docs: contract completed-frame capture repair and safe self-gate launch

docs/research/2026-09-01-overhaul/476-completed-frame-capture-contract.md

@@ -0,0 +1,92 @@
+# #476 — capture the submitted current frame, never uninitialized readback
+
+Lead contract2026-09-05. NOT IMPLEMENTED. Read the full
+476-capture-lifetime-diagnosis.md first: it records the lead's production
+trace, failed PNG and named-retail/paired-binary screenshot boundary. This
+is a modern Vulkan retained-copy lifetime repair, not a new AC algorithm.
+
+## Bounded change
+
+Move the existing IPrivateFrameScreenshot invocation from
+PrivatePresentationRenderer's still-open render phase to the outer
+RenderFrameOrchestrator's successfully closed frame boundary. The sequence
+is resource/world/private drawing, GPU measurement close, ONE GPU frame close,
+capture, CPU diagnostics, shadow comparison, lifecycle checkpoint. Keep the
+existing immutable outcome fields truthful: ScreenshotCaptured reflects that
+post-close capture, while portal/private presentation facts remain those of
+the frame just drawn. Wire the same optional capture adapter through
+FrameRootComposition; no second frame owner or screenshot queue.
+
+Preserve the existing begin/render/measurement/recovery/close error contract:
+Begin failure does not close; every successful Begin attempts exactly one
+End, including render/measurement failure; recovery remains before End when
+render fails; aggregated errors retain their original ordering. A close
+failure must not execute capture/post-close consumers or cause a second End.
+Capture/diagnostic/post-diagnostic failure AFTER close must propagate without
+trying to close/recover a finished GPU frame. Zero-area input opens nothing,
+captures nothing and leaves queued requests for a later renderable frame.
+Do not alter the idempotent GpuDeviceFrameLifetime clearing rule.
+
+In VulkanGpuDevice's existing capture fields/methods and EndFrame, ensure
+readback is valid only after an actual successful submission containing the
+retained image-to-buffer copy. Allocation/resize or command recording alone
+is not validity. Prevent an open/unsubmitted frame, missing copy, skipped
+acquisition, failed submission or new-size buffer from being advertised as
+the newly completed requested frame. Use minimal local state on the existing
+owner; reset it at the actual relevant lifetime edges. Preserve the correct
+bring-up caller that already reads after frame close. Do not read the
+swapchain after present or change semaphore/flight/resource retirement.
+
+Retain opt-in capture and zero copies/allocations when off. Keep the existing
+DeviceWaitIdle readback behavior for a valid submitted capture; do not add
+waits elsewhere. No sleeps, first-image discard, guessed warmup, retry-until-
+pretty loop, shader/quality change or screenshot naming/orientation rewrite.
+The controller's one-request-per-frame, complete/failed status and atomic PNG
+behavior remain intact. Do not bundle unrelated stale GL comments/cleanup.
+
+## Files and ownership
+
+Scope: RenderFrameOrchestrator, PrivatePresentationRenderer,
+FrameRootComposition; capture-only portions of VulkanGpuDevice and its
+Resources partial; existing focused orchestrator/private/composition,
+screenshot/lifecycle and Vulkan capture tests; relevant architecture paragraph.
+No Runtime/world/lighting/mesh/walk edits. Any intentional behavioral deviation
+needs its register row in the landing commit; none is expected.
+
+Use a fresh codex/s5-476-openai-impl scratch at the commit containing this
+contract. #477 is in a DIFFERENT scratch and touches other portions of the
+same Vulkan partials. Do not copy/edit its source. Root integrates narrow
+accepted changes and preserves both; do not perform a whole-file overwrite.
+No agent commit/graphical launch or extra agents. Both AcDream.App and retail
+acclient must be absent before EVERY build/test.
+
+## Proof
+
+Exercise the REAL orchestrator plus real PrivatePresentationRenderer and
+FrameScreenshotController composition with a fake GPU/frame whose distinct
+frame pixels become readable ONLY at End/submission. Do not merely assert a
+manually written phase list. Startup and actual-size-change first captures
+must carry the newly submitted pixels, with exact orientation/size. Reuse
+existing fake-GPU/test fixtures rather than a new recording framework.
+
+Verify real phase order and every failure boundary above, queued requests,
+zero-area then resumed frame, and capture-off behavior. For native validity,
+exercise the actual production state transitions/call binding, not an unused
+Boolean policy helper. Capture before first submit, after resize before its
+first copy, while open and after failed/no-copy submission must not read old
+bytes as a valid current frame. Existing bring-up/format tests remain green.
+
+Two actual-production mutations: restore capture before End; then mark the
+new buffer/recorded-but-unsubmitted copy as valid. Each must fail its intended
+pixel/validity assertion, not compilation or test setup. Restore exact raw
+hashes, rebuild Release0W0E and run affected hermetic lanes with zero new skips.
+Return commands/TRX/hashes and precise claim boundaries. Root independently
+verifies, then sequential behavior/retail and production/lifetime review
+lenses follow within the owner's ten-review limit.
+
+Graphical closure only after #477's synchronization run is clean and no agent
+is building/testing: fresh Release, ACE9000 up, both clients absent, one
+bounded real Vulkan run. Capture first requested frame at startup and across
+two REAL different framebuffer sizes, plus a settled portal/world frame.
+Preserve every PNG and record PASS/FAIL paths, provisional until owner looks.
+This repair alone cannot close #477 or the failed G4 owner matrix.

docs/research/2026-09-01-overhaul/478-480-readonly-diagnosis.md

@@ -99,3 +99,36 @@ 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.
+
+### Lead static follow-up (no lighting repair or live witness)
+
+The lead independently read CellManager::ChangePosition, minimize_envcell_lighting,
+DrawEnvCell and SetStaticLightingVertexColors, then checked the paired exe
+above using its PE sections. Exact body ranges/hashes:
+
+| VA / bytes | SHA-256 |
+| --- | --- |
+|004559B0 / 555|DDF73D34A7A8099672191EB5A18A9C21F73D4A3A0B379B94840D709AD6E64D49|
+|0054C170 / 58|C948472BF8FE01BD318F4EB2BE0A6FD0FB1D87BE19114F776C613CCE3737250D|
+|0059F170 / 293|408D9818DD02750825F36D05824D3562F9BB62D3FEFABF1DD172375A3C6C91B2|
+|0059CFE0 / 759|E0148A7FA69A56ED2D4E49201A8284AA5907AE314FF9A8CCC9E77B3FC3DDD340|
+
+The ambient branch loads0079B7FC (raw0039B7FC), bytesCDCC4C3E =0.2f,
+and passes whiteFFFFFFFF. Static/dynamic capacity data0081EC94/98 is
+40/7 respectively (bytes2800000007000000). DrawEnvCell's0059F1F6
+E8E5DDFFFF calls0059CFE0. That routine's static loop increments/compares
+against num_static_lights at0059D17E before looping; it does not share an
+eight-entry EnvCell selector with the dynamic hardware lights. This confirms
+the relevant difference, not its activation in the failed Town Network frame.
+
+Lead also read the current filmic shader and its gamma helpers/default
+settings, then independently evaluated its scalar formula with no bloom,
+rays or volumetric contribution, saturation/contrast1. At screen center,
+display inputs0.20/0.10/0.05 map to approximately0.140/0.052/0.023 under
+exposure0.8/fullACES; corner strength0.245 further dims them. These are
+formula examples, NOT measured current scene pixels. No parameter is changed.
+
+Before the next unattended run, the existing self-gate runner was narrowly
+hardened to refuse either client and explicitly use WindowStyle Hidden.
+It already uses a real hidden Vulkan window through exact-framebuffer mode.
+No route/rendering behavior, graphics profile or quality was changed.

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

@@ -4706,3 +4706,11 @@ old route labels "Sawato"/"Rynthid" are unreliable identifiers. Evidence and
 candidate elevated2D31 complex are in479-rynthid-location-evidence.md. A
 read-only object/part trace is in progress; transform cause still unproven.
 All #478/#481, #476, exact route, C1c/C2/full-G4 obligations in§49 remain open.
+
+Confirmed #476 capture-before-submit now has its separate bounded contract,
+476-completed-frame-capture-contract.md. A fresh scratch may implement it
+independently; root must integrate only accepted capture portions alongside
+#477, never overwrite its unrelated Vulkan barrier changes. Graphical closure
+waits for477 validation and an exclusive no-build/test window. The self-gate
+runner now explicitly starts hidden and refuses either graphical client;
+PowerShell parse/diff check pass. No rendering policy/quality change.