Golf: one reviewed original deal, 2026-11-14

One untouched default deal reached the native loss “Game over - 2 cards left” after 33 personally verified tableau removals and 16 ordinary stock draws. Native Game → Exit then ended the process with exit code zero. The remaining cards were Q♣ and 10♥. This is a native loss, with no win claim. Evidence: scratch/runs/20261114-golf-reviewed-outcome/result.json.

Route: lib/apps.js registered golf, using the existing local test/binaries/entertainment-pack/golf.exe (23,824 bytes, SHA-256 498937a1379c613a50735ca3ed49e4b09c81596f7540abc92499c981fcd550f9) and cards.dll (157,456 bytes, SHA-256 38caf8318e9a11f1f1d323a4b20544c76a2a6b29827aaf5ea178caed837f19e0). The registry declares exactly that DLL and no extra files, with cooperative runSlice 25000 / compat 100. No media download or replacement is involved.

Rules reviewed before gameplay

The local native EXE is the rules authority. Its static function at 0x010023f0 first checks that the selected tableau column is nonempty (0x010023f9–0x01002426). At 0x0100242b–0x01002478, it masks the waste card's suit bits with 0xfffffffc and rejects rank encoding 0x30, using native string resource 1002, “Can't play card on King”. At 0x0100247d–0x010024bb, it shifts both card encodings right two, takes each remainder modulo 13, and requires the absolute rank difference to equal one. Failure uses resource 1003, “Not adjacent to card on pile”. The acceptance branch starts at 0x010024f4; an actual removal advances the waste and decrements the selected column's exposed-card index.

Thus there is no Ace/King wrap. King blocks subsequent tableau removals; ordinary stock draws remain the way to replace the waste. This inspection does not read live card memory, a deal seed, or any hidden card. Full static disassembly is retained in the run directory. Ranks and legal choices during the original are to be read personally from each fresh complete scene.

Controller and evidence

tools/golf-controller.py supplies one mouse click per invocation, a permanent PID/start/window owner, fresh complete scene reviews (5 seconds for removal or stock, 30 seconds for menus), one optional ordinary bootstrap Start, and an irreversible 180-second Quit reserve. Each click independently prearms a 1750ms fallback UP, accepts a finite pulse at most 1500ms, checks the exact nonce/owner/epoch receipt, and waits for release before audit or capture. UNKNOWN, mixed identity, cancellation, child death and late ACK are terminal. No solver, card recognition, move generation, action queue or private deal inspection is included. Synthetic tests establish transport behavior only.

Run evidence: scratch/runs/20261114-golf-reviewed-outcome/. All 17 policy, 9 HTTP and 10 actual CLI controls passed. These include real endpoint UNKNOWN after a consumed click and dropped reply, cancellation, replaced window identity, late ACK, child death before DOWN, parent death, independent 1750ms fallback, stock click release, and capture only after UP. The initial CLI fixture duration was incompatible with the reserve guard and failed before input; the fixture recipe was corrected to 240 seconds without weakening the guard. Failed fixtures have actual native/watchdog/supervisor waits and both EOFs; their unjoined ancillary wrapper waits remain unknown. The successful rerun joins its wrappers too. Orphan release-child waits in parent-death tests remain explicitly unknown.

SkiFree's native/watchdog/supervisor waits were all zero with both EOFs and released heavy ownership. The GL lifetime original has native exit 1, watchdog/supervisor zero, both EOFs, released heavy ownership and absent exact PIDs; this is termination evidence, not successful GL gameplay. ThreadSentinel 9b8b7f3f was source-ready with explicit no-original. The sole Golf original followed those predecessors, the full doubly locked build (35.408 seconds), and publication of controller/tests at 55b720c9c434789e70424310c0dbda32236d2c78.

Personally reviewed original

The complete default board appeared automatically. No Start, New, Backup, reset, redeal, configuration change or second deal was used. Permanent native PID 4110805/start token 274425434, main HWND 0x10001, and 800×600 scene were bound once from baseline.png. Initially the seven exposed cards were 7♥, A♠, A♣, 5♣, 8♦, 4♠, A♥; waste K♠; stock 16.

Every move below was chosen personally from the current complete pixels and its consequence was personally inspected. This is a retrospective evidence table, not a generated or submitted action list. Column numbers run left to right. “Empty” means no underlying card remained. No covered rank or suit was counted as verified until exposed. All removal choices were adjacent in rank to the visible waste.

Removal Column Card moved to waste Newly exposed
1 6 4♠ 8♠
2 4 5♣ 3♠
3 1 7♥ 9♠
4 5 8♦ 10♣
5 1 9♠ 7♠
6 5 10♣ 3♣
7 6 8♠ 7♦
8 1 7♠ 4♣
9 7 A♥ 8♣
10 4 3♠ 4♥
11 4 4♥ 8♥
12 5 3♣ 2♦
13 5 2♦ Q♣
14 2 A♠ Q♠
15 3 A♣ J♦
16 7 8♣ 2♣
17 6 7♦ Q♥
18 4 8♥ J♠
19 7 2♣ K♥
20 1 4♣ 6♥
21 2 Q♠ 10♦
22 3 J♦ 9♣
23 2 10♦ 5♦
24 3 9♣ 10♠
25 3 10♠ 7♣
26 4 J♠ Empty
27 6 Q♥ 5♥
28 7 K♥ 10♥
29 2 5♦ 6♠
30 2 6♠ Empty
31 6 5♥ Empty
32 1 6♥ Empty
33 3 7♣ Empty

The sixteen personally inspected stock cards, in actual draw order, were 5♠, Q♦, J♣, K♦, 3♥, 6♣, K♣, 9♦, 2♠, 4♦, 2♥, 9♥, A♦, 3♦, J♥, 6♦. Every draw was justified by the then-visible exposed ranks, including the King blocks; no draw skipped an available legal removal. After the last draw the stock disappeared. Subsequent legal removals continued until the native status line in remove33.png explicitly declared “Game over - 2 cards left”. Q♣ and 10♥ were visible; waste 7♣; five tableau columns were empty.

Some personal reviews expired before the five-second gate and were rejected before an intent or input was sent. They did not renew or replay any sent gesture. One rejected early review also supplied an incorrect manual identity; the actual binding was never changed, and subsequent review geometry used the capture's exact identity metadata. All 51 submitted clicks had distinct nonces and independently prearmed UP; 49 gameplay clicks, then Game menu and Exit. Earlier clicks used 400ms, later ordinary clicks 120ms. No live input fault or UNKNOWN occurred in the original.

Native shutdown and limits

The current complete native Game dropdown visibly offered Exit in openquit.png; the reviewed Exit click was consumed. The native process exited normally before final UP. That release is OWNER_EXITED with physicalUpACK false, verified against the exact native wait; no physical UP acknowledgement is fabricated. Native Exit via visible UI is separate from host teardown; no host-close command was used in the original.

Native PID 4110805, watchdog 4110804, supervisor 4110801 and outer caller 4110800 all have actual zero waits, both EOFs and absent exact PIDs. The watchdog ended unfired; the control port closed and heavy ownership released. The CLI reported 418.488 seconds. One held main critical section was logged at exit, so this is not a claim that all thread resources were clean. All 336 runtime/media/read-only pins matched at release. Git SHA and relevant binary patches record source provenance; no source tree, media or WASM copies were archived. Ten selected, individually named original PNGs are retained; 65 unselected original/fixture captures were pruned with hashes and metadata. No frame bundle, automated board reader, solver, private-card access, browser, worker, deployment or new box was used. Box ID/expiry are null; state NOT_CREATED; stop receipt not applicable. The bounded task ends here.