Half-Life 2 destination focus lifecycle
Task evidence: scratch/runs/20261117-halflife2-destination-focus-normal-install/.
The prior run's destination Next was consumed and physically released. Its
next capture correctly refused focus 0, thread 0, main 0x10025, modal 0.
That run made no later HTTP request, so its focus persistence remains unknown.
Public metadata shows Wise's age, connection and destination pages are modeless
children within the persistent main 0x10025. The authored fixture now follows
that topology with real CreateDialogParamA, EndDialog, timer/message dispatch,
and main destruction/recreation. Ending the child clears its dead focus while
the main remains live. An ordinary guest timer creates the next page and seeds
its focus. The native proof actually observes main-live/modal-zero/focus-zero
before the next page. There is no host focus assignment or controller pump.
The controller observes only this narrowly diagnosed candidate inside the original three-second capture clock. The actor, process and screen must match the permanent binding. Readiness allocates no image or input authority. Once ready, exactly one strict capture runs with the remaining time. Capture and input still require live current HWNDs/generations and exact reviewed metadata. A real modal owner mismatch is terminal, as demonstrated by the rejected modal fixture. A five-second authored modeless gap produced 71 null-focus polls over 2.955 seconds and terminal UNKNOWN at the unchanged three-second limit, with no strict capture during pending and zero subsequent HTTP.
Both working and published full authored workflows passed two changing-page Next/UP intents and Cancel/UP/native Quit with one binding, three actual main replacements and current control generations. Focused tests cover stale and recycled HWNDs, actor/process/thread/module/runtime/device/queue ownership, strict capture refusal and immutable capture/review/input clocks. Native consumption remains 500 ms; independent physical release 1750 ms; complete capture/aggregate intent 3000 ms; menu/game review ages 30/15 seconds.
The exact retained module is
928f8bd42c76b6435ec59d8b289b5a40a28da8b597a6e98e378b467311f11abb.
Unowned diagnostic sources remain pinned to the previously proved
fe074afb6341bb9b1294924638424184a441ae08 graph. Only the owned scene/controller
bodies and exact reviewed-owner runner block are adopted. Canonical main builds
are separate from the unchanged module used by this sealed HL2 run.
Prelaunch refusals preserved evidence of unrelated USER-adapter graph drift and run-specific abbreviated/full Git pin disagreement before media invocation. The adapter now resolves aliases to an exact 40-character commit before sealing and receipt checks. The final original outcome is recorded below after the published complete workflow and native Quit proof.
Published code commits are b7d01871d, 509a290ec, and 8a04a857d.
The final fresh full workflow at exact 8a04a857d passed native Next/UP and
Cancel/UP/Quit. Both native post-intent observations saw null focus; the first
capture additionally recorded five pending observations. The second capture
started after focus returned, so the stricter test requiring pending observations
on both captures failed and is preserved. A separate current native sequence
check and the ordinary full workflow verifier passed. The previously completed
diagnostic proof was reused only after equality of all 5,004 executed source
hashes, source-pool/index hashes and the module; its actual wait/EOF and native
Quit proof were verified. Repeated diagnostic attempts that hit 500 ms remain
failed evidence. Canonical flock builds passed; temporary foreign census
refusals and subsequent owner corrections are retained.
Exactly one original installer was then invoked, with unchanged media SHA,
module and source inventory verified before execution. It received zero inputs
and produced no complete PNG. Passive startup observations still showed hidden
main 0x10001, focus 0x10003 under top-level 0x10002, modal 0, and failures
main.visible, focus.mainTopLevel, owner.ordinaryUSER. The run-specific
ctl.py wrapper expired its eight-second communicate timeout while the inner
startup operation permitted fifteen seconds. The startup helper's actual wait
and both EOFs were not retrieved. That lifecycle gap is explicit evidence,
not an exit-0 or completion claim. The worker latched permanent UNKNOWN; no
later HTTP, capture, rebind or input occurred.
The original native runtime, persistent owner, outer supervisor and watchdogs were actually waited/drained, and known owned groups are empty. Native cleanup recorded 2,571,825,152 free bytes, above the 1.5 GB floor. This was host cleanup. Original agreement acceptance, destination progress, installation and native Quit were not demonstrated in this task. Earlier destination focus persistence remains unknown. The new blocker is the run-specific startup controller's mismatched timeout and missing helper wait/EOF, with hidden-main readiness as the last actual native observation. No native focus implementation repair is claimed. Fix controller lifecycle with an authored delayed-start regression before any separately authorized fresh original attempt. No box was created; ID/expiry are null and stop receipt is not applicable.
The separately authorized controller repair is published as d5259f9e5.
halflife2-restored-native.py --controller RUN OP ... validates finite call
schemas and owns the reviewed helper through the existing parent-process
supervisor. Startup keeps its 15-second inner bound and has a 20-second outer
bound; capture/intent keep three seconds internally. Independent watchdogs arm
before the helper. Cancellation first signals the helper alone, allowing its
existing independent physical release to drain, then performs bounded scoped
cleanup. TERM/INT unwind through the reviewed controller's exception cleanup.
Every path records actual helper/watchdog waits and both pipe EOFs. Seven real
process regressions passed: nine-second startup, permanent unready at fifteen
seconds, helper error, outer timeout, TERM, release drain, and actual parent
process death. The earlier missing helper wait remains unreconstructed.
Evidence is in
scratch/runs/20261117-halflife2-controller-deadline-join-normal-install/.
Canonical flock build and focused ownership/capture/clock guards passed.
The retained 928f8bd4... module and 5,004-source graph were reused; only the
reviewed controller's signal-handler source changed. Existing main replacement,
focus and diagnostic proofs were reused by hash. Fresh native startup and a
personally viewed Cancel/UP/native Quit passed in 2,132 ms with actual normal
native wait and both EOFs; fresh Next sequences were not repeated.
One fresh ordinary original invocation passed readiness, then whole-image
reviews preceded Welcome Next, Internet Required Next, and I Agree inputs.
Their aggregate completion times were 314, 589 and 749 ms, with native DOWN
and independent physical UP within the unchanged 500/1,750 ms bounds. After
agreement acceptance, the observed snapshot had quit=true, main/modal/focus
all zero, with only hidden old install controls. The next passive readiness
observation refused runtime.running and ordinary USER ownership. No age,
destination or installation image was obtained. UNKNOWN latched and there
were zero later HTTP requests or inputs. The native runtime was host-cleaned
(exit 0, actual wait/both EOF), not proved to Quit through original UI. The
cause of the terminal native scene is unknown; this task makes no native
focus/quit implementation attribution and performs no replay. Installation
and original native Quit remain open. No box was created.