fix(overhaul): integrate completed-frame capture and host visibility

docs/architecture/acdream-architecture.md

@@ -1092,8 +1092,18 @@ the exact retail `Client::UseTime` order. See the completed Slice 6 ledger and
 `RenderFrameOrchestrator`. Its frozen graph begins per-resource frames and
 render-thread uploads; draws world/PView and its two shared-alpha scopes;
 renders portal and paperdoll private viewports, retained gameplay UI,
-devtools, and screenshots; then closes the GPU-flight transaction in
-`finally`. There is one UI stack: the retained tree.
+and devtools; closes GPU measurement and exactly one GPU-flight transaction;
+then captures the successfully submitted complete frame before CPU diagnostics,
+shadow comparison, and lifecycle checkpoint publication. Zero-area and failed-
+close frames do not consume queued captures. Vulkan retained-copy validity is
+earned only by a successful submission containing that frame's copy, never by
+allocation, resize, or command recording alone. The retained-copy command emits
+an exact-range COPY/TRANSFER_WRITE to HOST/HOST_READ buffer dependency; capture
+uses the existing device-idle completion wait before reading. This opt-in path
+requires the actual capture allocation to be host-coherent and rejects an
+unsupported noncoherent allocation before wait/read (AD-121); it does not change
+shared heap selection or generic buffer invalidation. There is one UI stack:
+the retained tree.
 
 ---
 

docs/plans/2026-09-01-campaign-overhaul-world-solidity.md

@@ -759,6 +759,13 @@ are deferred, not passed. Continue crash/stability, dungeon lighting, object
 placement and hand-item verification, with #476 capture as their immediate
 prerequisite. C1a remains correctness-only. No main merge or G4 acceptance.
 
+**#476 current (packet§55):** reviewed scratch02b707133 integrates with this
+ledger, both lenses closed including narrow F1 re-review3. Root Release0W0E,
+affected179/179, portableGPU307/307, two native mutations independently fail/
+restore. #477/#480 preserved; AD-121 added with AD-118 still retired and92active
+AD rows. Fresh integrated Release/tests and first-current-frame/resize/hand
+graphics next, not yet PASS. No FPS work or main merge.
+
 ## 14. Risk register (condensed)
 
 - **R1 second scene graph.** Graph records own only topology/membership facts;

docs/research/2026-09-01-overhaul/476-lead-verification.md

@@ -81,3 +81,55 @@ capture evidence. Both toggles use the real input action; no attack or inventory
 mutation is requested. Gate after accepted integration and green Release, with
 both clients absent, ACE up, agent builds/tests paused, one runner invocation,
 same session-v10/High profile, core+sync validation retained to guard #477 merge.
+
+## F1 lead completion and independent verification
+
+James froze the two F1 source/test edits before compiling or running them;
+no implementer F1 gate is claimed. Root's first Release build found one test
+handle type error CS1503 (uint passed to Device's nint constructor). Root
+corrected only that cast, then six capture tests passed. Root completed the
+bounded architecture/AD-121 documentation; no other source behavior changed.
+
+Root independently read the actual native copy/barrier/read methods and all
+four added tests. RecordBackbufferCapture emits an image transition, real
+copy and exact-buffer/range COPY/TRANSFER_WRITE→HOST/HOST_READ dependency in
+that order. Copy-off/mismatch emits none. Actual CaptureBackbuffer reads and
+swizzles distinct coherent mapped bytes after one native idle wait; its
+noncoherent allocation is rejected before wait/read. These are intercepted
+native-call witnesses, not physical GPU results.
+
+Two separate root production mutations, with exact raw restoration:
+
+- Omit the real host-read barrier call: actual native command collection
+  fails expected3/actual2 (image barrier and copy only).
+- Invert the actual coherence guard: the noncoherent witness fails because
+  the expected NotSupportedException is not thrown.
+
+After restoration full Release0W0E4.98s; the exact twelve prior affected
+classes give179PASS/0FAIL/0SKIP (175 plus four native witnesses). Portable GPU
+filter `FullyQualifiedName~AcDream.App.Tests.Rendering.Gpu&Lane!=Vulkan`
+gives307PASS/0FAIL/0SKIP on this pre-#477 scratch base. Do not compare this
+count to integrated#477's314 as a regression: that repair adds different tests.
+Existing early-capture/premature-validity pins remain in the affected run;
+their original root mutations above were not repeated again for F1.
+
+| Frozen F1 item | SHA256 |
+| --- | --- |
+| Resources.cs | 8782D8DDE8201FFCAFE0DFD55E949E5E058F6CEE20A0774D2B24989A4613070E |
+| VulkanBackbufferCaptureValidityTests.cs | BFAE818A74BC4C4C09F643BE9885ED3F42B42CDA21CC225AF2D8B80FDF95E541 |
+| Architecture | 06543E3145D26D3576BF8770F22C895DD0F6DB69F74B7CFDCCA305E525E4BFF2 |
+| Divergence register | 61E8AAE1D8A483A817B77114E51ED93BD299369344F8689126532C0353BBD37C |
+| artifacts/476-root/f1-m1/root-476-f1-m1.trx | 7EF4A88E0B579BC4009E441D3E63DA4299F05CF264AE93E375C78E556CE9509B |
+| artifacts/476-root/f1-m2/root-476-f1-m2.trx | FC24D369242E0179756AF5F0B342EB204F27DD6FBBFF4C58A37704D59956548C |
+| artifacts/476-root/f1-final/root-476-f1-affected.trx | BB98BEB2108981AC0453175AFC646D8F4A38FDF491E25B355CF5CE6C0541D232 |
+| artifacts/476-root/f1-gpu/root-476-f1-gpu.trx | E84997D224244F06CA2B980239FD7C22D8437670D97472B6053449CE316690CC |
+
+Wegener's narrow F1 re-review PASS, no findings; unchanged broad retail review1
+was not restarted. Reviewed scratch02b70713338f5513a2fd19c26c90298e474f57dc
+lands with this evidence. Nine nonoverlapping source/test blobs are identical;
+the two Vulkan partials differ only by the preserved#477 changes, independently
+read in the integration diff. Architecture preserves#477/#480; AD-121 lands
+alongside preserved AD-118 retirement, counted105total/13retired/92active.
+Fresh integrated Release/tests and current-frame graphical PASS are still owed.
+Owner12:53 FPS deferral remains binding; no profiling/optimization is part
+of this capture repair or its graphical correctness run.

docs/research/2026-09-01-overhaul/s5-consumers-material-closeout-packet.md

@@ -4840,3 +4840,21 @@ assertion, isolate the setup difference, and do not weaken/pin a golden over
 the mismatch. No production change follows from it without a separate proof.
 G4 remains FAIL; FPS deferral is not acceptance of the other owner findings.
 No merge to main.
+
+## 55. Completed-frame capture accepted for integration
+
+#476 reviewed scratch02b707133 lands with this ledger. Rawls retail review1
+PASS; Wegener production2's inherited host-visibility gap is closed by F1 and
+narrow re-review3 PASS. Root six native-state tests, affected179/179 and
+portableGPU307/307 on scratch; full Release0W0E; two actual missing-barrier/
+coherence mutations fail as intended and restore exactly. Root corrected one
+test handle cast before compiling; no implementer F1 execution is claimed.
+476-lead-verification.md gives the exact source/TRX seals and scope limits.
+
+Integration preserves#477 GPU synchronization and#480 held-item publication.
+AD-121 records opt-in capture's explicit noncoherent-memory rejection; combined
+register105total/13retired/92active preserves C1c's AD-118 retirement. Root's
+fresh integrated Release/tests and first-request/two-resize/held-item graphics
+are next. Rawls confirms no command active; all agent builds/tests are paused
+for the graphical window. Owner FPS deferral still applies. No graphical PASS
+or G4 acceptance is inferred from the offline checks.