Army Men: actual paired BltFast capture

Historical source last updated, 2026-10-20. Runtime and implementation claims below describe that dated source; see current app notes and source for subsequent changes.

One authorized execution on bx_6czhvg6m, 2026-10-20 12:37:56–12:38:10 UTC, accepted eight real pre/post pixel pairs. The single bounded assignment is complete. Stripe causality remains unknown.

Review and exact tested identity

Independently reviewed diagnostic correction 2c076caf4089c602654d1ad5aa4e5264f2af5c76, verified its saved evidence hashes, and extracted its test closure directly from Git. The retained original paired fixture is byte-identical to the failed original capture. The original callsite pushes six words including this; the pinned handler reads flags at ESP+24 and cleans 28 bytes. The diagnostic metadata copy preserves the production table and strict return validation.

All four focused suites passed:

node tools/army-bltfast-observer.test.js
node tools/army-bltfast-captured-abi.test.js
node tools/army-bltfast-preload.test.js
node tools/jig-owning-api-observer.test.js

The corrected fixture fails with the old adapter at assert(r.accepted). The captured-CPU replay still rejects missing pixels; its inferred sixth word is not used as actual capture evidence. Runner trace predicates also passed.

The existing /home/user/icm-fullvm build was retained without rebuilding:

Single execution and actual observations

The prior ordinary Boot Camp input schedule was retained: click (320,180) at batch 279, Numpad6 down674/up706, screenshots at 200/450/670/713 and exit. The runner used 200,000-block batches, 750-batch/90-second bounds, --no-build --quiet-api --no-threads, and --trace-api=__army_bltfast_observer_only__. Exact arguments and controller are retained in recovered/launch.json and recovered/run-capture.js.

It completed 750 batches in 14.258 seconds, exit0, without controller timeout or log cap. The log is 1,468,793 bytes and includes all input acknowledgements. No second execution was made.

Eight selected calls at return address 0x407506 have actual six-word arguments, complete source/destination metadata, pre/post bytes and HRESULT0. All 24 retained raw records have ESP delta28, unchanged dispatch EIP, zero yield reason and valid completion. Owner identity is pinned, pending depth0, no ambiguity/errors, and imports restored. The 663 additional selected calls were dropped at the documented raw-record quota; only eight have pixels.

The same source/destination surface has pointer 0x08221080, slot16, 640×400 dimensions, 8bpp, pitch640, format0 and backing 0x1c1d4000. Copy rectangles start at x2 or3 and y0 or1, end at (640,386), and copy to (0,0). Metadata remains unchanged across each invocation. All 72 sampled eight-pixel spans (576 pixel observations) changed at the destination, and every destination-post span equals its immutable source-pre span. Source-post bytes are also retained. These are actual guest-memory observations, not synthetic replay.

Visual inspection of the real 640×480 captures:

These samples support correct copying at the retained positions for these eight invocations. They do not prove full-surface correctness, correctness of later calls, or the absence/cause of the previously reported stripes. Much of the pre-input screenshot is obscured by the welcome panel, and there is no independent clean reference. Initial corruption remains unknown; no renderer/game fix or broad playability claim follows.

Evidence, recovery and release

Self-contained local evidence: scratch/runs/20261020-army-paired-capture-abi/. All 29 remote artifacts (6,299,676 bytes) were recovered individually and verified by size and SHA-256, without a large archive. Important paths:

report.md
review.json
replay.js
logs/
snapshot/
verification.json
verify-evidence.py
recovery.json
recovered/paired.json
recovered/launch.json
recovered/exit.json
recovered/output.log
recovered/{menu,bootcamp,turn-before,turn-after,final}.png
recovered/{preflight,postflight,overlay,source-manifest,original-manifest}.json
recovered/SHA256SUMS.json

node replay.js replays the source tests/negative control. python3 verify-evidence.py verifies saved artifact hashes, Git source identities, actual ABI/return/pixel assertions, input log and image identities. Neither launches the game. recover.js and remote.js retain the authenticated file-copy recipe; its source remains /home/user/army-paired-capture-abi-evidence on the runtime. Recovery requires an authorized accessible box and should target a new directory if repeated. Ignored evidence is retained locally and remotely, not backed up by Git.

All 170 hashed pre-existing Army/ICM evidence files remain unchanged. Postflight process inventory has no surviving guest/controller. The runtime lease was released at 12:39:31 UTC, leaving the box running for callback qualification. The corrected diagnostic adapter remains in the older runtime directory with its original saved in the evidence directory. No further runtime action follows release.