Original WEP3 Tetravex, 2026-11-14

This bounded task concerns registered wep16_tetravex, selecting existing test/binaries/wep16/WEP3/TETRAVEX.EXE. The community Visual Basic program in tetravex.md is a different executable and supplies no native input authority.

The original NE module imports KERNEL, GDI, USER and WEPUTIL. Static native input code captures the pointer at 1:2a78, moves selected tiles through 1:2b56, and releases capture at 1:2c00. This establishes a drag route; it supplies no current tile values or move choices. The older JavaScript Help text decoder produces damaged phrases, so its output is diagnostic only. The built native Help decoder read all ten topics successfully. Topic 2 explicitly specifies dragging and rejects an illegal destination; topic 3 requires matching every neighbor. Topic 8 provides ordinary Game → Exit. The actual current board is reviewed only in the single original run.

The controller accepts one manually reviewed gesture per complete scene, with a permanent PID/start-token/window binding. Placement reviews expire after five seconds; menu reviews after thirty seconds. Each drag has a personally specified source, destination, visible edges, and every occupied neighbor. An independent process arms UP before DOWN. Normal UP is sent to the chosen destination; independent cleanup always follows even if that UP fails. The total bound is 1500 ms, with a 1750 ms independent fallback. There are no captures, model calls or audit writes while held. Any uncertain, mixed, cancelled, late or identity-changing receipt closes input permanently.

A handler receipt does not establish legal placement. A separate personal review must establish the source vacancy, persistent destination tile, matching touching edges and native acceptance. Completion requires ordinary native completion UI. The final 180 seconds are reserved irreversibly for native Quit. No reset, second puzzle, shuffle, size change, hint or solve route is admitted.

Evidence directory: scratch/runs/20261114-tetravex16-reviewed-puzzle/. Qualification passed before the original: 25 policy tests, 11 HTTP controls, 12 actual CLI controls and the complete locked build. The CLI controls prove native MOVE → DOWN → destination MOVE → destination UP order, the finite release bound, native menu behavior, and refusal to turn a consumed gesture into a native placement claim. Late acknowledgement, UNKNOWN, cancellation, identity changes, child death, parent death, fallback and capture ordering are covered. All 48 fixture native/watchdog/supervisor/outer waits have actual pipe EOF evidence. The killed-parent release child retains a separate unknown wait; its physical release effects are observed by the native fixture.

The initial HTTP late-response fixture mistakenly delayed every gesture edge; its failed log is retained. The final fixture delays the intended DOWN only. No original game was involved in that correction.

The full build uses the shared worktree, including the existing foreign file size-table changes; relevant uncommitted runtime hunks are recorded as Git binary patches with the launch pins. This is not a clean-HEAD closure claim. The controller and tests were published before the original as aeca71687574839b4c719bfe3e13455993fb3c5c.

The single original ran on 2026-11-15 UTC after the Tutstomb and Quake3 predecessors' actual native/watchdog/supervisor waits, EOF, release and PID absence were reviewed. It used the default 3×3 puzzle, real ticks and the default 1000-instruction batch setting, with a 1200-second native limit and an independently prearmed 1220-second watchdog. The initial native window already displayed the board; no Start action was needed. One complete bootstrap/title/board review bound PID 14143, start token 274953602 and HWND 0x10001 permanently. There was no second puzzle or rebinding.

Nine personally chosen drags completed the puzzle. Each choice followed a fresh complete scene and a manual review of its visible numbers and occupied neighbors. The table records those observations, in top/right/bottom/left order; it is a report of completed actions, not an action queue.

Drop Source supply cell Tray cell Edges Matching occupied neighbors
1 r2c1 r1c1 8/6/4/6 none yet
2 r1c1 r1c2 6/9/4/6 left 6
3 r1c3 r1c3 3/6/4/9 left 9
4 r2c2 r2c1 4/1/5/3 above 4
5 r3c2 r2c2 4/4/5/1 above 4, left 1
6 r3c1 r2c3 4/9/6/4 above 4, left 4
7 r3c3 r3c1 5/9/5/2 above 5
8 r1c2 r3c2 5/5/4/9 above 5, left 9
9 r2c3 r3c3 6/8/4/5 above 6, left 5

Drops 1–8 have separate post-UP native persistence reviews: the source was vacant, the tile stayed snapped in its chosen cell, and every occupied adjacent edge matched. Seven of these have occupied matching neighbors. The ninth drop immediately replaced the tray with the native yellow smiley and stopped the timer at 10:19. It therefore has native completion evidence, rather than a fabricated ninth persistent-tile review beneath the completion image. The completion image remained on a later complete capture. All nine destination-UP receipts were acknowledged, with intent-to-UP times between 0.478 and 1.212 seconds; independent release children all actually waited with exit 0 and both EOFs.

Some local reviews expired before any input intent was created; they were refreshed and personally reviewed again. One read-only fourth-placement acceptance submission used different wording for the same neighbor, failed the exact-neighbor assertion, and was corrected from its original review. Neither case sent another game gesture. There was no original UNKNOWN, illegal placement attempt, reset, new game, size change, Arrange, hint or Solve action. No automated tile reader, solver or move generator was used.

After recording completion, the irreversible Quit phase opened the actual Game menu and selected its visible Exit row. Native [Exit] code=0 followed. The final release child observed OWNER_EXITED with actual native wait 0; it does not claim a physical UP acknowledgement after process death. Native PID 14143, watchdog 14142 and supervisor 14140 have actual wait 0, both EOFs, heavy release, absent PIDs and a closed control port. The ancillary tool wrapper returned 143 and remains separate from those actual child waits. This was native Exit, with no host-close fallback.

All 442 exact launch pins matched at postflight before releasing the runtime slot. They include registered EXE/HLP/WEPUTIL, every existing WEP3 sibling, known auto-import DLL candidates including MSVCRT, fonts, the system type library, runtime sources and both built WASM hashes. Native startup logs confirm WEPUTIL loading and eager mounting of the 15196-byte Help file. No media was downloaded or substituted. Exact sizes/SHA-256 values and the relevant shared uncommitted Git binary patch are retained in the run directory.

result.json names five selected raw complete PNGs: initial board, four persistent placements, eight persistent placements, native completion and the native Exit menu. Working capture PNGs were pruned after review; their review metadata and original hashes remain, with the pruning inventory. The run contains compact unique logs and no source-tree archive, media copy or WASM copy. No box was created (ID/expiry null, state NOT_CREATED, stop receipt not applicable). The runtime slot was released for the queued Win16 cursor worker; this bounded puzzle task is finished.