Forsaken first-white tracing attempt, 2026-11-12
The combat-white cause remains unproved. This attempt stopped at a concrete caller blocker before combat: command 0 delivered one native Enter and completed its bound capture, then exceeded the immutable 20-second work deadline while awaiting the new post-capture diagnostic drain. No later gameplay/input was sent. There is no runtime fix or changed-path before/after claim.
The predecessor's two-owner combat evidence was read before allocation, including retained state/API records, the natural render audit and full after-secondary/first-white captures. Its postwhite ONE/ONE draws, depth-only clears and sampled white uploads remain evidence of the saturated path, not of the first producer transition or a native crash.
Evidence: scratch/runs/20261112-forsaken-first-white-1540/ (ignored).
runtime-pins.json records staged main
0284c9d79c20872b21fa8dac60443a9ad7353b46, relevant-source.patch and
untracked-runtime.patch. owned-inputs.patch records the staged helpers;
final-owned-helpers.patch preserves the final local recovery/analysis helpers.
Source/media trees were not archived. The original executable and existing
873 media pins were referenced and checked without copying media into the run.
The recovered actual module is 2,189,863 bytes, SHA-256
04c534192ee3fe39e388536c2ce710c12c47a5cd2ca4d95c19f9d00b1e8f0a17.
Diagnostic path and observed state
The existing lib/d3dim-draw-trace.js producer observer was used on the actual
legacy WebGL adapter. It records natural clears/uploads/readbacks and selected
covering draws, with bounded retention and explicit drops. Its GPU sample reads
can synchronize rendering; these observations are not performance evidence.
It does not add guest state setters, input, draw flushes or guest API calls.
A run-local cpu-observer.js wraps existing log, log_i32, log_api_exit,
dx_trace and gpu_gl_call imports before ordinary guest instantiation. Original
imports run once with their original returns/errors. Selected public D3D calls
retain entry arguments and return registers; existing kind-29 hooks identify
changed states in $d3dim_set_render_state, including execute-buffer states.
The observer uses pure register getters and direct-window stack reads, retains
at most 4,096 events between drains and reports overflow. CPU and GPU orders are
separate; natural command boundaries are retained, not a global ordering oracle.
No production source was changed.
Before original launch, a real x86 Device2/3 public-COM pixel fixture against the
built module emitted 1,701 observer rows with no drops/error, including public
arguments/returns and kind-29 changes. The fixture used the async instantiation
entry point observed by the helper, instead of its usual new Instance; its
COM calls and pixel assertions were unchanged. This is diagnostic fixture
validation, not original-game validation. Exact helper import/input closure
passed locally and remotely for all 1,547 staged files before launch.
Five compressed original drains were recovered (diagnostic-summary.json):
one startup drain and four subsequent drains. Startup retained the earliest
4,096 of 16,879 CPU events and explicitly dropped 12,783; therefore it cannot
prove an uninterrupted startup transition. All four later drains have zero CPU
or GPU drops and no reported CPU observation error. The secondary guest Worker
has no selected events in these drains.
In the last successful menu drain, the owning guest emitted 200 public
IDirect3DDevice_Execute entries and twelve each of BeginScene, EndScene,
Viewport_Clear and Surface_Flip. Changed-state hooks show eleven changes each
to blend-enable 1 then 0, destination blend 2 then 1, and depth-write 0 then 1.
The owning GPU drain records twelve accepted flags-3 clears and 22 selected
covering draw records with blending false, source 2, destination 1. The backend
maps these factors to ONE/ZERO. Thus this original menu path demonstrably
restores opaque state through execute buffers; additive state is not permanently
stuck from initial launch. This does not identify what happens during combat.
No combat, incoming-damage flash, respawn or first-white transition was reached.
Actual guest Worker 533B67761F4045B2B42C20CAA100C04E and RenderWorker
DBBAD01140C8329E623CC286E79A0DA0, endpoint 1, were bound to their recovered
module and observed WebGL context. Requested GPU flags were not treated as
actual-owner proof. This attempt did not launch software; the predecessor's
GPU-off/no-active-port inline software evidence remains separate.
First concrete blocker
The command was published with workEnd 2026-11-12T15:44:42.222Z and a separate
ACK deadline. The runtime was draining a periodic diagnostic before consuming
it; admission began at 15:44:27.707Z. Native Enter down/up were acknowledged at
15:44:30.491Z / 15:44:30.517Z, with timer-driven release armed before down.
The subsequent full 1024 × 712 PNG and its before/after owner/module binding
completed at 15:44:38.952Z. The browser then awaited drainDiagnostics('menu');
that file was never published. The work deadline fired at 15:44:42.228Z.
await-trace.jsonl, protocol-timing.jsonl, the completed capture and the
absence of the final diagnostic file locate the pending phase; they do not
attribute the entire delay to one CPU/GPU operation.
The resulting command refusal is a harness deadline, not evidence of a native Forsaken fault. Caller outcome is known exit 1; both actual Enter edges are trusted in the retained menu-state record. Later admission is sticky-closed and no command was replayed. The run-local review text incorrectly calls the startup image a splash: coordinator inspection of both immutable full PNGs shows Single Player selected on the main menu. The second capture does not establish entry to START. These labeling errors are retained explicitly; no splash, successful menu transition or combat progress is inferred from the input ACK.
Next: remove diagnostic drains from the command ACK critical path and bound telemetry independently without weakening identity, hash, full-frame, freshness, release or no-replay checks. Exercise that exact combined trace/input/capture path before another original launch. Then retain the producer transition around the first combat-white frame. Do not patch blending/clears/textures from this menu trace or repeat unchanged enemy-frag attempts.
Verification and shutdown
Locked tools/build.sh, the production bounded renderer trace tests, inline
software owner/rejection guards, three timer-release tests, thirteen API custody
tests, the run-local observer controls and actual emitted-path fixture passed.
verify.py checks 51 output paths retrieved twice and rehashed, two full PNGs,
before/after identity continuity, actual module SHA, the balanced native pulse,
trusted edges and terminal records. Compressed drains remain exact bytes.
No guest AI/config/health/clock/position writes, forced gameplay calls,
substituted media, workers, public deployment, full-27/parity/audio/FPS or broad
crash-free claim was made.
Chrome exited 0 with actual parent wait and separate stdout/stderr EOF. Its browser harness parent and outer API command exited 1, with the deadline failure retained; the browser parent has separate EOF evidence. All unique outputs were retrieved and hash-verified before delegated normal sandbox stop.
Private box bx_z7nsayjm had an independently observed watchdog prearmed before
allocation, initial expiry 2026-11-12T17:40:31.443Z. Normal stop completed
2026-11-12T15:46:44.929Z, receipt stop_8e93d36395320d01, forced=false,
error=null. Final state archived, expiry null. Lifecycle client/owner and
watchdog actual waits are 0 with EOF receipts. Do not resume/delete this box.