Stronghold natural setup cause, 2026-11-05
The actual setup transition does not establish an input/runtime defect. The concrete progression dependency is an observation transaction that the current route cannot receive: an owning native whole-cycle checkpoint, the actual selected-primary fallback copy and its layer-to-display publication lineage. Adding only an RPC operation ID cannot repair this case. The existing linker has zero native surface intervals and zero identified renderer members to join. No recognizer, input, guest state, clock, cap or readiness policy is changed. Gameplay remains unverified and no original/browser retry is authorized here.
This investigation uses published 331c216924d16740884620b44b040755480a926c,
its actual run report, and
the existing native draw, control-state, render-completion and presentation
notes. Evidence is in
scratch/runs/20261105-stronghold-natural-setup-cause/ (E below).
Retained transition and original control flow
inspection.json binds the existing executable, native report, host report,
command-15 failure, captures and current source by path and SHA-256. It reads
retained evidence and PE bytes only. It does not execute the guest or fabricate
historical native events. Eleven instruction anchors are checked directly.
The three completed original draws are:
| Entry → return | Entry time minus saved timestamp | Refresh at entry | Fade entry → after update | Suppression entry → return |
|---|---|---|---|---|
| 43 → 59 | 9088 ms | 1 | 32 → 32 | 0 → 0 |
| 76 → 90 | 688 ms | 0 | 32 → 15 | 0 → 1 |
| 108 → 122 | 3536 ms | 0 | 15 → 0 | 0 → 1 |
The first draw's refresh branch skips fade advancement. The next two use the
original signed arithmetic: 32-trunc(688/40)=15, then
max(0,15-trunc(3536/40))=0. The refresh-zero bypass preserves timestamp
169132 after the first draw's late clock. Thus the earlier possible mechanism
of repeated sub-40-ms gaps losing residual time is ruled out for these actual
updates. There is no evidence here for changing time, scheduling or fade math.
Original block 429313 tests refresh and direction before 429321 writes
suppression 1. The negative clamp at 429358 clears fade and direction, not
suppression. Consequently a complete fade-zero draw that returns suppressed is
expected original behavior. The main-loop store at 576c56 clears suppression
separately; a later zero is a per-cycle observation, not a draw-completion fence.
A full-rest host ACK cannot perform that native store.
The premise that actual numbered selection changed is not supported. Every one
of the 140 retained samples has setup manager [692f48+38]==0; pending69165c
also stays zero in those samples. The control observer's text “unknown; no
matching setup selection” means no sampled selection, not a foreign selection
fault. Its native fault is null. No preset activation, queue or application was
recorded. There is no absence claim between sampled records.
The source also reveals an ordering dependency not captured by the current
observer. The input gate at 576b9a–576bc0 reads the preceding cycle's
suppression before the reset at 576c56. With fb4694/efe4b4 zero it calls
4e95f0 only when suppression is zero, unless alternate input flag
[1a431a4] permits the call. Otherwise it skips to 576bc5. Thus fade reaching
zero inside a later draw does not retroactively admit that cycle's earlier
input traversal. These three gate operands are not retained in this actual
report; its precise input branch cannot be recovered. New observation must
sample the executed gate/successor and manager pointer [150978c], rather than
interpret null selection as a broken gesture or a disabled preset.
The source has distinct ordinary selection operations: manager routine
4e95f0 clears +38 before input traversal; 4e97a3–4e97b7 writes a child's
address only after its separate field/global gates. Widget-local disabled16
+38 is another object's field. Neither a null manager selection nor the
suppression-induced zero feedback warrants inventing a selected preset or
changing input handling. Hover routine returns remain description callbacks,
not activation or queued action.
The final draw returned at boundary40820404 to 4e91cf; numbered widget draws
subsequently returned at40842013,40847488,40852951. The final ACK/capture is at
41586629, with capture PC 44fd7b and last logged PC 44fd63, inside the
original RGB565 sprite pixel loop. No fourth scene draw was entered. In this
run the terminal pending array is empty because the observed callback intervals
have returned, not because the entire native scene/render pass completed.
The existing observer does not cover that whole pass or retain the final pixel
loop's actual caller chain. It cannot identify that loop's owner from PC alone.
This differs from the predecessor run's still-pending fourth scene draw.
The setup draw is one callback among the manager calls at
4e91b6, 4e91c9, 4e91cc; widget traversal and subsequent graphics work are
separate. The main loop calls the scene manager at 576db6 -> 4e9180, then
widget traversal at 576dcf and 576ded, and surface presentation at 576e30. Source reachability does not prove the
retained run executed those unobserved callsites. A new observation must cover actual
entry, saved caller/ESP, scene manager, whole-pass return and suppression-reset
history, rather than assign all later pixel work to the preceding draw.
Why the actual route rejects
I personally viewed final actual route-native-boundary-10.png and the
historically recognized setup reference. The actual image is still translucent
setup artwork with an empty map area, unavailable text, numbered circles and
partially drawn player/configuration panels. The historical reference has a
populated map and opaque panels. Neither is gameplay.
The audit compares every opaque pixel of the fixed game rectangle, without alignment, masks, fitted thresholds or derived screenshots. The final actual capture differs from the recognized historical setup at395139/574480 pixels, including124223 Human/Lord pixels,49449 map pixels,69461 Deathmatch/gold pixels, 47024 player-list pixels and91 numbered-button pixels. This is not one tooltip or hover discrepancy. The preceding actual transition differs at397531 pixels.
The actual image classifier returns unknown, and
diagnosticScene(bytes, known) returns unknown too. Its complete inputs are
PNG bytes and the image classifier's label. Native suppression, selection,
completed draw and host relation are not parameters to this decision. The
current attachment persists its native report and invokes the passive host
linker during capture, then returns no native observation to the route. The
first unknown-scene failure therefore is an image/dataflow decision, not a
native suppression/selection rejection. Observer-only sampling cannot change
that decision. The immutable command-15 failure must remain terminal; it cannot
be resumed to test a prospective correction.
The missing host relation is more than RPC identity
The actual page report contains four successful upload/fallback pairs for
surface id 0x200006, canonical storage474234880, HWND65537, presentation2,
canvas3/window4, versions1372/1374/1376/1378. All37 renderer completion records
have empty draws. There are zero native surface-* records in the actual
140-row native report. Host ownership and upload success are observed facts;
the selected fallback destination and its completed display membership are not.
Current native collection is scoped to four conditional guest Blt CALLs inside
467b90. The join starts by finding a native surface-import-request within
one such interval. Therefore this actual history cannot produce even a
structural candidate. A new cross-realm RPC identifier alone leaves that empty
set unchanged.
The production fallback path is materially different. _presentDxIfDirty calls
presentBestDxOffscreen(true), which chooses the primary/offscreen source and
copies its memory in _presentDxSurfaceToMainWindow. The ordinary window path
writes to _dxFrameLayer.ctx when that layer exists, otherwise its back canvas.
The renderer then composites the layer. The observer instead accepts only a
canonical presentation's exact p.canvas passed to renderer.ctx.drawImage.
Its fallback completion explicitly has surfaceRelation:null; it records
neither the selected source/slot generation nor the actual copy destination.
Intermediate canvases provide additional unobserved
paths. A completed repaint and a canonical upload cannot fill these gaps.
The host deliberately allows a retained video-memory pointer to be displayed after three deferred lock checks. That source policy is not a guest whole-frame fence. The retained evidence does not include actual lock state at final copy, so it would be incorrect to assert that this particular capture crossed that branch, or to “fix” it by disabling normal retained-pointer presentation.
The new dependency fixture
executes on both shipped modules. It uses the intact original input gate and
reset windows, 32 independent flag/fade cases per module, and an actual active
4e95f0 CALL parked inside a synthetic callee. While that Stronghold CALL is
active, the newly shipped MDK2 ledger reports valid1/depth0. This is a concrete
counterexample to using that ledger as Stronghold pass-absence proof. The
separate original reset clears suppression; fade zero alone does not admit
input through the gate.
The composed arms execute synthetic native REP STOSW pixel writes, actual
WineAssembly.getImports() and _presentNow(), the real dirty/lock policy,
production host imports, renderer and observer. Windowed and exclusive modes
run with and without observation on each module. Synthetic inputs are the
surface/process table, window, presentation-HWND getter facade, broker and
paint program. Existing test/binaries/notepad.exe is loaded only for address
mapping; its application entry is never executed. This is not a game retry or
a reconstruction of the original capture's lock state.
One canonical upload precedes red, half-blue/half-red and blue publications.
The half-written native primary is published on the actual policy's third
held-lock check; the first two return0. Independent RGBA oracles check the
mixed frame and completed blue frame. All three layer/display hashes differ,
while the canonical upload version stays fixed and its canvas differs from
the actual fallback destination. The exclusive arm has empty identified
renderer members; the windowed arm identifies the old canonical backing only.
Fallback completion still supplies surfaceRelation:null in both modes.
Normal completion and canonical membership therefore cannot substitute for the
missing selected-copy/layer receipt. Observed/unobserved publications and
pixels are identical; pixel acceptance and readiness remain false.
The first two fixture attempts incorrectly required empty renderer members in the windowed mode. They failed and are preserved in E. The fixture now models both genuine compositor modes and preserves the canonical positive. A third fixture attempt correctly rejected its incomplete synthetic HWND setup; only the explicitly declared presentation-HWND facade was restored. No production code or historical evidence was changed to make a fixture pass.
The existing native-draw suite passes 20 draw cases and 30 rejection controls per module, with unobserved/trace/observed full-memory, register and command parity, plus 24 host structural-join negatives and block-forwarding checks. The full locked build passes and reproduces the exact tested module bytes. No production Stronghold runtime, recognizer or immutable original-run binding is amended. The native predecessor's MDK2-specific binding does not authorize a new Stronghold original on this module pair.
The newly shipped native pending_context_version()==1 primitive does not
supply this Stronghold proof. Its fixed relevant target set is MDK2-specific,
including 4522d0/443070 and the other MDK2 contexts; it omits Stronghold's
4e95f0, 4e9180 and 4e9400. Its valid zero depth is not absence evidence
for those passes. A future extension must bind an explicit Stronghold target
profile and original image/owner epoch before execution, preserve native tier
and callback/unwind completeness, and invalidate gaps. The historical 64-return
ring and a startup stack dump are also insufficient.
Precise next implementation dependency
The required observation transaction has three connected parts:
- A continuous owning native whole-cycle receipt, including the scene20 draw entry/return, subsequent manager/widget pass and ordinary suppression reset, including the input gate/successor and its actual manager, plus exact capture/ACK epoch, step and saved caller/ESP. No current terminal report may become a live prefix. If coverage is missing, return unknown.
- A passive receipt emitted at the actual fallback copy, identifying its selected surface allocation/slot, memory/storage range, destination frame layer or canvas, copy generation and any in-progress native write interval. A synthetic RPC upload result, nearest draw or same storage address is not that receipt. The ordinary fallback and guest-originated upload paths need separate provenance. No new forced present, wait or guest write is needed.
- The real compositor's source/layer lineage through normal publication, bound to that copy generation and the capture. Only then can capture adjudication consume the native observation transaction. A diagnostic native setup label must remain distinct from pixel acceptance, native input readiness and gameplay; unrelated/foreign/stale/partial/terminal/cap-exhausted cases reject.
The observation may explicitly describe a proven owning pass still in progress for a diagnostic setup-transition label. That does not assert a completed return, input admission, pixel acceptance or gameplay. A pending callback by itself, without continuous owner/copy/publication provenance, is insufficient. This distinction matters here: the image guard stops during later pixel work, before the missing whole-pass receipt could ever become a completion claim.
A completed ordinary unsuppressed fade-zero draw, identified ordinary host publication, fresh personally reviewed controls and actual application/world proof remain requirements for gameplay. This dependency grants none of them. There is no warranted direct input/runtime correction and no screenshot whitelist, increased cap, hover/pending-draw readiness or another original.
Validation, publication and recovery
Shared edits/tests/build began only after D3Dqueued0432 FINAL f1dc763eb then
MDKnative0505 FINAL 036088ef12eb90aed902efa67545b758d5c70ecc. E's
release-gate.json records fresh independent normal exit receipts, exact
PID/startTicks/profile absence, refused ports, no heavy process, free runtime
and build locks and sufficient disk. All final own fixture/build processes
exit0 normally; no browser/server/box was started.
Reproduce from repository root using existing inputs, without copying media or modules:
STRONGHOLD_DEPENDENCY_RESULT=scratch/runs/20261105-stronghold-natural-setup-cause/dependency.json node test/test-stronghold-natural-setup-dependency.js
STRONGHOLD_CALLBACK_RUN=scratch/runs/20261105-stronghold-natural-setup-cause/native STRONGHOLD_CALLBACK_RESULT=scratch/runs/20261105-stronghold-natural-setup-cause/native-draw.json node test/test-stronghold-owning-callback-observer.js --draw-only
flock /tmp/wine-build.lock bash tools/build.sh
Tested source: 036088ef12eb90aed902efa67545b758d5c70ecc plus the two
owned new-file patches in E. inspection.json records existing EXE,
retained reports/captures and relevant source paths/hashes. dependency.json
records the mapping-only PE and native paint program hashes.
relevant-uncommitted.patch and owned-new-files.patch preserve relevant
uncommitted/new source; Git is the source archive. Logs, JSON and receipts stay
in E and are not committed. publication.json records the own-path commit,
immediate strict pinned-SSH pull/push and remote main identity.
Final build identities (byte-identical to tested inputs):
build/wine-assembly.wasm:8605f6c4800e0286e9d863c599618d7d97cde67936f4b4e6c1528897576aa611build/wine-assembly.compat.wasm:ac64ace417f0af22b904213bf2f739b97a2132f7f631c261eccd855fae18c970
The actual original's input-gate branch, final caller chain, selected fallback copy and completed whole-pass publication remain unknown. Those are the next concrete observation dependencies; no native/input malfunction or gameplay has been established. Box ID/state/expiry: NA. Coordinator retains shutdown ownership; there is no test box to stop.