Half-Life 2 persistent launch owner and normal installer

Evidence: scratch/runs/20261117-halflife2-persistent-owner-normal-installer-progress/result.json.

The launch route now puts the unchanged actual-parent process supervisor under a dedicated waiting process. That process pins this single-task Codex worker's PID and kernel start token with a pidfd, verifies actual ancestry, clears the inherited transient-thread PDEATHSIG only after pinning, and checks the death race. It keeps the worker pidfd through its actual outer-child wait and both pipe EOFs. A separately armed owner deadline watchdog precedes outer startup. The outer retains its ordinary parent pidfd, scoped cancellation, native 600-second deadline and independent 620-second runtime watchdog.

The process tests actually return and reap a transient launcher, prove that the healthy dedicated owner remains, and inspect the actual supervisor's live parent pidfd in /proc/PID/fdinfo. Killing that dedicated owner must make the supervisor observe its parent death and cancel its child. A subreaper witness performs actual waits for the orphaned outer and owner watchdog. Separate tests exercise the independent owner deadline. Native route tests use the real DialogBox wizard, adapter signal cleanup, compiled sealed module and fresh main/CPUWorker/OS-child inheritance; native wait, both EOFs and descriptor release are required for normal completion and actual owner death.

The previous fully repaired wrapper, actor refusal, native-core and cleanup proofs are referenced by hash rather than regenerated. Source composition is explicit: base Git 70e85792eefac2489f655a415490f9e0a5c9297d plus the exact published Half-Life helper/component blobs. Every unchanged source must match the previous repaired qualification map, and the compiled module must match that proof. The owned component patch and Git identities recover the source; there is no source archive. Later unrelated emulator changes are not claimed as qualified by this run.

The prior UNKNOWN endpoint remains permanently closed. Normal installer input still requires a newly captured complete image and personal review for every finite intent. The actor/focus/modal/generation/context/target/nonce checks and 500/1750/3000 ms deadlines are unchanged. Successful normal installation and authenticated launch provenance remain prerequisites for any installed game invocation. No box or additional worker is used.

Initial authored route attempts are retained as failures. One test expected a CreateWindow line from a CreateDialog fixture. A second ran the dialog without the original's control port, so the tight fixture loop deferred SIGTERM and the outer's five-second grace expired, losing adapter-finally receipts. A third supplied a native exit receipt without its required invocation token and was refused at startup. The corrected authored command includes the original's control/event-loop mode and a matching UUID receipt token. These are fixture corrections, not evidence of a fault in the original installer.

Published route fix: 9945c2998002040fb5db955cf813258aa120dca9. Both successful authored native route source maps were compared to those published component blobs byte for byte. This exact match reuses the completed route tests after publication; it does not claim they were rerun under a later unrelated tree.

The one authorized unchanged original invocation then completed two personally reviewed normal Next intents: Welcome and Internet Connection Required. Both returned native-button-dispatch-return, consumed by actual native BUTTONs 0x10011 and 0x10017, with independent physical-UP acknowledgements and aggregate completions below 500 ms. The full scene was freshly captured and personally reviewed for each intent. Native logs then recorded creation of Steam Subscriber Agreement and its normal Agree/Do Not Agree controls.

The next complete-scene capture failed closed with control.unsupportedClass and control.textUnavailable. No agreement image was obtained and no agreement input was sent. This is the measured next boundary. The refusal reply does not identify its offending HWND or the reason native text was unavailable; the exact underlying control/text cause remains unproved. The tags alone do not establish an unsupported actual class because the supported-control predicate also requires available text. There is no original runtime or storage fault claim and no speculative admission change. Further diagnosis needs a separately authorized run with native refusal identities retained before the gate, followed by an authored faithful reproduction. Active shared capture/runtime files were not changed.

The endpoint permanently closed on this refusal. There were ten HTTP requests, two native DOWN edges, both followed by confirmed physical UP, and zero requests after the UNKNOWN latch. No retry, replay, rebind, second original, successful installation, installed game invocation or gameplay occurred. All original native, adapter, outer, owner and independent watchdog waits completed with both EOFs. The dedicated actual parent remained alive through join and its worker pidfd owner remained healthy. The native exit is a SIGTERM host-cleanup cancellation, not a native Quit. The installer allowance is spent.

Only the two named original screenshots are retained. Four duplicated normal extraction outputs were replaced after closure by exact original-media stream recovery recipes, with complete decoded-byte and CRC/SHA roundtrip verification. Unique runtime output remains. The original EXE hash was verified again after closure. No box was created; ID and expiry are null and stop is not applicable.

An additional authored route fixture explicitly sets inherited PR_SET_PDEATHSIG=SIGTERM before executing the same owner. All three actual process cases record inherited 15 and safely cleared 0, then prove healthy survival, parent-death cancellation, or independent deadline cleanup. During that diagnostic, a live /proc/PID/fdinfo descriptor closed between enumeration and reading; the test now ignores such vanished unrelated descriptors while still requiring a matching live parent pidfd. This changes test robustness only. The original's sealed source graph and executed component bytes remain those recorded under 9945c2998, before this follow-up test edit.