Aranna: first descriptor cache initialization, 2026-11-07
The first runtime initialization is now observed globally, before selected title
identity. OnStartChore and OnUpdate set their initialization guards, look up
an empty descriptor registry, receive EAX=0, and store that zero. This explains
the later guarded zero caches. It does not establish why the registry is
empty or demonstrate an emulator CPU, Win32, or memory contract mismatch. No
production code changed; no repair or menu/world progress is claimed.
Evidence is in the ignored run
scratch/runs/20261107-aranna-first-descriptor-init/ (started 06:28 UTC).
This extends the wrapper-boundary investigation.
The unchanged demo was launched once in one fresh box. Guest disassembly and
reads were evidence only. No guest stores, state forcing, synthesized callbacks,
media/config changes, or second original run were used.
First writes and matched lookup returns
The initial PE and startup observations on CPU owners 0, 1, and 2 have all four words zero. Store-helper observation starts globally at runtime startup, independent of title selection. It records four before/after store pairs on owner 0, below its independent quota of 64. Ordering:
| UTC | Actual guest instruction | Address | Original operation | Cache words afterward |
|---|---|---|---|---|
| 06:37:08.127 | 6aaac8 |
7b3684 |
byte OR 1, old 0 | [0,1,0,0] |
| 06:37:08.128 | 6aaae4 |
7b3680 |
dword store EAX=0, old 0 | [0,1,0,0] |
| 06:37:08.129 | 6aaa6f |
7b368c |
byte OR 1, old 0 | [0,1,0,1] |
| 06:37:08.132 | 6aaa8b |
7b3688 |
dword store EAX=0, old 0 | [0,1,0,1] |
The observer reports translated block PCs 6aaae1/6aaa88 for the two A3
stores; the instruction addresses above come from the unchanged guest bytes.
Other owners subsequently observe the changed shared words. Selected identity
arrives only at 06:38:45.308–309, over 97 seconds after initialization.
At 6aaaae, the first wrapper has caller 69b4d4, resource value 80010001,
and argument 1. At 6aaa55, the next wrapper has caller 69b67b, the same
resource value, and dt=0. Resource values are not asserted to be pointers.
| Name bytes / length | Name address | 450cf7 result |
Lookup callsite / matching return |
|---|---|---|---|
OnStartChore / 12 |
7a2008 |
6aaadb |
6aaadc / 6aaae1 |
OnUpdate / 8 |
7a1ffc |
6aaa82 |
6aaa83 / 6aaa88 |
450cf7 returns a caller code address, not an interned string. Actual
name bytes and caller-code bytes are recorded separately (events 176 and 177).
At 450dd3, the three arguments are that caller address, name address, and
length. Both calls follow:
450dd3 -> 454333 -> 452c47 -> 452c5a -> 452ca9 -> 450dea -> wrapper return.
At 452c5a, EAX is the header 086c00bc; at 452ca9, 450dea, and each
matched wrapper return, EAX is zero. Both returns have matching stack/return
context. These are 14 steps against an independent per-owner quota of 128.
See first-init-events.json, first-init-analysis.json, and original
observations.jsonl; validate-final.py asserts this sequence and all stores.
Registry state and registration ordering
The descriptor registry manager loaded from [7ac9d4] is 086e0034.
Its map starts at manager+20 (hex). At both lookups:
- map+0 = 0; map+4 = header
086c00bc; - map+8 = 1, the mode byte, not the count;
- map+0c = 0, the count;
- header left/right = header, parent = nil
086c0078.
Static 455866 initialization uses this empty-header arrangement, with nil
loaded from [7acabc]. 4582ac tests parent against nil and returns the header
on an empty tree, without a name comparison. 452c47 recognizes that header
and returns zero. The observed return agrees with those bytes and table state;
a bad comparison instruction is not demonstrated. A separate asynchronous,
read-only sample at 06:39:51.529 (live-table.json) confirms nil and the same
empty map/cache state. It is not a registration-return trace.
Do not confuse this registry with selected animator 09f4a828, its animation
manager 08c85834, or selected resource 80080008. The first initialization
uses resource 80010001 before selection.
The independent generic registration quota of 512 fills at 06:36:42.290,
before first cache initialization. It captures one constructor post-call at
451705 (06:36:42.054, count 0), 256 post-insert observations at 452348,
and 255 temporary-string cleanup observations at 45235d. Captured post-insert
and cleanup map counts are 1. The EAX-pointed words at 452348 describe an
iterator result on the stack, not a descriptor return. Thus earlier insertions
are positively observed, and later emptiness is positively observed; this
bounded trace does not prove how entries disappeared or that every registration
failed. No target-name predicate event 187–189 was admitted. That scoped
absence is not evidence of a missing registration.
The next concrete boundary is registration 4522f8: capture actual inputs,
caller/code key, name/signature bytes, ECX, and matching results. In particular,
45c5d7 returns at 4523b6 (nonzero EAX branches to erase), and 4512aa
returns at 4523d5, with AL tested at 4523d7 (zero also branches to erase). Observe
452448 -> 454b58 -> 452457, success AL=1 at 45241a, and final return
452420. Use separate bounded quotas for kept/rejected registrations across
owners, not just an empty local-name predicate. This would locate the empty-table
cause before choosing any emulator regression or repair.
Admission, original presentation, and recovery
Before the original, generated fixtures passed against the retained baseline:
identical guest bytes, registers, and PC; first guard/cache stores; nested
CALL/JMP/RET lookup; cross-owner observation without selected identity; nested
reflective dispatch; actual RPC slots 0/1/7/15; shared immutable identity; and
main/thread rendering import signatures. The observer wraps original store
helpers and calls each exactly once; no guest stores or flag evaluations are
added. Per-owner quotas remain independent from selected-wrapper observations.
These are observer admission checks, not an old-behavior-rejecting emulator
regression. first-init-check.json, observer-check.json, dispatch-check.json,
broker-check.json, and host-check.json record PASS.
A rendered Chrome 151.0.7922.108 shell smoke was personally inspected before
and after one balanced Start-button mouse down/up; the menu opens. The one
original then received ordinary EULA acceptance, passive observations, and one
balanced Escape after inspection of title doors. Captures show loading, intro
orb, logo, and title doors. response-04.png still shows doors immediately
after Escape. The later passive response-05.png shows a teal shell while the
UI says one app is running, with no title, menu, or world. The cause of this
presentation change is undetermined. The preserved final quit review mistakenly
said title doors; capture-review-correction.json explicitly corrects it without
rewriting the original command/review. Final broker records: 24,812, zero drops
in its reported counters. Selected-scope observations retain 64 guarded zero
cache entries and 960 zero wrapper returns; those bounded counts are not a
claim about every callback or runtime fault.
Recovery pins:
- Checkout at capture:
a59ac1821753ae20e868ce21ba89ad77e10a3f56; relevant dirty hunks inruntime-dirty.patch, retained untracked recovery inuntracked-wininit.patch; Git remains the source archive. - CPU source commit
344c54a2cebb0af51858c80b78c18e665dc52080, closure SHA-256393c95db18c4eb54d055867a661e397d3786549ea341205b90931ae47e903674; runtime closure commitd14a8e882f9cee83408ba67587591f90bd921957. - Final observer module SHA-256
40f3dd9c246ee44ab05cc1c89e38a774280004ff26b16be47e517aeb3c048f18(2,210,686 bytes), matched to the original's retrieved instantiated module.prepare-observer.cjs, retained transforms,observer-standard.wasm.source.patch,owned-inputs.patch, andruntime-pins.jsonpreserve reconstruction. - Existing media
scratch/game-media/aranna-demo-20261105/extracted/DSLOA.exe, SHA-256fced0961231248c09c10d86ff51935569db5f83ed3b2ba2739f0501dbc318589. Expansion.dsres SHA-2563ed0688349f33b2dd7b9458c18cfbbc00490353c5888a93cd02c10da7f3c3b10. Expansion was recovered intact by 54 independently hashed 4 MiB parts, concatenated and whole-hash verified; no whole-file upload retry. Parts were removed. Reference existing media rather than archiving/copying it into runs. - All 2,147 pinned runtime files passed pre/post validation. 65 terminal outputs (27,506,210 bytes) and two root artifacts were retrieved with verified hashes. No source tree, scratch tree, or media archive was created.
Fresh UUID-idempotent box bx_jqt84p35 was created at 06:32:50.711, with
original expiry 07:02:50.726. Explicitly delegated normal shutdown followed
retrieval: browser/runtime and supervisor exit 0; all 11 recorded processes
absent, both ports closed, browser profile absent, four locks released. Stop
completed nonforced at 06:40:55.363; inspected box state is archived, stopped
06:40:55.346, no expiry. Owned local watcher/smoke processes and ports are clear;
free /home remains above 2 GiB. No production build was run for this docs-only
finding. local-evidence-validation.json, publication.json, and final
result.json record validation and publication; result.json is written last.