The exact first divergence then needs independent CPU/API semantics or a native reference before a repair.

Another EULA/clock-read run or uncorrelated callback-delivery test would not fill this gap. No fresh original run was spent without that observation plan. No runtime is assigned or left running by this task; predecessor box bx_ean6556p was already archived with verified cleanup. A transient shared-disk drop below 2 GB was reported on the messageboard; large allocation was avoided and the coordinator restored headroom before further work.