Pawn second native turn: reproduced, commitment witness blocked
The second native reply remains UNPROVED. One new original reproduced 1.e2–e4 Ng8–f6, followed by displayed white d4/e4 and unchanged black Nf6. The focused observations establish delivered mouse transitions and a responsive native UI, but do not establish whether Pawn committed d2–d4 internally. No production defect or corrective patch is claimed.
Parent 0bfde3d97 and its report, starting/reply/stalled captures, native trace,
input helper and idle-loop evidence were reviewed first. All 67 sealed parent
files verified. Initial run 20261113-pawn-second-turn-transition-0810 froze
the descriptor/probe, then refused before watchdog or native launch at
1,498,787,840 available bytes, below the strictly-above-1,500,000,000 floor.
That refusal was published as fdcab4085; its evidence remains intact.
Before handoff, a distinct coordinator outcome arrived: completed EXACT8 closed
log recovery, with recorded raw hashes, restore checks and a fresh above-floor
sample. Its actual board entry is retained in the new run's
dependency-outcome.json. A separately frozen decision in
scratch/runs/20261113-pawn-second-turn-transition-0810-recovered/ then passed;
fresh prelaunch availability was 1,510,834,176 bytes. This was a new decision
after an observed resource change, without replaying the refused descriptor,
floor polling, RAM gating or foreign cleanup.
The independent observable 600-second watchdog was armed before native
PID960375. The unchanged qualified actual-owner guard accepted the legacy
software Bridge WorkerConsumer, worker thread3, device136454184/generation1,
instance, memory buffer and window. The original used the same current04
module and six media. Existing --trace-input-dispatch, --trace-input and
--trace-mouse-state observers and read-only turn-probe.js were frozen before
launch. The probe reads owner/window geometry, mouse mask/queue, guest
EIP/yield/post queue/capture/sleep/wait getters and render completion state;
it invokes no setter, input poll, callback, Present or scheduler operation.
The full starting board was personally reviewed before e2–e4. The full reply board was separately reviewed before legal d2–d4. Each finite drag retained independent mouseUP, actual wait0 and separate stdout/stderr EOFs. Fresh disk checks passed before both moves. No further chess move was sent.
The new trace closes the earlier dispatch observation gap:
- Both down/up pairs reached DispatchMessageA HWND0x10002 with correct WM_LBUTTONDOWN/wParam1 and WM_LBUTTONUP/wParam0.
- Host mouse polling observed each drag's actual coordinates, held mask1 and
final mask0. The second UP dispatched at batch137523, client
lParam
0x00c6009a, followed by a polled(508,340,buttons0)state. - At batch149790, EIP
0x40ca3ffollowed Sleep(10); yield0, quit0, post queue0, mouse mask0/input queue0, accepted owner, render submitted=consumed=completed 1584, entries0/pending0. The displayed lone c3 outline is not reply evidence.
Ordinary Game menu and Game > Save opened successfully, demonstrating a
responsive native UI after the stall. Save As was personally reviewed as a
potential authentic turn witness. Before focusing its filename field, the
fresh disk gate refused 1,496,494,080 bytes. No filename or save input
occurred. The helper's generic filename-focus-UNKNOWN.json records this
known-unsent refusal; it is not an uncertain delivered chess action. The
helper promptly requested normal quit. Internal move commitment, native search
state and the causal turn-transition diagnosis therefore remain unresolved.
One run-helper correction is retained: the first menu attempt bundled two
commands in a single cmd; its ACK proves only mousemove was scheduled, with
no down. The corrected menu helper used separate ACKed move/down commands and
a new claim. It did not replay uncertain input or change production behavior.
Native shutdown after 150.085 seconds/474464 batches and the independent watchdog both waited0 with separate EOFs. All five independent UP processes also waited0 with separate EOFs; all seven owned PIDs are absent. The closed 9,367,156-byte stdout log was losslessly compressed to 35,472 bytes after shutdown, with full decoded-byte comparison, raw/compressed hashes and restore metadata/recipe. This reclaimed 9,334,784 allocated bytes from this run's own telemetry. The alternating Save As window-rect trace accounts for most of that volume; the dispatch trace itself is bounded. No original was retried afterward.
All 320 current source/module/media pins and frozen helpers passed postchecks;
eight helper syntax checks passed. Module SHA-256 remains
04c534192ee3fe39e388536c2ce710c12c47a5cd2ca4d95c19f9d00b1e8f0a17.
Evidence preserves exact Git SHAs, binary source diffs, inherited untracked
source and owned-helper patches, captures and media hashes. No source/media
archive or repeated media copy was made. This is documentation-only; no
production build is claimed. Protected NFS, backup and LMarbles data were untouched.
Next: sustainable headroom through a fresh guarded original and native history/save witness. Observe whether the second move is committed before attributing the stall to input, computation or blocking. Keep the parent's and this run's second reply unproved; do not replay either uncertain action.
Box ID none, state NOT_CREATED, expiry null, stop receipt N/A.