docs: contract #477 synchronization repair from native validation
This commit is contained in:
Erik
parent
8fd5e434b6
commit
29730506c5
6 changed files with 330 additions and 2 deletions
|
|
@ -94,6 +94,15 @@ No driver replacement, TDR-delay changes, skipped waits or forced recovery.
|
|||
Exact stack, hashes, local Windows report and limits:
|
||||
`research/2026-09-01-overhaul/475-owner-regate.md`.
|
||||
|
||||
One bounded core+sync validation run at8fd5e434b completed/exit0 but FAILS
|
||||
three proven dependency families: acquire-to-layout transition, sampleable
|
||||
MSAA depth/stencil resolve masks and terrain buffer grow/copy transfer reads.
|
||||
Ten capped reports per family; no driver reset reproduced. Lead verified
|
||||
production sites and Khronos rules. `research/2026-09-01-overhaul/477-validation-reproduction.md`
|
||||
records exact evidence; `477-gpu-synchronization-contract.md` binds the repair.
|
||||
Eliminating these hazards is necessary API correctness, not proof that the
|
||||
original AMD reset has only this cause.
|
||||
|
||||
## #476 — First post-resize graphical gate PNG is horizontally corrupted
|
||||
|
||||
**Status:** OPEN, observed2026-09-05 during #474 regression; capture-lifetime
|
||||
|
|
|
|||
|
|
@ -713,7 +713,7 @@ Update immediately when a slice changes state. Chat is not the ledger.
|
|||
| S5-#470 | **LANDED + REVIEW-CLOSED 2026-09-05; LEAD GRAPHICAL A/B PROVISIONAL PASS.** Campaign implementation stack `15a796c3a` -> `7506e5f14` -> `b333edb4f`; route `19b44e5e3`; reviewed scratch `51f974da4` -> `2cad9c84` -> `98c004aa7`; packet §§27–30. | G4 UNPASSED | Retail/behavior pass 2/10 and production pass 4/10 closed the two evidence-only findings. Fresh campaign Release 0W/0E and focused 48/48; exact scratch hermetic 16,976/16,976 and canonical InstalledDat 386/10/1, manifests 30/30. Pinned gate `logs/selfgate-20260905-032132-s5-470-pinned-shadow-ab-r1`: retail/off -> High -> retail/off, 3/3 PNGs, exit 0, graceful, no client; visual PASS provisional. The stale recipe-8 pre-route launch is excluded and recorded in §30. |
|
||||
| S5-#469 | **LANDED + REVIEW-CLOSED 2026-09-05; LEAD GRAPHICAL TRANSITION PROVISIONAL PASS.** Contract `809887524`; reviewed scratch `c09b6cf0f`; campaign implementation `94ddde69a`; route `62efc72cb`; packet §§31–32. | G4 UNPASSED | Retail/deviation pass 1/10 and production pass 2/10 both PASS with no finding. The two atmospheric vertex receivers keep retail's authored unnormalized `uLights` direction across every shadow gate; celestial direction remains for opt-in shadow/volumetric projection; IA-24 corrected in the same commit. Exact scratch: Release 0W/0E, focused 126/126, Vulkan 2/2, hermetic 16,983/16,983, InstalledDat 386/10/1. Fresh campaign Release 0W/0E, focused 82/82, Vulkan 2/2. Gate `logs/selfgate-20260905-040449-s5-469-lighting-transition-r1`: five PNGs, active 2,500-caster/four-cascade High rows, no near-black relight, exit 0/graceful, no client; visual PASS provisional. |
|
||||
| S5-c5 | **CLOSED + LANDED 2026-09-05; LEAD GRAPHICAL SMOKE PROVISIONAL PASS.** Contract `b77989c32`; campaign `bf53e2ad6` -> `e625dc4e6` -> `1b7ee4e58`; reviewed scratch tip `158656f0d`; packet §§33–35. | G4 UNPASSED | Retail/deletion pass 1/10 PASS. Production pass 2/10 found one static-field hole in the owner guard; fix round 1 reproduced both static-owner mutations, and production pass 3/10 PASS. Exact scratch hermetic 16,921/16,921, canonical InstalledDat 368/9/1, manifests 30/30. Fresh campaign Release 0W/0E, App 146/146, Core 8/8. A stale recipe-8 preflight is excluded; a non-destructive recipe-10 bake produced 2,237,865 keys with zero failures. Corrected gate `logs/selfgate-20260905-052130-s5-c5-landed-v10`: five PNGs, exit 0/graceful, no fatal/deleted-prefix match, no client left; geometry matches S5-c4, visual PASS provisional. |
|
||||
| S5 | **#473 EXTERIORS OWNER-ACCEPTED; #474/#475 LANDED + REVIEW-CLOSED; recall timing narrowly OWNER-ACCEPTED. Latest owner gate FAIL on #477–#481 (2026-09-05).** #475 `de427d2c0`; c1–c5 plus #470/#469 landed. | G4 FAIL / UNPASSED | Packet §§7 and12–49 bind. #474 `220bda797` repaired the publication/placement ACK cycle. #475 exact eleven-file return `ab989d717` is blob-identical to integration `de427d2c0`, including architecture/AD-2; one test-only fix round, both review lenses closed. Lead independent Release0W0E, hermeticRuntime1891/1891, Headless48/48, App297/297; evidence `475-lead-verification.md`. Fresh campaign Release0W0E; owner-requested clean-binary run PID26920 completed five reveal generations but CRASHED with ErrorDeviceLost/Windows LiveKernelEvent141 while owner jumped in Neftet (#477). Network logout cleanup is NOT successful process exit. Other owner findings: dark Town Network/dungeons (#478), hanging Rynthid partly shifted (#479), invisible wielded items (#480), initial low FPS (#481). Recall timing seems OK per owner, not an exact benchmark. `475-owner-regate.md` records hashes/logs/events and limitations; no PNG this round. Owner AFK authorizes autonomous continuation. First priority crash diagnosis; four-agent maximum, OpenAI only, no client/build overlap. Both clients absent, ACE UDP9000 PID13340 up at11:09+02, recheck before action. No current root-cause or repair claim for #477–#481. Still owed: exact nine-hop timing repeat, separate confirmed #476 capture-before-submit repair, preserved C1c nine-file test/doc landing (both reviews closed; lead68/68), §45 real geometry/membership witness, full canonical lanes/C2 and owner G4. Never merge main before G4. |
|
||||
| S5 | **#473 EXTERIORS OWNER-ACCEPTED; #474/#475 LANDED + REVIEW-CLOSED; recall timing narrowly OWNER-ACCEPTED. Latest owner gate FAIL on #477–#481 (2026-09-05).** #475 `de427d2c0`; c1–c5 plus #470/#469 landed. | G4 FAIL / UNPASSED | Packet §§7 and12–49 bind. #474 `220bda797` repaired the publication/placement ACK cycle. #475 exact eleven-file return `ab989d717` is blob-identical to integration `de427d2c0`, including architecture/AD-2; one test-only fix round, both review lenses closed. Lead independent Release0W0E, hermeticRuntime1891/1891, Headless48/48, App297/297; evidence `475-lead-verification.md`. Fresh campaign Release0W0E; owner-requested clean-binary run PID26920 completed five reveal generations but CRASHED with ErrorDeviceLost/Windows LiveKernelEvent141 while owner jumped in Neftet (#477). Network logout cleanup is NOT successful process exit. Other owner findings: dark Town Network/dungeons (#478), hanging Rynthid partly shifted (#479), invisible wielded items (#480), initial low FPS (#481). Recall timing seems OK per owner, not an exact benchmark. `475-owner-regate.md` records hashes/logs/events and limitations; no PNG this round. Owner AFK authorizes autonomous continuation. First priority crash diagnosis; four-agent maximum, OpenAI only, no client/build overlap. Both clients absent, ACE UDP9000 PID13340 up at11:09+02, recheck before action. One core+sync validation run at8fd5e434b exited0 but reported three proven GPU dependency faults (30 capped messages): acquire transition, depth/stencil resolve and terrain migration. `477-validation-reproduction.md` records hashes/PNG FAIL/partial PASS; `477-gpu-synchronization-contract.md` binds the next repair, not yet implemented. Original AMD reset attribution remains unproven. #480 cold-cache publication omission is statically identified; #478 needs High/Classic comparison; notes `478-480-readonly-diagnosis.md`. Still owed: exact nine-hop timing repeat, separate confirmed #476 capture-before-submit repair, preserved C1c nine-file test/doc landing (both reviews closed; lead68/68), §45 real geometry/membership witness, full canonical lanes/C2 and owner G4. Never merge main before G4. |
|
||||
|
||||
---
|
||||
|
||||
|
|
|
|||
|
|
@ -0,0 +1,143 @@
|
|||
# #477 — repair three observed GPU synchronization dependencies
|
||||
|
||||
Lead contract2026-09-05. **NOT IMPLEMENTED.** Base is the commit containing
|
||||
this contract, on the campaign's reviewed renderer stack after `de427d2c0`.
|
||||
One OpenAI implementer, then lead verification and sequential API/behavior
|
||||
and production/lifetime lenses. No more than the owner's ten review passes;
|
||||
each code finding gets a bounded fix round. No main merge before G4.
|
||||
|
||||
## 1. Objective and evidence boundary
|
||||
|
||||
Eliminate the three native synchronization hazards in
|
||||
`477-validation-reproduction.md` without changing scene membership, draw
|
||||
order, shaders, quality, streaming budgets, frame pacing or resource lifetime.
|
||||
These are demonstrated Vulkan API defects. Their role in the owner's AMD
|
||||
watchdog/device loss is not yet proven; do not promise this contract alone
|
||||
fixes the reset. #478–#481, #476 screenshot lifetime and the exact #475 timing
|
||||
route stay separate. In particular, do not hide the first-frame malformed
|
||||
image by dropping it or adding a delay.
|
||||
|
||||
This chunk adds no AC-specific algorithm or new claimed retail behavior.
|
||||
Retail scene/walk/alpha ordering and all existing register deviations remain
|
||||
unchanged. The following primary API sources govern these native dependencies:
|
||||
|
||||
- [Khronos render-pass specification](https://docs.vulkan.org/spec/latest/chapters/renderpass.html):
|
||||
fixed-function resolves use color-output stages/access masks for both color
|
||||
and depth/stencil attachments.
|
||||
- [Khronos submit reference](https://docs.vulkan.org/refpages/latest/refpages/source/VkSubmitInfo.html):
|
||||
acquired-image layout transitions must chain to the semaphore wait stage.
|
||||
- [Khronos synchronization specification](https://docs.vulkan.org/spec/latest/chapters/synchronization.html):
|
||||
dependent memory accesses and layout transitions need execution and memory
|
||||
dependencies. In-queue order alone is not a transfer-write/read barrier.
|
||||
|
||||
Lead read these primary sources and the production native structures before
|
||||
contracting. No new named-retail/paired-binary claim is inferred from Vulkan.
|
||||
Any implementation deviation must be reported and registered in its landing
|
||||
commit; expected new intentional retail deviations: none.
|
||||
|
||||
## 2. Bounded scope
|
||||
|
||||
Required production sites:
|
||||
|
||||
- `src/AcDream.App/Rendering/Gpu/Vk/VulkanGpuDevice.Resources.cs`:
|
||||
TransitionBackbufferForRendering, depth-resolve entry and sampling exit.
|
||||
- `src/AcDream.App/Rendering/Gpu/Vk/VulkanUploadQueue.cs`:
|
||||
native device-buffer migration dependency at queue recording.
|
||||
- `VulkanGpuDevice.cs` only if needed to share the already-existing acquire
|
||||
wait-stage value with the actual barrier; no frame-lifetime or error-policy
|
||||
redesign. Passive native-barrier construction helpers are permitted if
|
||||
directly consumed by production and needed for genuine tests.
|
||||
- Focused App Vulkan tests under `tests/AcDream.App.Tests/Rendering/Gpu/Vk/`;
|
||||
reuse current test infrastructure. Relevant architecture paragraph/comments.
|
||||
|
||||
No edits to Runtime, terrain allocation policy, shader/SPIR-V, capture path,
|
||||
quality/settings, package/DAT format or other worktrees. No second queue owner,
|
||||
general recorder framework or synthetic model pretending to be production.
|
||||
Do not introduce a device-idle wait, change semaphore/timeline ownership, or
|
||||
replace narrow dependencies with new ALL_COMMANDS/ALL_MEMORY barriers.
|
||||
Existing unrelated masks are not a license for a broad synchronization rewrite.
|
||||
|
||||
## 3. Required behavior
|
||||
|
||||
### A. Acquired-image execution dependency
|
||||
|
||||
The first swapchain transition currently sources TopOfPipe, while the submit
|
||||
wait uses ColorAttachmentOutput. Chain the transition to that real wait.
|
||||
Preserve first-use SrcAccess=None, Undefined->ColorAttachmentOptimal,
|
||||
queue-family-ignored/color subresource, and later-pass read/write content
|
||||
preservation. The same production helper serves ordinary world and filmic
|
||||
presentation; both remain supported. Do not change the presentable image's
|
||||
ownership, acquire timeout, pacing or first-use discard semantics.
|
||||
|
||||
### B. Sampleable MSAA depth/stencil resolve dependency
|
||||
|
||||
Use ColorAttachmentOutput/ColorAttachmentWrite for the single-sample depth
|
||||
resolve destination's entry dependency and its sampling-exit source.
|
||||
Retain Early|LateFragmentTests/DepthStencilAttachmentWrite for ordinary
|
||||
non-resolved depth. Cover 1x sampleable depth, multisample non-sampleable
|
||||
depth and 4x sampleable resolved depth. Keep image identity, both aspects,
|
||||
layouts, SampleZero resolve mode, sample counts and fragment sampling masks
|
||||
unchanged. Do not treat the multisample attachment and resolve result as the
|
||||
same writer just because both carry a depth format.
|
||||
|
||||
### C. Device-buffer grow/copy dependency
|
||||
|
||||
Before a migration copy reads device-buffer data, make preceding transfer
|
||||
writes visible to that transfer read. This must work when upload and growth
|
||||
are in the SAME drain, A->B->C migration chains are in one drain, and the
|
||||
producer was recorded in an earlier drain/frame on the same queue.
|
||||
Adding TransferRead only to the trailing consumer barrier is insufficient.
|
||||
Retain exact ordered copy ranges and the existing draw/shader visibility
|
||||
barrier, stage-buffer lifetime, completion-serial retirement and queue clear.
|
||||
Distinguish device migration from independent host-staged writes so the fix
|
||||
does not put a new global barrier before every ordinary staging copy.
|
||||
No unbounded dependency/history map or per-frame duplicate copy ownership.
|
||||
|
||||
## 4. Proof and mutations
|
||||
|
||||
Tests must inspect native structures produced by the actual production
|
||||
construction/recording path. If passive helpers are extracted, bind their
|
||||
real call sites and command order using the project's existing source/IL
|
||||
guards plus effect-discriminating mutations; testing an unused helper is not
|
||||
proof. Prefer a narrow production seam to a new test GPU framework.
|
||||
|
||||
- Acquire: native barrier and actual semaphore-wait stage agree; first and
|
||||
subsequent pass masks/layouts/access preserved; ordinary/filmic callers.
|
||||
- Depth: distinguish the three configurations above; assert entry AND exit,
|
||||
exact image/aspects/layout, ordinary depth masks unchanged.
|
||||
- Copy: stage->migration, A->B->C and across-drain/frame producers; barrier
|
||||
must precede dependent read, with exact source buffer/range where scoped.
|
||||
Independent staging stays batched and the original draw visibility remains.
|
||||
Use real production queue recording or a narrowly extracted recording body
|
||||
that production actually invokes, not a separately reimplemented planner.
|
||||
- Existing frame-flight, memory/ring, swapchain, world-pass and Vulkan rendering
|
||||
tests remain green; zero new skips or changed tolerance/quality expectations.
|
||||
|
||||
Perform four separate actual-production mutations: restore old acquire source
|
||||
stage; restore old resolve entry masks; restore old resolve exit masks; remove
|
||||
the pre-migration dependency. Each must fail its intended assertion, not build
|
||||
or setup. Restore exact raw source hashes after each cycle and rebuild green.
|
||||
Return commands, TRX/output paths and hashes. Lead repeats discriminating
|
||||
checks before accepting, then the two independent review lenses.
|
||||
|
||||
## 5. Execution and graphical closure
|
||||
|
||||
Use a fresh exact-base scratch on a `codex/` branch. Keep the existing #475,
|
||||
#474 and C1c scratches untouched. Only the lead commits/integrates. Before
|
||||
EVERY build/test or launch, check AcDream.App and retail absence; no tests
|
||||
beside a running graphical client, even in another worktree. No agent launch
|
||||
or debugger attachment; lead owns one graphical process at a time.
|
||||
|
||||
Implementer returns Release 0W0E, new/affected tests, exact source list and
|
||||
production hashes, four mutation results with restoration, and a compact
|
||||
report. No full-solution PASS is claimed from a focused lane. No new test
|
||||
framework, ignored findings or undocumented scope expansion.
|
||||
|
||||
After lead verification/reviews and integration: fresh Release build, ACE9000
|
||||
up, neither client running, repeat the committed bounded Neftet route with
|
||||
core+sync validation (`VK_LAYER_VALIDATE_SYNC=1`), same quality/resolution.
|
||||
All three hazard families must disappear; inspect any new message instead of
|
||||
filtering it. Preserve every PNG and report PASS/FAIL with paths, provisional
|
||||
until owner inspection. A clean validation run is API evidence, not proof
|
||||
that the original AMD TDR cannot recur. Retain later ordinary (uninstrumented)
|
||||
Neftet movement/portal/performance and full owner-G4 obligations.
|
||||
|
|
@ -1,6 +1,7 @@
|
|||
# #477 — one bounded GPU validation reproduction
|
||||
|
||||
Prepared2026-09-05; **NOT RUN YET.** No production behavior change. This is a
|
||||
2026-09-05; **RUN COMPLETE: synchronization FAIL, no crash reproduced.**
|
||||
No production behavior change. This is a
|
||||
diagnostic recipe under the owner's AFK continuation authorization, not a
|
||||
repair contract, new review round, timing benchmark or G4 PASS.
|
||||
|
||||
|
|
@ -62,3 +63,71 @@ run does not close #477; preserve the first-failure propagation defect and
|
|||
choose the next narrow diagnostic based on evidence. Do not enter an
|
||||
unbounded sequence of GPU-reset reproductions. Other owner findings and the
|
||||
remaining campaign gates stay open.
|
||||
|
||||
## Run r1 results
|
||||
|
||||
Clean source `8fd5e434b68dd808ac45d16f73b612b2bef7978d`; production files
|
||||
unchanged from reviewed `de427d2c0`. Fresh Release 0W0E/exit0,21.06s.
|
||||
App.exe SHA256 `E37EC4579456295231440CA209F9EFA42074E435D73F071F013DEC8B5624F4C9`;
|
||||
App.dll `E3D29F7C7D547997AEED03AF4C5DDD307CBF30460E43C6FFA4184FFBFFDE6F23`.
|
||||
Recipe10 package unchanged. Exact command/environment above; actual High
|
||||
near4/far25/MSAA4/A2Ctrue/aniso16/completions4 confirmed by startup output.
|
||||
|
||||
Run: `logs/selfgate-20260905-111914-477-neftet-validation-r1/`, PID12428,
|
||||
started09:19:14.370058Z, final checkpoint09:20:15Z. Runner and client exit0,
|
||||
route complete, graceful logout requested/confirmed, no unhandled/device-loss
|
||||
or incomplete-shutdown match. Both clients absent on privileged post-check;
|
||||
ACE9000 PID13340 up. No debugger attach, concurrent build/test or second run.
|
||||
The layer explicitly reports the deprecated synchronization flag and its
|
||||
replacement `VK_LAYER_VALIDATE_SYNC=1`: validation really loaded. Use the
|
||||
replacement on future runs. No global settings were changed.
|
||||
|
||||
Thirty reported errors, ten per message ID (the layer caps repeats):
|
||||
|
||||
| Hazard | Actual operation | Confirmed production defect |
|
||||
|---|---|---|
|
||||
| WRITE_AFTER_READ | First swapchain-image layout transition vs acquire; vk-world and atmospheric-filmic | `TransitionBackbufferForRendering` uses TopOfPipe source while submit waits acquire at ColorAttachmentOutput |
|
||||
| WRITE_AFTER_WRITE | Atmospheric sampleable MSAA depth/stencil resolve entry and exit transition | Both barriers use ordinary depth-test writer masks, but the fixed-function resolve writes at ColorAttachmentOutput/ColorAttachmentWrite |
|
||||
| READ_AFTER_WRITE | Terrain vertex/index buffer migration reads previously copied data | Upload drain lacks transfer-write to transfer-read dependency before a migration copy, including same-drain growth |
|
||||
|
||||
Lead read the native error messages and independently checked the named
|
||||
production barriers, submit wait and upload-copy loop. The source diagnosis
|
||||
matches the messages. Thirty messages are not thirty distinct bugs and
|
||||
repeat suppression means this is not an exhaustive occurrence count.
|
||||
No evidence attributes the owner's AMD reset to one particular hazard yet.
|
||||
|
||||
All five checkpoints exist. Final reveal generation2/destination8763000E
|
||||
is ready/materialized/complete/visible, failures0, waitCueShown=false;
|
||||
all transit ownership counters0. Twelve jump press/release pairs ran and
|
||||
stamina fell, but no explicit twelve-airborne-transition assertion was
|
||||
captured. Do not equate injected inputs with twelve successful jumps.
|
||||
Validation frame times are not a performance gate.
|
||||
|
||||
### PNG inspection (lead; provisional)
|
||||
|
||||
All three images opened by the lead,1600x900, under `artifacts/screenshots/`:
|
||||
|
||||
- `477-00-first-resize.png`: **FAIL visual evidence**; radial stretched
|
||||
geometry fills the view, not a valid settled scene. Retained, not discarded.
|
||||
This run started at1600x900 already, so the route's same-size resize does
|
||||
not reproduce #476's extent-change allocation. Cause not established here.
|
||||
- `477-01-neftet-before.png`: coherent Neftet scene, but portal-space wait
|
||||
text still visible; not a settled visual PASS or equipment PASS.
|
||||
- `477-03-neftet-after.png`: coherent settled scene and wait text gone,
|
||||
**provisional PASS for this narrow settled-cue check only**. Hands remain
|
||||
empty; no claim that #480 or any overall renderer gate passed.
|
||||
|
||||
SHA256:
|
||||
|
||||
| File | SHA256 |
|
||||
|---|---|
|
||||
| client.log |1387BB25509264C22353739FAE4D61344082394EA07B1FF0C137BC9573D18380|
|
||||
| client.err.log |E66FB4C4E8B767F4AD7064D35A3DB33722D658FB7B570F329655ED6BC93CD880|
|
||||
| 477-00-first-resize.png |85B769BF1C47781667E651205E1AE02087F05D653F86C781A9A7F1294655F4CC|
|
||||
| 477-01-neftet-before.png |6ECC46345293BCADAC553FDC4BC9DC953A896830EA3F2552233BC6A9A8C8034A|
|
||||
| 477-03-neftet-after.png |9A0072B108B9180727A89A25CA31F8E2563E0D0D54FCE8B3734F50EBC7115614|
|
||||
|
||||
Next bounded code contract is `477-gpu-synchronization-contract.md` for
|
||||
these three proven API dependencies. First-failure propagation remains a
|
||||
known diagnostic limitation; do not combine unrelated recovery work into
|
||||
this repair or call a no-crash validation run a resolution of the AMD incident.
|
||||
|
|
|
|||
|
|
@ -0,0 +1,85 @@
|
|||
# Owner gate follow-up — lighting and equipped-item read-only diagnosis
|
||||
|
||||
2026-09-05, source `de427d2c0` / documentation HEAD `8fd5e434b`.
|
||||
No implementation, test execution or visual acceptance in this note.
|
||||
#477's proven GPU synchronization faults are the immediate repair priority.
|
||||
#479 object identity/transform and #481 entry performance remain unmeasured.
|
||||
|
||||
## #480 — missing publication before equipped render membership
|
||||
|
||||
James's bounded read-only trace identified an absent production edge; lead
|
||||
independently read the controller, ordinary materializer, prepared publisher,
|
||||
registry and walk consumer to confirm the chain:
|
||||
|
||||
1. EquippedChildRenderController resolves the actual GfxObj from DAT for its
|
||||
template/availability, then BuildChildRenderParts (1202–1214) relies only
|
||||
on PhysicsDataCache.GetGfxObj/GetVisualBounds.
|
||||
2. Ordinary DatLiveEntityProjectionMaterializer explicitly calls
|
||||
LiveCollisionAssetPublisher.CacheGfxObj before membership construction.
|
||||
The attached path does not. Attached Position-null hydration bypasses the
|
||||
ordinary materializer; the cache getters do not lazily load the package.
|
||||
3. ShadowShapeBuilder.FromSetupRenderParts skips a visual part when cached
|
||||
physics and visual bounds are both absent. This can return an empty list
|
||||
despite DAT-backed drawable mesh/template data existing.
|
||||
4. AttachChild (controller693–698) accepts that product, and the controller
|
||||
logs successful attachment regardless of the ignored Boolean result.
|
||||
Even when the root CELLARRAY is inherited successfully, an empty part
|
||||
array emits zero RetailPartEntry rows.
|
||||
5. WalkProductionWorldData.ResolveCellView (286 onward) reads only those
|
||||
rows, so that child contributes no draw or downstream selection part.
|
||||
Parent-cell fallback would conceal, not repair, the missing producer.
|
||||
|
||||
The owner log has local-player attachment messages (lines212,882,905), but
|
||||
does not expose the current GfxObj cache state. Thus the reachable production
|
||||
defect is established statically; its occurrence for each owner item remains
|
||||
to be demonstrated by a genuine composed regression or live observation.
|
||||
The proposed repair is to reuse the existing strict prepared-asset publisher
|
||||
before constructing child render parts, never invent bounds/collision or a
|
||||
second cache. Composition already owns that publisher.
|
||||
|
||||
Proposed discriminating test: extend the actual controller OnSpawn fixture
|
||||
with a nonempty child Setup/GfxObj and cold production-style cache, backed by
|
||||
prepared collision content. Give the parent a real retail CELLARRAY. Assert
|
||||
attachment success as control, then exact child local-ID/part/GfxObj rows in
|
||||
the registry and the same child's EquippedChild record through the walk.
|
||||
Current code should fail at the first row assertion. Test nonempty fixture,
|
||||
effective part replacements and cleanup; do not manually inject membership.
|
||||
Existing controller fixtures use empty Setups; registry fixtures inject
|
||||
handmade child shapes, and selection tests inject visible parts directly.
|
||||
|
||||
Agent named-retail pointers (lead binary re-verification still owed before
|
||||
any AC-specific repair contract): CPhysicsObj::add_shadows_to_cells recursively
|
||||
visits children; CPartArray::AddPartsShadow visits every non-null visual part.
|
||||
Named pseudo-C lines282819–282901 and285933–285955. No new retail claim is
|
||||
accepted merely from a successful attachment log.
|
||||
|
||||
## #478 — distinguish local-light selection from High postprocessing
|
||||
|
||||
Rawls's bounded read-only report supplied named-retail/paired-executable
|
||||
checks; the lead has not yet independently repeated its complete retail
|
||||
proof. Treat the following as investigation targets, not an authorized fix:
|
||||
|
||||
- Current sealed-interior ambient .2 and no sunlight appear to match
|
||||
CellManager::ChangePosition; WorldRenderFrameBuilder593–612. Do not raise
|
||||
global ambient to compensate for another error.
|
||||
- LightManager590–645 fills one eight-slot EnvCell set with dynamic lights
|
||||
before statics. Existing AP-16/AP-35/AP-85 record the divergence from
|
||||
retail's distinct dynamic hardware lights and all-static vertex-lighting
|
||||
product. Enough dynamics can exclude every static fixture; the owner log
|
||||
lacks the current selected set, so this cause is not yet proven in-frame.
|
||||
- High's AtmosphericPostProcessGraph995–1053 applies filmic/exposure/vignette
|
||||
indoors. IsOutdoor gates rays/volumetrics, not the final filmic pass. This
|
||||
may amplify a low local-light contribution. Compare identical frozen pose
|
||||
in High versus Classic before changing a lighting formula.
|
||||
- Existing meshless Setup-light publication appears intact. Town Network's
|
||||
Setup02000365 carries a warm intensity100 light; actual selected membership
|
||||
must be observed. #469's directional correction does not cover this path:
|
||||
EnvCells use LightingMode1 and do not receive the outdoor directional shadow.
|
||||
|
||||
Next bounded evidence after #477: stable00070143 High/Classic image pair plus
|
||||
actual registered dynamic/static counts, snapshot membership and the eight
|
||||
selected cell indices/contributions. Reuse current diagnostics/test seams;
|
||||
do not create a global telemetry framework or resurrect camera-flood light
|
||||
selection. No current PNG/light-membership witness exists for the failed
|
||||
owner session. If a genuinely required retail fact cannot be recovered
|
||||
statically, follow the owner's stop rule; do not invent it.
|
||||
|
|
@ -4644,3 +4644,25 @@ silently dropped, but the next graphical run must serve the confirmed crash
|
|||
investigation. Both clients absent; ACE UDP9000 PID13340 remains up. Check
|
||||
again before builds/launches; one run per background command, close only our
|
||||
own gracefully. G4 FAIL; no merge main.
|
||||
|
||||
### #477 bounded validation result and next contract
|
||||
|
||||
Clean8fd5e434b, same productionde427d2c0, fresh Release0W0E. One run
|
||||
`logs/selfgate-20260905-111914-477-neftet-validation-r1`, PID12428, core+sync
|
||||
validation, High/near4/far25/MSAA4/A2Ctrue,1600x900. Route completed and
|
||||
client/runner exited0, no device loss reproduced; both clients absent afterward.
|
||||
Validation FAIL: three confirmed native dependency faults, ten capped messages
|
||||
each—swapchain acquire transition, sampleable MSAA depth/stencil resolve,
|
||||
terrain migration transfer read-after-write. Lead verified source and primary
|
||||
Khronos rules; this is actionable API evidence, not proof of the AMD TDR's
|
||||
specific cause. Exact hashes/PNG outcomes: `477-validation-reproduction.md`.
|
||||
First capture is malformed (retained FAIL); later settled Neftet cue clears
|
||||
(narrow provisional PASS), not an overall renderer/equipment acceptance.
|
||||
|
||||
`477-gpu-synchronization-contract.md` binds only these three dependencies,
|
||||
native-structure/recording proofs, four actual-production mutations and the
|
||||
same validation repeat after review/integration. No new retail algorithm or
|
||||
intentional deviation, no frame-idle workaround, no screenshot/lighting/items
|
||||
smuggled into the chunk. Other read-only findings and next proof targets are
|
||||
in `478-480-readonly-diagnosis.md`; #480 has a concrete missing prepared-cache
|
||||
publication edge, while #478 needs an identical-pose High/Classic witness.
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue