Klotski modal bootstrap observation
Evidence: scratch/runs/20261115-klotski-modal-bootstrap/.
The predecessor route 549253e1a and report 01e9657fc remain closed.
That original failed before PNG, binding or input. Its combined assertion
saved neither public main-window equality nor capture HWND. A held physical
button or native defect is unproved. This task never retries that route.
The derivative records each observed operand with request nonce, native PID/start token, request start/end time and observation status before evaluating a predicate. Main-window mismatch, nonzero public capture owner and failed transport are separate evidence. Capture ownership alone does not establish a physical button state. An unresolved controller gesture, including an UNKNOWN UP receipt, prevents observation before any HTTP call.
The main-window equality and zero-capture predicates remain intact. Native qualification did not reproduce either failure. No engine or bootstrap admission relaxation is warranted, and zero fresh originals launched. The predecessor's actual rejected operands remain UNKNOWN. This result does not establish that Klotski gameplay, the first puzzle, or native Quit works.
Heavy work began only after personally checked actual releases from
Sims133024, Win16quitcrash147239 and the subsequently prioritized
Sims-thunk167634. The current module includes the separately published Win16
quit repair 466774bfb; this worker changed no engine files. Relevant foreign
changes remain in the shared checkout and are recorded as Git binary patches.
Actual native observations
A small generated NE executable calls genuine imported USER APIs. It creates a STATIC parent and EDIT child, then USER.1 MessageBox. A second variant opens an unowned MessageBox before creating an application main window. A repeated MessageBox variant proves actual modal destruction and replacement. These are synthetic native fixtures, never original Klotski media or gameplay.
On the current module c775a317 (2202556 bytes), the owned startup observation
was main 0x10001, modal 0x10003, capture HWND 0, all five physical mouse
buttons up. Snapshot/public main equality passed. The unowned startup had
main 0, live modal 0x10001, capture HWND 0, and all buttons up. Its
capture also passed. A main-less modal capture does not authorize permanent
input binding: the existing native guarded-owner admission remains intact.
During a real finite modal gesture, left-button state was 32768 while capture
HWND remained 0. After independently prearmed UP, button state returned to
0 and capture HWND remained 0. Thus zero native capture is not physical-UP
proof either. The first fixture incorrectly expected modal HWND to equal
capture HWND after UP; its failure is retained and explicitly corrected from
the observed operands. It was a fixture assumption, not an engine defect or a
reproduced predecessor failure. No screenshot was taken during the held phase;
the public held observation was persisted after UP.
The derivative now separately requires known physical-button state and refuses any held button before PNG/input. It retains the local unresolved-gesture guard, including refusal of an UNKNOWN UP file. Each physical choice compares the current main/modal live generation with the reviewed scene and refuses pointer coordinates outside the current active modal. This uses public USER identity, geometry and host USER physical-key-state observations; it reads no private puzzle state. Ordinary modal ownership is preserved.
Validation and result
The focused policy suite passed 42 tests; HTTP controls passed thirteen;
actual CLI/native controls passed seventeen. The native controls include owned
and unowned startup, held versus released physical state, capture after UP,
actual modal replacement, stale transport owner, outside-modal refusal,
unresolved own gesture, and a real Win16 parent/EDIT drag. Public EDIT selection
changed after the drag. Separate Notepad controls cover hover, finite
MOVE/DOWN/destination/UP, late and mixed receipts, dropped ACK without replay,
parent death, capture names, occupied destination and the 80 ms lower bound.
Parent-death controls use a test subreaper to obtain an actual wait for the
surviving independent release child; PID absence alone is not used as a wait.
Actual modal replacement changed HWND/generation from 65539/3 to 65542/6
under the same owner 65537; the old scene was refused before intent/input.
The Win16 EDIT drag changed public selection from 0 to 720896.
flock /tmp/wine-build.lock bash tools/build.sh passed under an independently
prearmed deadline watchdog, with actual native/watchdog/supervisor/outer waits
and both EOFs. The built module retained the qualified hash. Installed native
Help also decoded successfully through that module. The 6571-file inventory
includes resolved PATH Node/Python, source, Git patches, mapped Python/native
canvas dependencies, default DLL autoimports, TLBs, fonts and media without
copying them. Postflight found only this worker's expected report edit; source,
module and media had no unexpected drift. There is no original-process mapped
host inventory because no original was launched.
The final recheck additionally recorded the declared Q3 worker's unrelated
light test edit in test/quake3-startup-task-controller.test.cjs. It is not
imported by these controls; runtime source/module/media and this controller
remained unchanged. Both before/after hashes are retained rather than claiming
a zero-drift whole-checkout inventory.
The selected complete captures were personally reviewed: owned MessageBox, unowned MessageBox, ordinary Win16 parent/EDIT, and blue EDIT selection after UP in both the Win16 fixture and Notepad. They are fixture results, not screenshots of a Klotski board. Exact rejected and accepted operands, identity/time, causal controls, source/module pins, actual waits and cleanup are in the ignored run directory.
No native capture defect was proved, no capture predicate was deleted, and no source-engine correction or unchanged original retry was made. There is no new original binding, puzzle selection, complete-board review, drag, native counter reading, persistent placement, exit-block completion, or native Quit. The five-move objective was not attempted because the bootstrap admission correction lacked causal warrant. A future task needs evidence reproducing the actual window/capture inconsistency before authorizing another original.