erik/acdream
1
0
Fork 0
Code Issues Pull requests Projects Releases 6 Packages Wiki Activity Actions

docs(overhaul): verify #476 capture boundaries against retail

Browse source
This commit is contained in:
Erik 2026-09-05 10:59:28 +02:00
parent 0057b97443
commit df4396ee89
1 changed files with 48 additions and 0 deletions
Download patch file Download diff file Expand all files Collapse all files

48
docs/research/2026-09-01-overhaul/476-capture-lifetime-diagnosis.md
View file

@ -88,3 +88,51 @@ No #476 implementation is authorized by this diagnostic note. The lead will
commit a bounded contract and dispatch one implementer after #475, with both
review lenses and the campaign's normal gate/ledger rules. C1c, the geometry
witness, full lanes, C2 and owner G4 remain outstanding; never merge main.
## Additional lead static checks while #475 tests are corrected
No #476 implementation was started. The lead read the complete real outer
frame transaction, private renderer, capture controller and GPU frame adapter.
The existing `IPrivateFrameScreenshot` is already a narrow capture seam;
moving its invocation does not require a second GPU-frame owner. Preserve
GPU measurement closure before GPU submission, and exactly one attempted
frame close on every successful Begin, including render/recovery/close errors.
The actual adapter clears CurrentFrame before calling its End.
Post-screenshot consumers inspected: RenderFrameDiagnosticsController is
CPU diagnostics; WorldLifecycleAutomationController reads borrowed world/
reveal/resource snapshots and writes checkpoint JSON; the shadow comparison
reads the completed CPU render oracle and clears its own dirty markers.
None of these bodies requires an open GPU command frame. Their mutual order
(capture, diagnostics, shadow comparison, lifecycle checkpoint) still matters.
Do not move world/private draw phases or introduce another frame tick.
Vulkan retention is explicitly opt-in. ConfigureBackbufferCapture creates
the buffer only when `_retainBackbufferCapture` is true; retain that default-
off cost. RecordBackbufferCapture records the copy while the swapchain image
is acquired, then QueueSubmit2 submits it. A validity marker must reflect an
actually submitted copy, not merely allocation or command recording; resize,
failed submission and skipped/zero-area frames must not label old/uninitialized
bytes as the newly requested complete frame. The bring-up caller already
reads after ending its frame.
Named retail's screenshot boundary was also checked statically against the
same paired executable SHA256006FFEADC5D679C871497112A5BD1F87714D0E273E2166BAE5052DDE369297B1:
| Function/window | Length | SHA256 |
|---|---:|---|
| Device::GetScreenshotSurface00439160, complete body |298|7717277646A9283ECBD10594177A5EC70213194E19B8F61BBC587C20BC3A7907|
| Device::SaveScreenshot0043A540, initial call window |32|00C51A0CD7CCA7D4C8075352BDA05683B196516604719243D01C349B4C408B4A|
| RenderDeviceD3D::GenerateSurfaceFromFrontBuffer005A0210, complete body |105|BA36CC608770994E60F9566C36D269160AE4936765CAD103DAF9DE061B7A4653|
Lead PE-section mapping gives raw offsets00039160/0003A540/001A0210.
The call at0043A553 is E808ECFFFF, resolving to00439160; the initial
GetScreenshotSurface vcall is offset3C (GenerateSurfaceFromFrontBuffer).
The D3D function's005A0256 call is FF9384000000, the named
IDirect3DDevice9 GetFrontBufferData vtable slot. The named body and bytes
agree on allocating a front-buffer-sized surface, then calling that capture.
GetScreenshotSurface crops to the client rectangle in windowed mode.
This is evidence for capturing completed front-buffer contents, NOT a claim
that Vulkan's acquired-image retained-copy technique is literal D3D retail.
The current PNG/location conventions and typed RHI lifetime remain modern
adaptations; no screenshot numbering, format or orientation change is proposed.