Screen Savers recovery source review — rejected
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.
Blocking finding: ordinary control input is never consumed
P1: the recovered no-PE pump is incompatible with current renderer input.
src/09ca-screensavers.wat:397 drains WAT posted messages and paints, explicitly
omitting host input polling. host.js:6791 calls only that pump and repaint. However, current lib/renderer-input.js:1300 _dispatchMouseEvent queues events
in renderer.inputQueue; it no longer synchronously sends them to a native
window. The ordinary-window child path (:2203 onward) uses that function. The applet is created through host_create_window / renderer.createWindow,
not the dialog registration path. Its frame has no isDialog, isFindDialog
or isAboutDialog flag, so the direct dialog route at :2111 is inapplicable. Neither its listbox (class 4) nor buttons (class 1) qualify for the combobox or
ThunderCommandButton exceptions. host.js:2523 exposes the queue consumer as
check_input, but the applet never calls it.
Minimal retained reproduction:
node scratch/runs/20261020-screensaver-source-review/reproduce.js
The script reads the exact candidate renderer input blob from Git, runs its
actual handleMouseDown and _dispatchMouseEvent, and compiles the unmodified
candidate applet using its retained small fixture. Geometry and unrelated
device plumbing are doubles, as are the original fixture's controls. A press
inside Preview queues WM_NCHITTEST (132) and WM_LBUTTONDOWN (513) for HWND 207. After three actual applet pumps, both events remain queued. No control receives
the press, and no saver launches. A direct posted WM_COMMAND/IDOK positive
control launches WIN98.SCR. This proves the undrained input boundary, not
whole-browser execution. List selection and the Close button use the same
unconsumed path; the title-bar system-button path is separate and is not claimed
broken by this reproduction.
The existing test masks this defect by injecting WM_COMMAND directly into the
WAT post queue, skipping renderer input and the control notification that must
produce that command. Its host fixture replaces init and stop; its painting,
window destruction, allocation and ShellExecute doubles do not qualify actual
lifecycle integration. All 12 original focused checks replayed successfully
without editing the retained source or original evidence. Passing those checks
does not resolve this failure.
Required next assignment: adapt applet input delivery to the current owned input/activation architecture, with a regression starting at renderer input and observing real native list/button behavior. Do not accept direct post injection as the sole user-interaction test. No fix was made in this review.
Other reviewed boundaries
- Historical
40193154applet data, control IDs, geometry, and behavior are faithfully retained; differences in the applet itself are class 33→35, host-window failure cleanup, and checking ShellExecute's result. Class 35 does not displace MDI 33 or Animate 34. IDs remain list 0x500, Preview 1, Close 2, statics 0xffff. - All 19 catalog entries are unique and registered. The static table terminates before UI strings; invalid index/field fixtures pass. The walker relies on its immutable NUL-terminated table, not a general corrupted-memory bound. SCRSAVE_TEXT is allocator-declared at 4096 bytes; all 263 preexisting region records remain unchanged. Current generator/declaration checks pass.
- The post-drain Close path reads its arguments before calling the wndproc,
clears applet globals, frees MSG once, and exits before another scratch read. Production
wnd_destroy_treerecursively destroys child state; host stop defers memory release until the current call unwinds. This is source-level reasoning, not real native teardown qualification. - Selection text is copied by STATIC and its temporary is freed. Initial
control captions (four) and list insertion temporaries (19) are allocated
via
wat_str_to_heapwithout matching applet frees: production controls duplicate/copy them andctrl_create_childfrees only CREATESTRUCT. This is inherited bounded per-open retention, released with process memory, not an observed unbounded idle leak. The fixture'swat_str_to_heap: p => pcannot measure it. Later allocation failures are also not covered by the initial MSG-allocation-failure test; the helper itself does not guard null allocation. - Cold startup uses current font preparation and suppresses guest Worker startup; the browser avoids system payload mounting. Missing exports/open failure route to stop. Cancellation during real async init and actual renderer/reference cleanup remain unqualified by the replaced init/stop test. In particular the early canceled-init branch precedes compile-status timer cleanup, so cancellation UI behavior needs real-path coverage.
- Preview uses the real host ShellExecute entry, and the new
.SCRfallback resolves registered savers with registry arguments. Registered launches are separate shell processes, notprocess_spawnchildren in_stopChildren. In single-app modelaunchExequeues until the applet exits; concurrent Preview-then-close behavior needs explicit desktop/mode qualification. Host bare-name failure remains lenient (33) if no launch resolves; the fixture's failure return 2 does not test that production behavior.
Evidence and limits
No full build, full VM, browser, remote runtime, heavy local build, asset registry qualification, saver rendering, or saver-survives-applet-close test was performed. MIDI's runtime was untouched. Shared HEAD/index/config, the candidate source tree, assets, prior ignored evidence, and game lanes were preserved. This is a failed source review with a small executable boundary reproduction, not browser or gameplay acceptance.