Q3 original initialization: offline instruction diagnosis
Evidence: scratch/runs/20261110-q3-initialization-cause-0342/.
Result: unresolved original input/state dependency; no supported runtime
repair. The retained final EIP is an image preparation scan, and its exact
instructions terminate correctly in the canonical compiled interpreter. No
browser, box, native execution, new game startup, clock/tier/deadline change,
media copy, or source archive was used. Shared runtime files are unchanged.
The starting main commit is d3651d79085448bb73a4961c59cdc3f05d649f9a.
The actual retained EXE is
test/binaries/candidates/quake-3-arena-demo-installer/installed/quake3.exe,
SHA-256 2c0dd4d06e3282e0abc41ebc8e84e2d9868a17182ac175b232a9bb604a8269a3.
The tested canonical CPU SHA-256 is
4f2c69a1528e01a9ab5127201d5e8a222da2a35236377a1351f444f55fa5a7d9.
The prior native evidence is the 0328 run described in
the retained-producer report.
What the retained execution actually establishes
The 29 producer replies establish the correct WebGL owner/context but zero presentation/write sequence/draws/vertices as completed batches advance 2–139. The exhausted console64 lane ends at ARB multitexture; later text remains unknown. Re-reading the retained console and transport containers verifies their original SHA and size against the pre-stop retrieval inventory. The new analysis retains all 191 nonnull slice samples and deduplicates 378 surviving GL boundary records without filling overwritten gaps.
The six sampled EIPs in 0x456a7a..0x456af1 belong to a texture pixel scan.
The final three samples have unchanged ESP 0x074fabb4:
| Page time (ms) | Slice sequence | EIP |
|---|---|---|
| 49050.295 | 3218 | 0x456ae0 |
| 49300.985 | 3625 | 0x456a7a |
| 49550.295 | 4014 | 0x456a9a |
Before these, samples pass through file/CRC work, a mipmap filter at
0x4571d6, and other routines. The final interval advances 796 slice
sequences; these observations do not show an unchanged loop counter or
repeated image identity. Every sampled register payload is null. There is no
recorded trap, but this does not establish eventual completion.
Original instructions and caller/API dependencies
Disassembly locates the containing function at 0x456780 (texture upload
preparation, conventionally Upload32). Its static caller is the call at
0x4566cf, return 0x4566d4, inside image creation beginning 0x456590.
This is a static caller relationship; the actual retained stack return was
not captured. Image creation also has callers for built-in textures and
loaded images; the final image identity cannot be chosen from EIP alone.
After power-of-two/resampling work, the function clamps output dimensions at
0x456a2e..0x456a40 using dword 0x7cb1c0. Renderer initialization populates
that dword at 0x44a4d1 from glGetIntegerv(GL_MAX_TEXTURE_SIZE=0x0d33):
the indirect query is 0x44a4c5, pointer slot 0x7bdec0. The caller supplies
an output pointer into its stack and copies that result into the global;
nonpositive results become zero. The adapter's production _getState reads
backend.getParameter(0x0d33) and writes an Int32 through the guest address
translator. The partial boundary record for opcode103 proves a query entered,
but retains neither pname, destination contents, nor its returned limit.
The observed EIP is after this clamp and scratch allocation. At
0x456a53, ECX becomes ESI×EDI (source/resampled pixel dimensions, distinct
from the clamped output dimensions). The loop at 0x456a7a loads each RGB
byte with FILD, compares to three stack maxima with FCOM/FNSTSW/TEST AH,0x41,
updates maxima for greater values, and checks alpha against 255. The final
0x456ae9 increments EBP by four, 0x456aec decrements the remaining count,
and 0x456af1 branches back while nonzero. 0x456a9a is the green-channel
load, not an API wait or error handler. Later calls at 0x456bad and
0x456cef use indirect pointer 0x7bdf50 with the nine glTexImage2D
arguments. Six surviving opcode158 _glTexImage2D boundary entries show some
canonical texture uploads already entered; ring loss prevents a full history.
At the observed scan ESP, the exact missing inputs are:
| Location | Meaning |
|---|---|
| ESI / EDI | Width / height used to form scan count |
| EBP | Current pixel pointer plus two |
[ESP+0x24] |
Remaining pixels (after the first iteration starts) |
[ESP+0x2038] |
Caller return of upload function |
[ESP+0x203c] |
Source/resampled pixel pointer |
[ESP+0x2040], [ESP+0x2044] |
Width / height arguments |
[ESP+0x2048], [ESP+0x204c] |
Mipmap / picmip arguments |
[0x7cb1c0] |
Limit produced by original GL query |
The original scan input can change during resampling; the pointer at
ESP+0x203c reflects that work. None of these values or an image-name stack
chain is present in the retained register-free slice lane. The console cap
cannot recover them. Consequently, an oversized/corrupt original count,
repeated image creation, and finite legitimate preparation remain distinct
possibilities, not established causes.
Compiled causal controls and outcome
node tools/q3-initialization-cause.test.cjs scratch/runs/20261110-q3-initialization-cause-0342 loads the hash-verified Q3
PE into the real compiled CPU, extracts the original scan bytes
0x456a53..0x456af3, and runs them with controlled pixels/stack. Internal
branch offsets remain unchanged; only a terminal return pad is appended.
These are isolated test inputs, not writes to the original native guest.
Five controls cover 1×1, 17×13, 256×256, 1024×1024, and zero count, opaque
and nonopaque alpha, RGB maxima, x87 comparison/pop behavior, advancing pixel
pointer/decreasing remaining count, and eventual return. All pass. The
1024×1024 control consumes 839 calls of run(10000), with decreasing counts
at slices 1/400/800. Thus the original final 796-sequence scan interval is
compatible with a finite ordinary scan; this is not a recovered original
dimension or timing measurement.
Exact original clamp bytes also terminate for limits 2048/4096/zero. Zero halves dimensions down to zero; it does not make the clamp infinite. A deliberately injected negative limit reproduces nontermination after the dimensions reach zero, but original initialization normalizes nonpositive query results to zero. There is no evidence of an original negative value or later corruption, so this fault control supports no repair.
The production GL render adapter, command stream and bridge execute context creation/current binding and the integer query against real compiled memory with the canvas-free software consumer. It writes 2048 and reports no GL error. Additional production query-writer controls forward backend values 4096/8192/zero unchanged. Those backend answers are controlled; no native WebGL value or hardware correctness is claimed. No hardcoded software limit is substituted for the original WebGL limit.
Focused controls and flock /tmp/wine-build.lock bash tools/build.sh pass;
the canonical CPU hash remains unchanged. Only this test and report are
published. The 52b51b920 native cancellation admission repair and existing
producer selector are preserved without edits.
The concrete next step requires original state absent from this evidence: under a separately authorized original execution, replace an existing passive scan sample with a bounded read-only snapshot of the table above at two successive scan slices, plus the image-creation return/name chain. Keep the existing total observation bounds, console cap, selector and cancellation policy. Decreasing count with advancing EBP distinguishes finite work; repeated entry/name identifies repeated initialization; dimensions and the actual query result permit an authentic full Upload32 control. First try any already-retained memory snapshot if one becomes available. Do not infer a GL or CPU repair from EIP repetition or relax presentation readiness.
No task box exists, so there is no new expiry or stop receipt. The inherited
0328 box bx_2qbqdc6h remains archived, expiry null, receipt
stop_3bfd33cc00fe9668; it was untouched. This single-task worker releases
its paths and stops after publishing the unresolved instruction dependency.