Aranna EULA/title diagnostic, 2026-11-07
The original registered demo passed EULA and rendered the title, but did not reach Main Menu or a map. The repaired run-local route is validated. The remaining causal boundary is the title animator's completion path to its installed state-5 callback. No emulator CPU, Win32, input or clock defect was demonstrated, so no emulator implementation was changed.
Evidence is retained in scratch/runs/20261107-aranna-eula-title-diagnostic/.
This follows aranna-title-transition-20261107.md and the earlier successful
aranna-acquire-before-arm notes. The earlier rejected guest-coordinate click
was a harness error, not evidence of an emulator defect.
Reproduction identity and controls
One fresh UUID Idempotency-Key created box bx_ean6556p at 05:08:07Z. Base source
commit: d14a8e882f9cee83408ba67587591f90bd921957;
runtime-dirty.patch and untracked-wininit.patch retain relevant dirty runtime
source. runtime-pins.json identifies 2,087 runtime/helper/media files by size and
SHA-256. Media is referenced by existing path and hash, not copied into evidence.
The retained qualified observer WASM SHA-256 was
e7d6617cecdf3275db05454763c07cad7e6eb557afc2a14ca3413d960328fd31;
the owning Worker module was inspected, rather than inferred from a staged file.
Puppeteer 25.7.0 and Chrome 151.0.7922.108 passed rendered smoke before the original.
browser-smoke.png was personally inspected. The predecessor used a different
Chrome patch release; these are this run's actual bytes and version.
Before the single heavy original launch, two generated-shell sessions used the actual retained browser driver, queue, completion protocol and observer:
- Positive route:
canvasclick:0.2481,0.7093was admitted and completed. Fractions mapped through canvas rectangle(10,194,661,568)to page(173.9941,596.8824). Guest button down/up counts were balanced (one each). The generated fixture's transient black response is not claimed as visible button activation. Quit completed, driver/Chrome exited 0, all ten tracked PIDs, both ports and the browser profile were absent. - Negative route:
canvasclick:2,2failed schema validation withError: queued fractionsbefore admission or effects. There was no action, admitted receipt or done receipt. Driver exited 1 and Chrome exited 0; all ten tracked PIDs, ports and profile were absent. No rejected/UNKNOWN input was replayed. - Missing generated fixture and optional API/TLB/icon inputs were added to the run-local pin closure before original launch. A premature capture read and a mistaken expectation that schema rejection produced an action receipt were corrected by inspecting authoritative output, without retrying input.
See control-route-approved.json, controls-approved.json, both generated
session directories, and their preserved remote command envelopes. The complete
2,087-file closure passed before and after the original.
Personally inspected original progress
The original used the corrected EULA fractions, mapping to page
(173.9941,586.9986) for its actual canvas geometry. It passed EULA, loading,
Microsoft/GPG/Mad Doc introductions, and reached the Aranna title over stone
doors (response-07.png). Ordinary balanced title Escape, Enter and center
canvas click completed. response-10.png still showed the title. Subsequent
passive observations response-11.png and response-12.png showed teal desktop;
the app remained reported running with no trap, focused, visible, and owning
relative pointer lock. This disappearance is observed, not diagnosed as a
renderer defect. No Main Menu or map capture exists for this run.
Two earlier review descriptions incorrectly named the intro images:
response-05.png is the scientist/Mad Doc image and response-06.png is the
Aranna title on black. review-corrections.jsonl records the corrections; these
were distinct completed actions, not replays. All final interpretation uses the
actual screenshots and matching state receipts.
Read-only causal finding
The prepared owning Worker diagnostic identifies slot 0, endpoint 1, target
EC5692A21A4A03EF62D15B3AE5977859. It reads guest state without guest writes,
state forcing, injected guest calls or guest-byte changes. The previous run
already proved owning Escape/Enter callbacks completed; that is not a new cause.
The new center-click evidence reaches frontend member 0x4bccac with receiver
150089148. The frontend's one-shot byte at +0xe4 is 1. Following the actual
handle graph resolves frontend handle 0x40090009, wrapper 162711416, object
167032348, animator handle 0x00120012, animator 167032084.
Two exact-owner snapshots, at 05:17:53.220Z and 05:20:32.154Z, retain the same:
- frontend state 4; immediate request 0, delayed request 0, deadline double 0;
- animator bytes
+0x7c=1,+0x7d=1,+0x7e=0,+0x78=0; - animator clip fields
+0xc4=13,+0xd4=13,+0xd8=0,+0xdc=1; - callback record at animator
+0xf4:[0x4c0236,0,150089148,0x4f3cab].
The guest clock advanced from 545.7833295758464 to 862.8066514279344 over those 158.934 wall seconds. It is advancing; the roughly 2x ratio alone does not prove an emulator clock mismatch. There is no pending delayed state request waiting on a deadline in these samples.
Read-only disassembly shows 0x4bf1bb starts the title animation and installs
that binding. Callback 0x4c0236 requests state 5 via 0x521f84 with zero delay.
The animation update at 0x69e7d0 skips its main body when byte +0x7e is zero.
This establishes a concrete next boundary: determine why this animator remains
in that condition and whether/where its completion callback is dispatched.
An installed callback and sampled flags do not prove it was never invoked, or
that the emulator executed an instruction incorrectly. A native/reference
comparison or bounded observation of that boundary is still needed before a
CPU/API repair. Repeating Win32 callback delivery alone will not resolve it.
The disassembly and both resolved snapshots are retained alongside the trace.
observation-summary.json counts 26,612 records, 24,812 broker completions, no
observer error/mismatch records. These are observer findings, not proof of
whole-program correctness or full callback coverage.
Validation and shutdown
The generated positive/negative routes, rendered smoke, full pre/post pins,
normal original quit, and evidence SHA-256 retrieval passed. Local
flock /tmp/wine-build.lock bash tools/build.sh passed; it is a checkout sanity
check, not a claim that its newly built module was used in the retained runtime.
No emulator fix or implementation regression test is claimed.
Original quit completed at 05:21:03.779Z. Driver/Chrome/supervisor exited 0.
Cleanup verified all eleven tracked original/supervisor PIDs absent, HTTP 44603
and CDP 39895 refused, profile absent, all four runtime locks free. Terminal
state records the owning Worker as retired-during-read with TargetCloseError
after quit; generated-shell terminal reads also lack live Worker state. Live
matching captures are retained without fabricating terminal CPU state.
172 files / 17,679,652 bytes were retrieved with matching SHA-256, plus the two
approval receipts. Large unique traces/logs were gzip-verified before retrieval;
no source/media archive was made. /home stayed above 2 GB free.
Explicitly delegated normal stop completed: box bx_ean6556p is ARCHIVED,
forced=false, stopped at 05:21:48.269Z, no remaining expiry. No test runtime was
left running. The run ends with a precise animation-completion blocker and no
unproved emulator modification.