FoxBear original: visible scene, control binding blocked

One authorized registered dx_foxbear original reached a personally reviewed complete 1024×712 frame: forest platforms, fox at lower left, bear at lower right, pit and HUD. initial.png is the selected original frame in scratch/runs/20261118-foxbear-original-visible-gameplay/. An app-specific input consequence and ordinary Quit were not established.

The earlier investigation is apps/foxbear.md, including resolved asset-name padding and DirectDraw object-capacity startup failures. No local original C/C++ sample source was located. Embedded help and the original PE key jump table supplied this task's control evidence; Donuts gameplay semantics were not imported. Numpad 2/3/4/5/6/7 select crouch/apple/right/stop/left/jump per embedded help. The jump table at VA 0x402008 maps their WM_KEYDOWN routes to 0x401b97/0x401ba5/0x401b6d/0x401bb3/0x401b89/0x401b5f. F12 maps to 0x401b42, posting WM_CLOSE to the game window.

The designated source is the existing registered local pair, unchanged before, at dispatch and after the run. These hashes identify local bytes, not an authenticated upstream distribution. The older app note labels a DX5 SDK sample; no upstream receipt or source provenance was located. Nothing was downloaded or replaced:

File Bytes SHA-256
test/binaries/dx-sdk/foxbear/foxbear.exe 389120 2ca53c3bf88977f56393466dad3f17a2892e8717066e6cf396bdc111ccbc452c
test/binaries/dx-sdk/foxbear/foxbear.art 1644824 ad7199964ab49488a5dbb647899eda51caeae52aa7720642870a93590206ac3e

The finite controller adaptation and native tests were published as 934b81cfc, with immediate strict pull/push. Native authored-guest checks passed all six numpad keys and F12, stale-review refusal, independent UP, lost-ACK closure, changed-focus cancellation and late-confirmation closure without subsequent endpoint calls. The locked full build passed. A separate queued-expiry test failed its coverage assertion because admission expired before DOWN; it proves no queued-expiry execution and is retained explicitly as inconclusive. Tests used authored media, never another FoxBear original.

The original used the existing frozen indexed-source loader and native-supervise.cjs, normal guest Workers, real ticks and native raster rendering. Source Git was 6f25a1ea09e87cfb67dfb42a1306482f01d64a7c; the independently rebuilt module was eb2395e1c115980dd78dbb646a26f755a4119a2a1e086942e715240e07c34bc6. The compile recipe initially hashed Buffer source bytes after the compiler normalized non-ASCII comments through a shared Buffer slice. Raw Git hashes and normalized compiler-input hashes were subsequently separated and every normalized hash independently regenerated. No module bytes changed. All126 observed loads matched the frozen inventory; source reads were observed in main and Worker threads0/1/2. Only main's module-byte read was observed. Workers receive the compiled module object through the pinned host's init message; there is no independent per-Worker module-byte hash receipt.

Preparation constructed the actual renderer, encoded/decoded a complete PNG, started GuestThreadHost without a PE, and verified registry/media in main and a real Worker. Final audit found an important preparation limitation: sealed media entries used test/binaries/..., while the registry's actual reader uses the top-level binaries symlink spelling. No original media read appeared in the sealed-read log. Matching local before/dispatch/after hashes are retained; immutable actual media-read coverage is not claimed. A future preparation must audit the actual registry host-path spelling before spending an original.

At epoch1794970376.512, the one permanent native-scene-v1 bind failed with reviewed native paint changed. The reviewed capture began at1794970362.709; the bind intent was recorded at1794970376.145. This route hashes the current entire encoded canvas against the reviewed PNG before binding. The original passed the scene-identity comparison and failed the paint equality check. Which pixels changed is unknown; no second frame was requested. No keyboard DOWN, independent release or input-owner lease was created. The helper's sticky terminal record closed the scope. No subsequent endpoint request, rebind or original replay occurred. The guards were unchanged. This is a concrete control-workflow blocker, not a proved guest rendering/input defect; no shared-runtime repair or conditional second original was justified.

Owned host cleanup was requested after closure, so it is not ordinary F12 Quit. The invoking launch tool actually exited143 before writing original native wait/EOF and containment receipts; the signal source is unknown. Those receipts remain missing/null. The original and launcher group were absent; an owned orphan watchdog was identified and removed by a separate finite supervised cleanup. That cleanup's actual wait and both EOFs passed, all three known original/watchdog/group PIDs were absent, and port37099 had no listener. No native exit0 or complete original lifetime proof is claimed. Observed disk checks stayed above1.5GB, with no RAM gate; original watchdog minimum telemetry was lost with the interrupted owner.

The run retains exact Git references, relevant unapplied worktree patches, owned-file patches, selected whole PNGs and unique module/input evidence; result.json is published last with a failed goal. No source archive, frame series, capture bundle or media copy is retained. BloodII lifetime, NFS admission and OrganicArt ownership were preserved. No workers, agents, boxes, Python, Wine or deployment were used. Box ID/state/expiry/stop receipt: null/NOT_CREATED/null/null. Ownership is released and this task stops.

Moving-frame workflow repair, fresh task on 2026-11-18

Run 20261118-foxbear-moving-scene-reviewed-controls-repair repairs the controller and preparation workflow, but does not establish original gameplay consequences or ordinary F12 Quit. Its one fresh registered launch stopped before PE loading. The predecessor scope above remains closed; its missing original wait/EOF receipts remain missing.

The existing native-reviewed-context-v1 route in lib/cli-native-scene.js and test/run.js already supports animated paint. FoxBear now uses its atomic native-reviewed-capture, retained exact PNG bytes/hash and native capture ID, original wall/monotonic times and per-intent native target verification. No runtime input or paint guard was removed or changed. A changing authored native GDI pixel reproduces the old paint-equality bind failure; the same real native renderer/USER/Worker fixture succeeds through the retained-frame route while the PNG changes. This is an authored canvas animation control, not another FoxBear run or a fabricated native identity.

At binding and consumption, the route checks the reviewed full native window metadata, process/thread, HWND generations, focus/modal/capture/physical state, screen and pointer mapping, target ancestry/visibility/enabled state and key focus. The permanent input owner retains renderer/queue/device/actor references and epoch; the controller also validates the native PID start token and owned loopback endpoint. Each retained frame reserves one finite nonce, and the guard checks the actual parsed input at consumption. Changing paint alone need not invalidate a reviewed frame whose native control context remains unchanged. The route's existing gameplay/menu freshness limits are 15/30 seconds from the original capture, without renewal; the fresh launch selected these existing limits. Native tests also use a stricter five-second controller gameplay bound. The 500ms confirmation, 3000ms operation/observation, 1750ms independent-UP and fixed Quit reserve remain enforced. No original was paused or frozen.

Real native controls pass six numpad keys and F12 delivery, wrong epoch, unretained frame, wrong hash/window, changed focus, replay, stale review, lost ACK with independent UP, late confirmation and sticky terminal refusal without later endpoint calls. Queued-expiry now actually applies DOWN to the renderer queue, expires it before the slow authored native USER pump can consume it, purges it and confirms independent UP. An initial late-confirmation test used a 1000ms pulse whose finite admission refused before dispatch; its coverage assertion failed. The corrected test uses a 100ms authored pulse and reaches the deliberate late ACK. This changes test stimulus, not guard budgets. Existing retained-context unit controls cover wall/monotonic freshness, generation/control metadata, invalid targets and single-frame nonce reservation.

Preparation now places canonical media identities and allowed aliases in the authenticated sealed index. The actual registry binaries/... spelling maps to the same sealed bytes as test/binaries/.... Each alias must still resolve to its pinned canonical path; retargeted or unlisted media cannot fall back to mutable reads. Sync, Buffer/URL, callback, promise and stream reads pass in real main and Worker threads. Live file-content mutation cannot alter sealed reads; fresh preparation rejects changed designated media hashes. Authored retarget and missing-input controls reject. No media copies are retained.

The existing detached blood2-durable-launcher.cjs and native-supervise.cjs are reused. Native wait receipts now also retain the process start token. Separate authored submitter TERM and KILL controls pass with actual native and watchdog waits, both EOFs and no containment leaks. The frozen renderer and GuestThreadHost preflights pass before the fresh registered launch. Code was published as 084802a67 after focused tests and a locked full build, followed immediately by strict pull/push. The independently rebuilt published WAT closure still hashes to eb2395e1c115980dd78dbb646a26f755a4119a2a1e086942e715240e07c34bc6; raw Git and normalized compiler inputs are independently recorded.

The fresh original's frozen source graph is 084802a67; the EXE/ART hashes in the earlier table are unchanged. Both actual registry paths were read from the frozen inventory in main and a real Worker before dispatch. However, the first inventory omitted the CLI-global system type libraries. Native PID2350139 failed on test/binaries/tlbs/stdole2.tlb at the ordinary VFS boot mount, before a PE, module instance or original guest Worker started. No original scene was captured, no input owner or DOWN was created, and no original control endpoint was contacted. The actual native exit is 1, signal null, with actual wait and both EOFs; watchdog exit0 and its actual wait/EOFs also survive. Containment owner exit0 is not the native exit code. The owned tree is retired with leaked:false, minimum observed home free space2171752448 bytes and no RAM gate. The helper's terminal-scope check rejects locally with zero transport calls. The83 observed main-thread reads include the sealed EXE at its actual binaries/... alias and the eb239 module bytes, both with matching hashes. ART was not reached, and no original Worker-read receipt or original module-instantiation/guest-execution proof is claimed.

That preparation gap was repaired offline, without a replay. Collection now seals the normal boot type libraries and Git font data; rendering preflight executes the actual VFS system-data mount and required font reads in both main and Worker before any PE. Deliberately omitting stdole2 fails there; the complete inventory passes. A full frozen authored animated renderer/input run with normal guest Workers also passes using the retained eb239 module. The locked full build and shared preparation regressions pass after this correction. This is a preparation defect, not a proved guest runtime defect; it authorizes no affected additional FoxBear original under this task's one-original limit.

The concrete blocker is original validation after the corrected boot inventory: app-specific visible input consequence and ordinary F12 Quit remain unverified. Evidence retains exact Git and owned patches, the unique module, logs, media hashes and selected named authored PNGs; result.json is written last with a failed goal. No original source/frame archive, media copy, workers, agents, boxes, Python, Wine or deployment were used. Box ID/state/expiry/stop receipt: null/NOT_CREATED/null/null. Ownership is released and the task stops.

The final image audit found that some early automated captures sampled the desktop before deferred window painting, although their native window identity and changing GDI pixel were genuine. They are supplementary input controls, not evidence of a fully painted fixture. The tracked test now uses the normal --live-present route, waits for both renderer windows and requires more than eight decoded colors before review. visible-atomic-main and visible-atomic-old-paint pass on the same published 6308e7a3b graph and eb239 module: the whole painted window was personally inspected, moving paint rejects the legacy binder, and the retained-frame route accepts bounded native input.

The slow-pump queued-expiry negative intentionally delays initial WM_PAINT. It proves native queue cancellation and independent UP only; it is not a visible scene-review positive. A stricter image assertion initially rejected that negative before dispatch; the final test makes this distinction explicit. The full painted positive and legacy-rejection pairs each decode to 205 colors and differ at exactly the real guest GDI pixel (544,264). The current-target refusal also passes with a fully painted window.