OrganicArt bounded startup observations: fresh SOFTWARE startup succeeds
Run scratch/runs/20261118-oasaver-bounded-startup-progress-cause/ publishes
observer 47095dcb6 and performs the one explicitly authorized local SOFTWARE
causal reproduction. The fresh original binds in 4,108 ms and renders a
scene. The earlier 15-second failure does not reproduce; its cause remains
unknown. No emulator behavior repair or second original is justified.
The predecessor's ready:false conjoined PRESENT and composition and did not
prove actual zero PRESENT. Accepted historical Falling Leaves pairs remain
accepted and were not rerun.
What the fresh original measures
The ordinary registered scr_oasaver, original OASAVER.SCR /s, uses all 88
registered-media/DLL pins without substitution or registry changes. Source is
published main 47095dcb6c84178f63dde8bb61e9fd142aa1f4b9; relevant foreign
working-tree differences are recorded but not applied. The immutable module is
eb2395e1c115980dd78dbb646a26f755a4119a2a1e086942e715240e07c34bc6
(2,408,519 bytes). Deferred actual instantiated byte inventories from page,
guest Worker and D3D render Worker all match that hash. Debugger inventory runs
only after startup and the selected bound capture, never during startup.
| Observation | Fresh result |
|---|---|
| Independent completed presenter PRESENT | Initially 0; first sampled 1 at bind +3,568 ms; then 2 |
| Ordinary native window/composition | One visible Screen Saver HWND 65537, process 1000, composed throughout all 32 samples |
| Native execution/loading | Initial owning-worker yield 5 (load_library), then changing EXE/D3DRM EIPs and remaining slice budgets |
| Initial native UI | Selected whole frame is black while PRESENT is still 0; no visible dialog candidate in sampled window tables |
| Focus/modal limits | Focus 65537 from normal owning-worker slice replies, age unknown; modal HWND/API/loading-completion flag remain unknown |
| Renderer progress | Decoded device records 2 → 36 → 188 during binding; later actual software executor records 10 completed PRESENT and 152 DRAW_UP |
| Binding/capture | Unchanged readiness and geometry guards pass at 4,108 ms; selected whole bound frame shows a colored 3D scene and the known lower-right white rectangle |
| Errors/input | No page exception or bridge worker error observed; no input sent or recorded |
The initial D3DRM loader EIP 0x747f4a96 maps through its logged runtime base
0x747a4000 and original base 0x64780000 to original VA 0x647d0a96, after
an indirect import call. The following EXE samples include 0x7440f5b5,
0x7440f76b and 0x7440f7bc, then D3DRM work around runtime 0x747d51xx.
Selected original-byte disassembly is retained. These establish live native
startup work, not which asset/decoder operation completed at each sample.
Remaining slice steps are a budget, not a cumulative instruction counter.
Sample timestamps mark request start; the native getter reply and subsequent
presenter/composition reads are separate observations, not an atomic clock or
an exact first-frame latency measurement.
The initial same-directory D3DXOF request returns 404 at bind +8 ms and the registered DLL-directory request succeeds at +12 ms. UI/manifest and optional DLL 404s remain visible, unsuppressed. Their presence is not proof of the older startup failure. The current window is already composed before PRESENT, and native loading proceeds to real software rendering. This separates transient initial zero PRESENT from missing composition in this execution. It does not retrospectively identify which predicate failed in either closed predecessor. No native-modal HWND getter is used to infer absence of a modal.
The initial black frame and bound scene are selected startup checkpoints, not a moving-scene pair. No current WebGL original, scene parity, motion-pair, FPS, audio, logo correctness, ordinary native Quit or full27 claim is made. The known authentic-logo/installation-input limitation remains unchanged.
Observer and native proof
The opt-in helper records independent presenter count, native-window table and
current composed membership. One existing guestWorker.readExports operation
per sample reads only reservation-approved EIP, yield and remaining-budget
getters. Focus comes from the existing normal slice reply. Modal yield and
visible owned dialog candidates are recorded separately from unknown modal
HWND and API/loading-completion fields. No reservation guard is waived.
The original 15-second bind, 300-ms stability requirement, no-resize inspection, post-binding ownership/geometry checks and terminal closure remain intact. The passive initial whole-frame observation never establishes an input baseline. If startup geometry changes during it, the image is omitted with both states recorded; the ordinary binding candidate still must satisfy all original guards. Late observation replies cannot append samples, launch a capture or issue a new endpoint request after closure. No early Debugger, forced viewport, rebaseline, missing-asset suppression, fake paint or guest input is introduced. NFS3 startup, Donuts and BloodII loading helpers are untouched.
Selected same-route native authored controls establish:
- Ordinary and observed positives render the expected blue geometry with completed DRAW_UP/PRESENT, identical expected whole-frame pixels, zero input and no page errors; binding takes 4,056 and 3,750 ms respectively.
- Zero-PRESENT control retains composed initial UI and PRESENT 0, then closes at 15,002 ms.
- Native MessageBox control exposes its visible dialog, title and owning HWND; the final safe control closes at 15,012 ms. Its ordinary peek loop can yield 0/15, so yield 6 alone is not a modal detector.
- Hidden-window control completes PRESENT, then separately shows absent composition/visibility and closes at 15,001 ms.
Focused tests cover separate predicates, explicit missing values, the exact
reservation-safe getter group, deferred module inventory cleanup and finite
late-response/capture closure. The locked build passes before publication.
An initial diagnostic capture incorrectly treated changing unbound startup
geometry as terminal; that observer defect was corrected before the original.
An overly narrow authored modal expectation was also corrected. Those failed
controls remain explicit in controls-audit.json. Two authored attempts were
canceled by the independent disk floor during external shared growth; no
original ran in those attempts. Final selected controls and the original ran
above the floor. No failed/closed scope was queried or replayed.
Provenance, cleanup and stop
The observer and focused test were committed together and immediately strict
pulled/pushed. observer-publication.json, original-preparation.json, exact
Git pins, relevant binary patches and original media readbacks distinguish the
actual immutable run from concurrent foreign WIP. Git is the source archive.
The identical native-control/original module is retained once as runtime.wasm;
module-deduplication.json records byte equality and recovery of the original
preparation filename. No source snapshot, media copy, frame series or capture
bundle was made.
The durable run owner uses tools/native-supervise.cjs, an independent finite
watchdog, disk sampling and worker-owned cleanup. Original native and command
waits are exit 0, both log streams reach actual EOF, the independent watchdog
is canceled and awaited at exit 0, and no tracked owned process or server
listener remains. The original's minimum free home disk is 3,138,093,056 bytes.
Earlier authored disk-floor cancellations are retained rather than hidden;
there is no RAM gate. The board watcher returned actual exit 143; its signal
sender is unknown. No box was created: ID null, state NOT_CREATED, expiry
null, stop receipt null. Final artifact verification and cleanup precede
result.json, which is written last. No second original is run because no
applicable native old-fail/new-pass behavior remedy exists. Ownership is
released and work stops.