Screen Savers fixed candidate: independent source-slice acceptance
Historical source last updated, 2026-10-20. Runtime and implementation claims below describe that dated source; see current app notes and source for subsequent changes.
Reviewed candidate 95d2f43d868d60d84d945b20f393863176622d64 against main
c918629780ca97b871886e417fc46a800ab3203b on 2026-10-20.
Accept the repaired renderer-to-native input source slice. No remaining
blocking production defect was found in this bounded review. Proceeding to
full-VM/browser validation is justified; this is not full lifecycle acceptance
or permission to skip integration/build gates. No production changes or main
merge were made.
The retained clean wine-screensaver-input-boundary-fix was read-only. Original review and fixed evidence were read without repeating the recovery
search; all 8 and 11 respective manifest entries verified. Evidence is in
scratch/runs/20261020-screensaver-fixed-independent-review. A durable copy of
the tests, identities, reconciliation and results accompanies this report in
screensaver-fixed-independent-review-20261020-evidence/.
Input ownership and control behavior
The applet now uses the existing host check_input owned dequeue, followed by
check_input_hwnd with live focus and separate full-width wParam/lParam reads. The browser host prefers renderer process identity, falling back to the
process HWND range/instance policy. The test puts a foreign-process event
ahead of owned input and proves it remains queued while owned input proceeds. It covers that ownership branch, not all Worker/thread ownership variants.
Actual renderer handleMouseDown, handleMouseUp, _dispatchMouseEvent and
takeInput feed the production host input block. The compiled closure runs
the applet, native dispatch, BUTTON/LISTBOX/STATIC state and wndprocs,
notifications, mouse activation, capture and USER focus helpers. It observes
list row 2, matching STATIC file text, and Preview of ARCHITEC.SCR only on
release. It does not replace control input with a synthesized WM_COMMAND. The small applet fixture still uses direct post injection and is only
supporting coverage.
The unchanged original candidate 4c98907328917f5917a59445bb4d5ce3e2caca09
fails the new regression at renderer DOWN selects real listbox row 2. All four mutations fail at specific behavior assertions: omitted poll,
wrong HWND, omitted mouse activation, and dropped native button notification. Compilation failures or unrelated exceptions cannot satisfy these oracles. The original applet regression and its three controls also pass, as does
the host wiring fixture with its six failure/cancellation cases.
The pump snapshots event fields, constructs the MSG tail, performs USER mouse activation, then sends accepted input to the native target. Work is bounded to 64 input events; the 65th remains queued. Renderer release deferral is honored and the host schedules another step after 50 ms. Independent checks enqueue DOWN and UP before any pump: the first pump acquires capture without launching, the next release launches once and releases capture. Original outside-release and WM_CANCELMODE cases pass.
Fixture focus gap found and addressed in review evidence
The supplement changes only this adapter to return 1 and adds wrapper exports for existing production USER focus/capture helpers, in memory. No production form or retained file changes. It then verifies real focus transfer, keyboard DOWN routed using live list focus, cancellation of a mouse-origin button press through USER focus loss, and cancellation through real capture transfer with the new capture owner preserved. All pass. The keyboard event is queued directly; this does not qualify DOM keyboard handling or dialog Tab navigation. The original fixture gap is a coverage correction, not evidence of a production owner-thread bug.
Close and lifecycle boundary
Native BN_CLICKED calls the class-35 parent synchronously through
ctrl_notify_parent/wnd_send_message. The input pump checks scrsave_hwnd
after dispatch and returns immediately if Close cleared it; the post path
also stops after Close. Both original native Close and the independent queued
WM_CLOSE test prove one MSG free and no subsequent dequeue, even with another
event queued behind Close. Additional pump calls return zero.
The source clears applet globals before destruction, recursively sends native
WM_DESTROY through wnd_destroy_tree, removes records, calls host window
destruction and sets quit. The fixture's recursive removal merely deletes
map entries; its free/read guards are not a production allocator/UAF oracle. Production destruction, HWND focus/capture cleanup, host stop/unwind and shared
renderer cleanup remain full-VM/browser obligations. The separate
wnd_destroy_recursive path must not be mistaken for the applet's tested
wnd_destroy_tree path.
Hit testing, geometry, window-record storage/owner metadata, allocation, painting/GDI, default-button border work, recursive teardown, and shell execution remain adapters. The launch array proves the selected filename reaches ShellExecute, not that a saver starts or survives applet Close. The inherited bounded temporary-caption/catalog allocations, low-memory construction behavior, asynchronous startup cancellation/timer cleanup and single-app queued launch semantics identified by the original review remain explicit limitations; this input repair does not resolve or qualify them.
Reconciliation with current main
Common ancestor is 4d57bb66dae7da0a488b9546ccb0b4f4f7ba2434. The only changed path common to both sides is host.js. Main adds MIDI
callback transport/dispatch wiring in the wave callback and audio setup areas. Candidate adds no-PE startup/Worker suppression, ScreenSaver shell resolution,
font/startup cancellation checks and the applet run branch. The changes are
in separate functions/locations; the owned-input import block is byte-identical
on candidate and main. Native input/control/USER source dependencies are
unchanged by main's MIDI work. Main also changes audio/register/API/export
sources and generated import signatures, while candidate changes applet,
regions/manifest/native class dispatch and registry/shell paths.
Exact replay and next qualification
- Unmodified input closure: 23,280 bytes,
980e3a0cf223f2543d9152c21d0815abfbc9288ba2fe76458b79fd59dfc9d100. - Original defective applet closure in the same regression: 23,034 bytes,
7584941562a3498aee57180cff62e574c69eb268266fbdc3e9a4215bf18cec39. - Supplemental closure with three USER wrapper exports: 23,355 bytes,
b61e581ff93617a4466f5dac8e86daba74af53be997d8ee39123cbdabbcae7b8.
Next validation should use the reconciled full module: cold no-PE startup and real asynchronous cancellation; actual renderer hit testing, scrolling to later entries and visible press/release/cancel/focus behavior; native recursive teardown and all close routes; shell resolution and registered saver arguments; Preview launching a separate saver, followed by applet Close while the saver continues, explicitly recording desktop versus single-app mode. Actual DOM input, shared renderer ownership, GDI repaint and resource cleanup must be observed rather than inferred from this fixture.