NFS3 native window-present association, 2026-11-18
Evidence: scratch/runs/20261118-nfs3-native-window-present-association/.
All preceding original/UNKNOWN scopes remain closed.
The authored genuine Device2 API control creates its UI on a separate native thread and renders through normal guest Workers and a real render Worker. With the frozen c9 browser source and native 7d347 module it reproduced the zero window-present association. The publication existed, its primary was still type 2, its creator and cooperative window matched, and its CPU present stamp was unchanged. Its saved DIB differed from the primary's current DIB. This was a live Flip transition, rather than evidence of retirement or a foreign instance. The read-only observer samples existing native sync kicks; it neither submits a present nor changes guest memory, clocks or input.
The producer swapped the primary/back-buffer DIBs before entering
d3g_present, which could synchronously wait for the preceding in-flight
PRESENT. Throughout that wait the old publication failed the exact-DIB check.
The host correctly marked its retained layer stale and excluded it. The old
control retained 114 sync entry/exit observations, all with zero association
and mismatched DIBs. The corrected control retained 58, all with valid
associations and matching DIBs, while completing 60 actual presents.
The correction reserves PRESENT capacity before Flip changes the primary DIB. The ordinary PRESENT capacity check remains for other callers. Native reader checks for type, DIB, creator, target window, render target and sequence remain unchanged, as do CPU retirement and browser window/instance fences. No host composition, freshness, binder, inventory, physical input or guest clock/state semantics are changed.
The focused native regression queues actual records behind an asynchronous consumer and checks the preceding publication at its synchronous backpressure boundary. It fails against the old producer and passes against the corrected producer. Existing foreign-target, window-switch, retired-primary, same-address allocation reuse, CPU-write and whole-pixel checks remain in the same regression. Separate browser pressure controls and exact module recovery are recorded in the evidence; whole authored blue/white controls were reviewed as control pixels, without claiming game progress or observer neutrality.
The runtime repair was published as 1fe5e1baf after the focused tests and
locked full build. Exact recovery of the old 7d347 module was verified before
applying only the producer delta, producing b8c8fa6c. Both exact-module
light/pressure browser controls completed 60 presents with 62 valid sync
observations each and zero mismatched DIBs. All six authored whole PNGs are
byte-identical; this establishes control pixel equivalence, not game progress.
The one fresh registered original used the frozen c9 source, genuine chooser, 135 verified media files, published wait-source correction and b8c8 native repair. Actual Worker module inventories were read back. At 11 and 234 presents the native window-present association was positive with matching primary DIB, creator, window and CPU stamp. Its whole reviewed page was still white below the host strip. No binder or physical DOWN/UP was used. The remaining white composition cause was not established by these snapshots.
The attached submitting process then disappeared, its independent watchdog reported parent disconnection, and the native supervisor received SIGTERM. The sender is UNKNOWN; the coordinator explicitly reported sending no signal. The native receipt records cleanup without leaked descendants, and the tool wait returned 143, but the dead outer controller did not retain its intended watchdog/browser waits. This scope remains closed, with that evidence limit.
A preparation correction puts the existing driver and native supervisor under a detached custodian. It retains actual native/watchdog waits and both EOFs, with the same 900/920-second limits, 180-second Quit reserve and disk floor. Real process controls show attached-owner TERM closing the native/browser before an authored WASM result, while loss of the detached launcher's submitting process under TERM and KILL allows the browser to execute native WASM and finish normally. These controls do not identify the original signal sender. The driver, guest runtime, input confirmation/review/freshness fences and independent physical release path are unchanged.
The preparation correction was tested, built and published as 49af4ab72.
The authorized additional original retained the same frozen source, chooser,
135 media pins and exact b8c8 native repair, with actual guest/render Worker
module readbacks. Its initial whole loading screen had a 27 ms native scene,
positive window association and a selected, non-stale GPU layer. After loading,
the whole personally reviewed screenshot original-durable/post-loading-scene.png
showed the cockpit, road, rain and opponents at 0 MPH. Native window and RT
association were both 54032 with matching DIB, creator, window and CPU stamp 1.
However, that screenshot's native scene was approximately 1329 ms old and
ready=false. It did not authorize input. No bind or physical DOWN/UP occurred.
The independent watchdog then observed free disk at 1466826752 bytes, below
the required 1.5 GB floor, and closed the scope. The native supervisor retained
SIGTERM followed by SIGKILL cleanup, leaked=false and pollError=null; the
detached custodian retained actual native and watchdog waits and both EOFs.
The run remained closed when shared disk space subsequently recovered.
The tested producer repair fixes the demonstrated invalid association during Flip backpressure. Original driving, road consequence, Pause/Continue and ordinary Quit remain unverified: no reviewed driving screenshot coincided with an admissibly fresh native scene before the disk-floor closure. The cause of the remaining scene lag and intermittent white composition is not established by the retained snapshots. No additional original was launched, and no freshness, identity, generation or retirement check was relaxed.