AoWII Device7 observation qualified; original blocked before startup

Run: scratch/runs/20261117-aow2-device7-ground-inputs-render-repair/. No production repair. Terrain, fresh army response and native Quit remain unverified. The single authorized diagnostic original stopped on a required DLL body cancellation before game startup; it was not retried.

Actual Device7 observation

The existing kind15 hook covers Device3 only. Instead of extending that hook, the bounded passive observer now reads the still-live Device7 stack at the existing natural gpu_stream_kick import. It selects return PC 0x4459fe, FVF 0x1c4, a type20 device and its actual 1024x525 target. It snapshots raw vertices, packed vertices from the worker descriptor, the 4096-byte device state, bound texture slot and samples, render states, stage0 states, viewport, and strip triangles' geometric coverage of [600,300]. Geometric inclusion is reported separately from effective raster acceptance. Surface flags now read object+28; object+8 is the device render-target slot or surface misc0.

native-fixture.cjs executes the real WAT handle_IDirect3DDevice7_DrawPrimitive through that exact import, rather than a model or import bridge. Its covering textured strip renders RGB565 f800 at [600,300], with the expected supplied vertices, packed data, bound texture and state captured. A separate target's narrow strip excludes that point and is attributed accordingly. Seven whole-memory hash checks around passive callbacks pass; these checks run only in the fixture, never in the game. Copied receipts remain immutable and disposal restores the import wrapper. check-observer.cjs additionally passes bounds, ownership, lifetime, return/receiver preservation, read-failure isolation and real-error controls. Its stress maximum is 1.65us/hook. The final native fixture records 11 calls, 8740 copied bytes and 9.50ms observer time, including cold snapshot work. This proves observation coverage, not a rendering defect in AoWII.

Matching module and names

reconstruct.cjs reconstructs the requested sources in memory from Git 0a823095cd147c43d362384efd33a312cc2766c7 plus the retained 20261117-aow2-restored-headroom-matched-package-gameplay/final-relevant-source.patch. Every source matches its pinned hash. Recompilation initially differed only by the appended wine-region-layout custom section; all executable sections were already byte-equal. Retaining that exact section reproduces the complete module SHA256 87ef39648a61ed6f54ca4ab8de1926de02c651eb2d99655015dc7e5de5407009. No source tree snapshot was written.

The named fixture is derived from this source closure. The selected native function bodies match the original module byte for byte at their actual indices (native-function-map.json), avoiding the earlier mismatched source mapper. In particular, 7444 is gps_kick, 8393 d3g_kick, 8395 d3g_begin, 8403 d3g_upload_at, 8402 d3g_upload, 8423 d3g_upload_rows, and 8424 d3g_prepare. The predecessor's first captured kick therefore occurred during target preparation/upload while the Device7 call was live; it did not establish completed draw output or its last writer.

One original and the concrete blocker

The driver reused the prior sealed media, module, source adapter, audio and input controls. Offline preflight and the real Device7 qualification passed before launch. The whole original/registered-launch.png was personally reviewed before the ordinary Launch click. No game controls were sent.

original/first-fault.json records the required binaries/dlls/oleaut32.dll request as network-loading-failed, canceled, net::ERR_ABORTED. The strict controller closed with REQUIRED_BODY_UNKNOWN. There is no fresh game module identity, terrain capture or Device7 input receipt. The cause of cancellation remains UNKNOWN; no later query, unchanged retry, intended-color inference or speculative rendering change was made. The previous audio/two-move/native-Quit evidence in 20261117-aow2-required-dll-audio-gameplay-progress/result.json remains reference-only. The preceding run's unverified exit remains unverified.

Shutdown and next work

The independent finite watchdog was ready before the driver, with a 900s deadline and /home floor1500000000B, without a RAM gate. Custody records real waits and both EOFs: driver1, Chrome0, watchdog0, supervisor1 and keeper1. The watchdog did not fire. The supervisor reports clean shutdown with no remaining scope or children. Unique custody files were retrieved and hash-verified before removing the task's own tmpfs profile. No boxes were created: IDnull, stateNOT_CREATED, expirynull, stop receipt not applicable.

Only this findings note is committed; the helpers and evidence remain in the run directory. A canonical production build is not applicable to this documentation-only change. The next useful step is a demonstrated repair of the required-body cancellation boundary before any newly authorized original. The qualified Device7 observer can then establish whether the supplied inputs represent valid application black or an actual API/raster/coverage defect.