Deus Ex blank world: capture ownership/freshness is the first unmeasured boundary
No production rendering or input defect is proved. The demonstrated driver defect is that visual readiness can be approved without establishing which renderer publication the PNG contains. A run-only repair now binds approval to the latest capture, checks the owning actor Worker and visible process-owned world window and displayed DOM canvas, records publication/input boundaries, and rejects stale, pending, crossed or missing-owner evidence. It is tested offline against the exact historical JavaScript contracts and the recovered driver's actual capture and approval branches. These tests do not reproduce the live teal failure or prove that input reached an engine tick. Another gameplay attempt is not authorized by this result.
Evidence and executable repair live in
scratch/runs/20261104-deusex-blank-world-cause/. Existing captures, modules and
logs are referenced in place, not copied. Inputs are the feasible-region note
published at fade2000a, the supported-native-route note at 5f12207f, and
scratch/runs/20261104-deusex-feasible-region-native/recovered/.
Earliest observed divergence
I personally viewed 04-restored.png, 05-left1.png, and 06-left2.png in the
recovered directory. Restored and firstpulse show the same crate room, ceiling
light, crosshair, health display and inventory HUD. Secondpulse shows a teal
application area with the world and HUD absent; the surrounding toolbar and
runtime log remain. This is invalid world evidence. Color does not identify its
cause or prove that the game rendered a teal frame.
| Boundary | Retained observation |
|---|---|
| Load/readiness | One Enter restores count14; actor sample at 04:00:44.678Z, screenshot/state completes 04:00:44.782Z; ACK3 at .783Z |
| First Left | Trusted down/up 04:01:05.349–.375Z; requested25ms, hold26ms, operation28ms; settled native response zero; valid image/state at 04:01:05.898Z; ACK4 at .899Z |
| Second Left | Trusted down/up 04:01:18.542–.568Z; requested25ms, hold26ms, operation28ms; actor sample at 04:01:18.981Z; invalid image/state at 04:01:19.037Z; ACK5 same millisecond |
| Stop | Invalid-image finding retained at 04:01:20Z; command6 at 04:01:35.583Z contains only quit |
The first zero native response precedes the first invalid image. The earliest
observed visual divergence lies after the valid firstpulse image and by the
secondpulse image, not necessarily at second keydown. There is no intervening
frame history. Actor sample and screenshot are sequential operations, not one
atomic world snapshot. State timestamps bound screenshot completion, not guest
frame generation. Both pulses retain exact location
[7026.70458984375,-1707.887939453125,109.60284423828125], yaw250640,
Physics1, BaseSmall0 and velocity0. Instruction pointers differ across these
samples, and observer timers advance; this excludes a literally identical
serialized sample being replayed, but does not prove engine simulation/render
progress or that the observed actor is the viewport's current actor.
Six matching terminal ACKs complete, maximum6 steps, held keys empty. All14 trusted DOM events balance. These establish driver completion and browser event delivery. They do not establish event dequeue, shared key-bitmap reads, simulation consumption, window ownership or presentation completion. The only ArrowDown selects Load in the menu. No Down preparation, Up, Space, second Load, guest write, save/export, correction or guard weakening follows.
Exact historical contracts, not current-main inference
Retained canonical module: 1,898,260 bytes, SHA256
8459597c8d66e098d727fb4ab53e1cf46332241dc7eef2e158f3c475a78b1676;
companion ABSENT in retained before/after checks. Source recipe remains Git
a504fa248caaecbbb042b65e9ae5384a340ade26 plus nine-file patch
scratch/runs/20261104-deusex-supported-stack/historical-source.patch, SHA256
9415aea6dbced13143b9dcdceb9d2d3594219a467263f7b3cf2bb623da36c796.
retained-audit.json verifies the recovered module, all52 recovered unique
files against their final manifest,453 manifest entries, ACKs/events and the
exact hashes of six relevant JavaScript files against provision pins. This is
local verification of retained evidence, not a new remote453-file inspection.
The historical patch changes nine WAT/API files; it does not change these JS
contracts. Its DirectDraw hunk changes attachment matching by creator version,
not capture ordering or keyboard timing. No module rebuild/substitution occurs.
The recovered capture() does exactly: owning-Worker actor sample →
page.screenshot() → page state. Its page state includes isolation, threads,
guestWorker existence, DOM events and log text. It records no renderer/window
identity, frame/write sequence, pending dirty state, active slice depth, queue
delivery or key-state publication. Its old capture test only checks this order
and screenshot-error propagation. approve-ready checks native support and the
hash of any named inspected PNG; it does not require the latest capture or a
new committed frame. That is a demonstrated observer/readiness defect, not
evidence that it caused the teal pixels or zero motion.
In the exact historical source:
lib/renderer.js:_workerPublicationHeld()follows executing Worker slice depth.repaint()defers while held;_repaintOnce()composites the window stack;_presentDisplayCanvas()selects direct/filter presentation. The compositor can draw an empty desktop when no appropriate window is in its visible stack. Neither stack nor branch was recorded live.host.js:_presentWanted()includes DX dirty/lock-deferred/shared dirty and scheduled/deferred repaint. Presentation can await rAF or a slice boundary. A successful CDP screenshot is not a host presentation fence.lib/renderer-input.js: physical state and queued async message state are separate.handleKeyDown/Upenqueue;takeInputdelivers. Unlike mouse releases, keyboard releases have no minimum-frame hold. Down/up can drain in one pump; a later physical-state read can see up. This is a possible boundary under a26ms hold, not a demonstrated failure or permission to extend it.host.js_workerKeyBitmap()andlib/guest-rpc.jsget_key_down_state: the page publishes physical key bits; Workers can answer locally from that bitmap or cache while the appropriate event state is open. Renderer getter counts therefore include publication and DOM work and miss local Worker reads. They cannot be called guest-consumption counts.
The console has startup resource404s, normal thread creation, the startup thread's recorded exit, and C++ throws while opening Load before the valid restored image. No new console error identifies the later visual boundary; the last retained entry is an asset-cache message at 04:00:27.699Z. Absence of a pageerror file is not proof that every Worker/render endpoint was healthy. Worker identity enumeration selects original slot0/tid1 and records the three remaining guest Workers. It identifies the memory observer; it does not tie the screenshot's visible surface to that actor or prove ownership stayed live.
Tested minimal repair and executable handoff
The new boundary-observer.cjs wraps existing JS endpoints, forwards arguments,
receiver, return values and exceptions, and records only successful completed
composite/display calls, window/process ownership, pending flags, bounded key
dequeues and bitmap transitions. It reads shared DX dirty/lock state through
the original broker. It calls no guest export for painting/input and does not
force repaint, inject events, consume key press bits, pause/freeze, change a
clock or edit guest memory. The capture helper additionally reads the selected
Worker's slot/tid and actor before/after the single screenshot.
prepare-driver.py generates recovered-driver.patch and optionally the
executable driver from the exact recovered driver. To prepare a later
authorized run (do not launch it as part of this offline task):
python3 scratch/runs/20261104-deusex-blank-world-cause/prepare-driver.py --output scratch/runs/20261104-deusex-blank-world-cause/next-driver.cjs
Stage that generated driver as the run driver and the two new .cjs helpers
inside its existing feasible-evidence/ directory. Reuse the predecessor's
remaining helpers and original runtime/media/save/config/binding pins unchanged.
The helpers use Node/Puppeteer already required by the original driver; there
is no new browser dependency or production asset to install. The driver loads
OUT/capture-boundary.cjs directly, so the new observer does not alter the453
HTTP/runtime pins. Record the new driver/helpers separately as run inputs.
Every capture writes *-boundary.json; the existing actor, PNG and state files
remain. Capture invalidates visual approval. approve-ready requires the latest
capture/hash plus successful boundary qualification and the original native
support check. A fresh publication, consistent owner and unchanged publication
counter across screenshot are necessary; a pending/held publication, old frame,
crossed publication or absent owner rejects approval. One PNG is retained even
when unqualified; no automatic screenshot retry or gameplay follows from that
failure. Existing human PNG inspection is still required. This conservative
gate can reject a valid busy frame: it must not be relaxed simply to obtain
another pulse. It does not synchronize browser compositor timing or establish
actor/viewport identity; those remain live observations.
Focused checks pass:33 boundary cases, including actual historical repaint and
input methods executed in Node VM fixtures; generated-driver syntax; actual
repaired capture/approval branch with fresh acceptance and stale/pending/missing
owner rejection; original queue/timing tests and retained evidence checks.
Fixtures demonstrate coverage of the observer boundary, not rendered pixels,
input consumption or live causation. Initial fixture failures (missing original
region-map dependency; approval-branch extraction/context errors) were corrected
in the offline tests only and retained in prerequisite-failures.json.
No production code changed, so no production build was required.
Remaining discriminating live observation
For the next original attempt, retain these observations before approving the first preparation pulse, then around its single postpulse capture: the live world hwnd/process and owning Worker/actor, last completed publication and its window surfaces, pending/held/DX state, actual key queue delivery and published Left-bit transitions. If the world owner disappears, distinguish window destruction/visibility/process replacement from a wrong observer. If the publication does not advance or is pending, retain that boundary rather than treating screenshot completion as a new frame. If an unchanged owning surface is committed and the PNG still loses the world, investigate that surface/display path. Even then, a teal committed frame alone does not prove why rendering failed. If published Left changes but native motion stays zero, the remaining missing observation is the owning Worker's actual local key query and engine tick/viewport state during the hold; page counters cannot supply it.
This task makes those initial observations executable but cannot reconstruct them retroactively. Stale/pending presentation, missing/wrong owning endpoint, genuine guest/render failure and an unobserved short-pulse input interval remain competing explanations. No geometry, point/duration chase or gameplay success is claimed. No browser/heavy CLI/game, remote access, allocation, resume or lifecycle operation ran. No box is assigned; bx_dzkgnet6's last retained state was RUNNING/APIidle, expiry2026-11-04T04:42:08.886Z, with coordinator owning shutdown. Its current state was not queried. Immutable count14 and original guards remain intact. This single task ends after publication.