Stronghold Custom Game drain capability, 2026-11-04
No new original was launched and no causal repair is claimed. The offline capability gate fails: a valid execution of the original renderer can exceed the unchanged observer boundary cap before returning. The exact cause of the historical DOWN drain remains unknown because its intermediate queue state and native render arguments were not retained. Repeating that original would not meet this assignment's prerequisite.
Input commit: 9707f95251b9ba1c1d25cee3bf22a8d2085c8792. See the
preceding original result.
Evidence: scratch/runs/20261104-stronghold-customgame-drain-cause/.
Actual producer and retained execution
audit.cjs reads the sealed route and applies the production composeRoute,
then executes its actual PhysicalRoute.drain method with the actual readiness
and ownership predicates. An empty queue with the owned active/last DOWN event
still waits. After the next fixture poll supplies explicit null/null, it
returns with held mask 1. Missing telemetry, foreign input and wrong mask
reject before another step. There is no demonstrated mask/default-argument
or return-value defect in this producer.
The production host.js input body is also executed with the exact retained
DOWN entries. It dequeues 132 then 513, retains the last/active 513 after the
queue empties, and clears both on an empty poll. The fixture dequeue is a
minimal queue adapter, not an original guest message pump. It proves this host
receipt lifecycle, not WndProc dispatch or callback completion. In the actual
Worker transport, per-slot event save/restore also preserves follow-up message
arguments; a host empty queue cannot certify guest dispatch.
The retained observer's 14 successful post-DOWN ACKs sample these PCs:
| Steps | Observer blocks | Native PC |
|---|---|---|
| 7872 | 163000 | 0x4554a5 |
| 7936–8128 | 227000–419000 | 0x4554d7, 0x45550a, 0x4554bf |
| 8192–8384 | 483000–675000 | 0x44fd80, 0x44fe0c, 0x44fe38, 0x44fd88 |
| 8448–8704 | 739000–995000 | 0x44f7d8, 0x44f7e6, 0x44f7dd, 0x44f810 |
These are native rendering loops, not callback 0x430e10. No native callback
entry or pending saved-return interval was recorded. ACK annotations contain
no queue/active/last fields. The final queue was sampled after cleanup UP;
it cannot establish when DOWN first became eligible for the drain predicate.
The next frozen advance to 8768 failed its observer check and is not a
successful ACK.
set_trace_esp(1,0,0) arms all native block entries. In src/13-exports.wat,
dbg_recompute sets both chain and tier guards for that flag, requiring blocks
to return through the run loop and preventing micro-op loops from hiding
entries. A host 64-step ACK is therefore neither 64 native blocks nor a
completed message/frame. The retained increments of 64000 boundaries per ACK
are consistent with this distinction; they do not prove input dispatch.
Executed capacity counterexample
native-render-capability.cjs loads the exact original EXE in each existing
normal/compatibility module and executes the unmodified function 0x44f6e0.
Only fixture caller bytes, stack and initialized data are supplied. The
800×600 RGB565 surface contains nonzero pixels; arguments are
(0,0,799,599,32), with deterministic lookup data. No native function is
patched, no callback is invoked or fabricated, and no scene is synthesized.
These fixture writes are offline setup, never writes to an original runtime.
Both modules execute 1,441,804 entries inside the original render function.
The complete fixture has 1,441,856 entries including its caller and stop-loop
padding. The actual observer is attached before instantiation and started
before execution. It fails at 1,000,001, while the function is still in its
pixel loop. The fixture sees that fault after its enclosing 64-block run
returns, at total 1,000,064 and PC 0x44f7d8.
The remaining execution is continued only in this offline parity experiment; the production driver still stops on the first fault. Full 512 MB page hashes, all sampled general registers and lazy flags, surface output, native PC histogram and every run-command result exactly match the trace-armed control. Neither arm has an input queue or game command stream: this is an isolated renderer capacity proof, not input-to-callback or gameplay evidence.
The loop compares X at 0x44f7d8, tests a pixel at 0x44f7dd, converts a
nonzero pixel at 0x44f7e6, increments at 0x44f810 and advances rows at
0x44f818. Its static caller at 0x469a43 passes runtime rectangle fields
after clipping. Those fields were not captured in the failed original.
Consequently this counterexample proves that the current capability contract
can reject valid rendering; it does not prove that the historical call used
the fixture rectangle or fully explain its DOWN-drain failure.
Validation and stopping condition
Both-module executable owning-Worker callback/branch/scene/setup fixtures, pre-instantiation binding negatives, actual attached-driver composition, local command/review contracts and route recognition checks pass. Test output roots are redirected in memory to this run; retained helper inputs remain at their original paths. An initial test-output adapter mistakenly redirected the command-client input too; correcting that adapter restored the passing local tests. This was an offline harness error, not a product repair.
flock /tmp/wine-build.lock bash tools/build.sh passes. Normal and compatibility
module hashes remain respectively
78a0b1d61b23428a1ae909d739440ae0de059bbbd84a69a2ad3e4c37471ad664 and
99b579cd17b57f94145df92ae6d28110f24fc284d13a1b2c0ebe3596bd10a77f.
EXE SHA remains
9024c533095f2227b0105b7e8a9e4cf28ec4fb206b165629787b2c97cf99f13f.
All 554, 418 and 523 predecessor sealed artifacts match their existing seals.
Source recovery uses Git SHA plus relevant binary diff/new-note patch;
media/modules are existing-path/hash references, with no source/media copies.
No old-fail/new-pass causal repair exists here. A complete genuine input/control-flow capability within the original bounds has not been demonstrated. Resetting, enlarging or dropping boundary history, skipping drainage, changing pacing or granting callback/scene readiness would not repair that missing proof and was not done. No engine change is justified by the missing historical native state. The task stops on this concrete capability blocker, with setup/preset/Start/world/gameplay unreached.
There is no new Chrome, server, profile, local-runtime flock or assigned box. Box ID/state/expiry are NA; the coordinator retains lifecycle ownership.