Forsaken threaded black-output discriminator, 7 November 2026

The registered software Direct3D route reaches the original menu and readable Volcano gameplay with threads enabled. The fresh run reproduces a black capture at 10 seconds, then the original Single Player menu at 20 seconds. Ordinary Enter, thrust and fire work. No WAT/JS disagreement was demonstrated and no production repair was made.

Evidence: scratch/runs/20261107-forsaken-threaded-black/ (ignored). Read the retained matched-module startup sweep and existing Forsaken investigations before allocating. This is one serial cooperative/threaded discriminator, not another corpus sweep or browser WebGL qualification.

Identity and observed boundary

Source is main 05babd650578b76572ca8f39ca975ce6f78e9d17 plus the recorded relevant diff and existing untracked lib/wininit-boot.js patch. Both actual standard-module reads match the retained sweep: 2,182,500 bytes, SHA-256 f9650607181301b5d2a9990420f437bd9f1ca10d40ac73067339f07499d53121. All 1,328 final runtime/helper pins and 873 original media paths match locally; remote pre/post checks and all 821 observed reads covered by those pins match. Native dependencies pass their 162-file postcheck. Node is v24.22.1. Existing native/capture controls were reused; remote headless-GL and WGL controls pass. No guest setting, clock, media, binding or private handler was changed.

The passive observer delegates existing host hooks unchanged, records their completed JS return, samples per-instance registers and 48 stack words, and hashes the actual surface bytes. coop-progress.json and threads-progress.json retain one-second observations. Hook completion is not a final guest HRESULT or completion of a WAT operation that follows the hook. Surface hashes are sampled reads, not atomic frame transactions.

The first demonstrated distinction from an unpresented-pixels hypothesis is inside the target itself:

Threaded observation Main EIP Execute boundaries Present boundaries Allocations Primary contents
9.069 s 0x4cd387 3,552 2,310 132 All zero
10.069 s 0x4cd32b 3,552 2,310 154 All zero
11.075 s 0x48f291 4,253 2,349 162 536,860 nonzero bytes

The 10-second compositor capture is black; its ordinary presentation call returns 1. Primary and back surfaces both contain zero bytes in this interval. Execute/Present counts stay fixed while allocation and Lock/Unlock activity continues. This supplies evidence of absent new geometry production during that resource-processing interval, rather than nonblack pixels lost solely by the compositor. Subsequent geometry, nonblack target bytes and reviewed menu prove recovery in the same original process. The cooperative run also has black-target intervals with unchanged Execute/Present counts and continued allocation, followed by nonblack targets. There are 22 black-primary and 83 nonblack-primary cooperative samples, versus 10 and 128 threaded samples; the latter includes gameplay and these totals are not performance comparisons.

The main guest is running with sampled yield/wait zero during these black intervals. It progresses through 0x4cd2c2..0x4cd40e, rather than remaining parked at one wait. Retained original disassembly shows byte reads, lookup tables and conversion loops there; its exact semantic function name remains unknown. Auxiliary workers start/exit and the scheduler continues. The threaded callback-lifecycle receipt has zero events/pending entries; sampled multimedia callback-active flags are zero. These observations do not prove correctness of every wait/callback or exclude an unsampled transient context error.

No first incorrect completed operation was identified. The old sweep's threaded final EIP 0x4cd32b lies in the same progressing loop, consistent with sampling a resource-processing interval. Its sparse retained log cannot prove the exact historical blackening operation. The remaining causal boundary, if transient black intervals themselves require a fix, is the first original clear/reinitialization and its guest caller before this conversion/allocation phase, with complete API returns and resource identities. The present evidence does not justify a scheduler, submission, clear, blend or compositor patch.

Reviewed original progress

Named captures below are under retrieved/actual/threads/; copies inspected during execution are under early-captures/threads/.

Three Enter pulses of 120 ms followed the reviewed menus. All five presses arm UP independently before DOWN. The threaded control channel has 22 exact success ACKs including captures and ordinary quit, ten balanced keyboard events, no UNKNOWN, no replay, and empty terminal held/pending sets. These are ordinary CLI renderer input events, not browser trusted-native-event evidence.

Verification, harness limitations and lifecycle

Existing second-instance D3DIM Worker pixel parity and control-stdin regression tests pass. Evidence analysis, module/read/pin checks and helper syntax checks pass. No production source change, fail-before/pass-after repair regression or new full build is claimed. Retained module and controls are used as requested.

Run-local setup mistakes are retained: a helper edit during staging tripped a local hash assertion before its transfer; final helper versions and remaining inputs were staged and verified before launching. One capture lookup preceded file creation. A read-only attempt to open the runner's stdin socket through /proc failed before writing an ordinary quit command. No guest input was sent by that attempt. The cooperative arm and its parent driver were then stopped before the first threaded launch; its exit is SIGTERM, not a normal quit/code-0 claim. A corrected file-driven ordinary control channel served the single threaded arm, which exited 0 after 138.348 seconds. There was no original replay or overlapping heavy runtime. The unused initial pulse path was never invoked.

The reused remote cleanup helper incorrectly says input was not injected. analysis.json explicitly corrects that wording with the complete balanced threaded command evidence. Its process census is valid and empty; no browser, port or profile was created. Local source identity and owned helper patches are retained without source/media archives.

All 76 remote references / 68 unique contents, 19,496,511 referenced bytes, including logs, PNGs, native modules and both WASM modules, were retrieved and SHA-256/size verified before normal box shutdown. Both launch floors exceed 5 GB. The independent detached 25-minute watchdog was armed before allocation and never extended. Box bx_ame2gydy, created 22:57:31Z with original expiry 23:22:31.967Z, is archived, current expiry null. Normal stop stop_7dd40d405afc5e0e completed 23:04:09.514Z, forced=false, error=null. Both lifecycle delegates exit zero and their supervisor has no remaining children. Owned runtime processes, input, ports/profiles and watcher are released. No stopped box was resumed/deleted and no public deployment occurred.

Reviewed software menu, movement and firing are established. The historical 20-second black endpoint alone is insufficient evidence of a persistent threading defect. Browser WebGL, scene parity, FPS, audio and level completion remain outside this task.