JezzBall reviewed wall progress, 2026-11-14

Evidence directory: scratch/runs/20261114-jezzball-reviewed-wall-progress/. The task permits one original registered wep16_jezzball default game, after TicTactics and then Quake II browser actually release their runtimes. Result: partial. One original reached natural timeout loss and native Exit, but no wall placement was admitted because the native orientation cursor remained unknown.

The existing registry uses local WEP4 JEZZBALL.EXE (49,664 bytes, SHA-256 0450d0559fa5957d813bdd7be1f71122bf23047869b610802386320f109cb705), WEP4UTIL.DLL (19,408 bytes, 40a405813946caac1aa2ec3e31fa570de1c418f8db43182d614e0bac4e1978cb), JEZZBALL.HLP (17,963 bytes, aab2b95e5eae8348d3aa3e092753f3196d2ac786d4c3fc9cb12d0c638c05fa32) and all 22 registered WAV companions. Exact sizes and SHA-256 for all 25 files are recorded in media-before.json; there is no download or substitute media. The registry's logical-frame declaration is left unused: this task uses ordinary real ticks and default batches of 1,000.

Original rules and lifecycle

The local original Help describes the double-headed wallbuilder arrow: up/down means vertical, left/right horizontal. Left click starts growth; right click changes orientation. Walls grow both ways and must anchor before an atom strikes. A single anchored half may remain after a collision, which costs a life; that partial wall is not a completed partition. With two atoms the second collision ends the game. Time expiry also ends it. Clearing at least 75% advances a level; bonus points are distinct from proof of completion. Scores count cleared squares and isolated empty regions.

The Help's strategy suggests centered vertical walls and dividing atom groups. This is rules information, not a trajectory predictor or a preselected move. Each actual wall choice must come from the current complete scene. original-help-prose.log retains static phrase-expanded original prose with formatting omitted. The existing generic fulltext helper misaddresses this old phrase table; the inspection instead interprets offsets relative to the offset table and two-byte references as ((first-1)*256+second)/2. No parser or runtime source was changed for that inspection.

The released Win16 lifecycle regressions were read before play. They cover owner disabling before WM_INITDIALOG, modal mouse activation/eating, nested dialogs, pre-disabled/absent/destroyed owners, cancellation, EndDialog during init, stack restoration and invalid/stale destruction. The modal destruction fixture also checks guest-owned handle retirement and reuse. The renderer regression preserves queued native-control UP into the modal pump. The existing WEP4 gameplay test proves live red/blue wall growth, not completed walls. It is not run as a substitute original and its old coordinates are not used for this task.

Guarded personal control

tools/jezzball-controller.py derives the released finite-click controller and Barrage identity transport. One bootstrap or active-board review binds the exact PID/start token/window once. Actual active level-one board admission requires personally reviewed rules, orientation, partition and counters. Wall choices require a raw complete scene at most five seconds old; menus allow thirty seconds. Each placement attests its visible orientation, reason, no existing growth, no modal and review of the preceding result. Completion reviews require a persistent partition; animation alone cannot qualify. Area alone cannot qualify as a natural outcome.

The CLI's canvas PNG omits the OS cursor. The owned capture route therefore also reads exactly renderer.canvas.style?.cursor || 'default', the public display output, without reading a guest handle, game variable or memory. It preserves the raw complete canvas unchanged, retaining the current cursor CSS and converting its existing CUR bytes to a separately named PNG with local ImageMagick. This is visual format conversion, with no orientation recognition. An absent cursor remains unknown. Cursor-based personal reviews check the native identity, complete-scene hash, cursor PNG hash and original freshness; the model must inspect both visual surfaces. No cursor image is manufactured. The headless canvas lacks a style object, so a small run-specific preload generated by the published controller wraps its canvas factory and retains canvas.style ||= {}. Existing native SetCursor output can then reach that public display property. This adds no guest calls, input hooks or game settings. A separately tested, single guarded pointer-only hover can expose that cursor before the first wall. Hover requires a fresh five-second complete scene and must produce no mouse-down. It is not a wall or an orientation-changing event.

Every wall and menu event is one finite left click, at most 1,500 ms, with independent UP prearmed at a 1,750 ms fallback. The controller waits for physical release before subsequent capture or audit. UNKNOWN, cancellation, mixed identity, parent/child failure and late ACK end the chain with no retry or rebinding. The 180-second Quit reserve is irreversible. Native Quit and host teardown remain separate evidence. A final native exit before UP can be recorded only with its actual normal wait and both EOFs; physical UP remains unacknowledged.

The production endpoint has no tested right-button guarded semantics. Right clicks, extra button arguments and raw bypasses are refused. No orientation event is enabled. No recognition, ball tracking, private game state, solver, action list, guest clock, configuration or seed change is used. Completion stays modal; the controller has no Next Level route or completion dismissal route. No owner click-through or second level is authorized.

Before the original, 28 policy, 10 HTTP and 12 actual CLI controls passed. These include real consumed-DOWN UNKNOWN, replaced identity, cancellation, late ACK, child death before DOWN, parent death, 1,750 ms independent fallback, capture after UP, public cursor retention and pointer-only hover with zero DOWN. The first actual CLI attempt exposed absent canvas style; its failure is retained. Its parent-death assertion also raced the child's later terminal-record write. Both were corrected, and the final actual CLI run passed all 12 controls. Orphan release-child waits after parent death remain explicitly unknown.

The full locked build passed in 30.157 seconds. Native Win16 WINDOWPOS/focus, modal lifecycle and modal destruction/handle reuse regressions passed. The older renderer modal and button fixtures failed on stale pointer metadata expectations; after an in-memory metadata-only diagnostic, the modal fixture also lacks its fake window table. These auxiliary failures remain failures; no production source or foreign test was changed. One attempted obsolete renderer test filename did not exist and is recorded as an invocation failure.

TicTactics's native/watchdog/supervisor waits were zero, both EOFs and exact PIDs absent, with heavy ownership released. Quake II browser's controller exit was one, Chrome/watchdog/supervisor zero, actual waits and both EOFs, exact PIDs absent and heavy ownership released. Those are termination facts, not game outcomes. Their ancillary outer tool exits 143 and separate UNKNOWN waits remain explicit, without imposing an extra wrapper gate. Source/runtime pins are taken only after these releases.

Original result

The tested route was published as bf1e170a102163f5bc6df89963f4c1a1093a3abc before launch. launch-pins.json records that commit, 420 exact file hashes, three system-tool hashes and relevant-uncommitted.patch (58,821 bytes of relevant shared-tree changes). All 25 registered media files matched their before sizes and SHA-256 afterward; all 420 runtime pins had zero drift. No source, media or WASM copies were staged.

One native process ran with real ticks, default batch 1,000, a 1,200-second limit and independently prearmed 1,220-second watchdog. The complete initial board opened automatically: level one, two red atoms, two lives, score 0, area 0%. It was bound once to PID 4151724/start token 274639749 and the same window. No New Game was needed. Two five-second reviews expired before policy admission; another had insufficient remaining freshness. All were refused before intent/prearm/input. A newly reviewed pointer-only hover was then consumed with its independent release ACK. The public native cursor output still read default, with no orientation image. The active board was therefore never admitted for wall input. No wall DOWN, right-button event, guessed orientation or retry of an uncertain gesture occurred. The specific missing native cursor display path remains unresolved; retaining canvas style alone was insufficient in this original.

The same game continued naturally. natural-timeout.png personally proves Time Left 0 and a modal “Game Over” asking whether to start a new game. This is timeout loss, with no wall collision, zero persistent walls, area 0%, score 0 and lives 2. It is neither first-level completion nor a win. The modal No button was selected once to decline restart. declined.png shows the same window with an ended blank client. The current Game menu was opened and its visible Exit selected once (game-menu.png, native-exit-receipt.json). No second game, reset, second level, Next Level or completion dismissal occurred.

Native Quit is supported by the consumed visible native Exit plus actual normal native exit 0 and both EOFs, not host closure. The native owner exited before the independently scheduled final release: its record is OWNER_EXITED, physical UP ACK false. Earlier hover, No and menu events have actual UP ACKs. Native PID 4151724, watchdog 4151723 and supervisor 4151721 all have actual zero waits, both EOFs and absent exact PIDs. Heavy ownership was released and the HTTP endpoint closed. The watchdog exited on actual native wait without firing. The ancillary invoking tool returned 143; its separate outer wait file is absent and remains UNKNOWN. It does not replace the actual supervisor wait. No host teardown or cancellation was used.

The requested two completed persistent walls and increased area/score were not achieved. Next work requires a separately authorized run after resolving and testing current native wall-orientation display; this task does not start another original. result.json names the five retained complete scene PNGs, compact receipts, waits, media and pin evidence. No box was created (ID/expiry null, state NOT_CREATED, stop not applicable), and no browser, worker or public deployment was used.