Half-Life 2 full wrapper and normal installation progress

Evidence: scratch/runs/20261117-halflife2-fixed-wrapper-normal-install-progress/result.json.

The prior installer endpoint remains permanently closed. Its correctly bound Next control was never sent input because the wrapper shadowed the helper source lookup key with the finite intent name; published 936ccca4f fixes that collision.

The authored full-wrapper regression now retains the complete original runtime metadata and an opaque source field, then executes capture, review, bind and an arbitrary finite intent name against a real native DialogBox. It requires the native BUTTON return and both title and body changes. Reusing the same capture must produce a non-consumed native cancellation; every subsequent capture, bind and intent must fail before HTTP. Existing wrong-target and actor/focus/modal/generation/context checks and 500/1750/3000 ms limits remain.

The adapter uses this task's separate single-invocation reservation. Completed published VB native-core component tests are reused only for the same compiled module and matching material source inventory. This does not reuse the VB original's later cleanup as an outer lifecycle proof. The adapter always runs fresh sealed main/CPUWorker/OSchild inheritance, focused changed components, the full wrapper and its own actual outer lifecycle before original startup. Changed module, material bytes and inventory are explicitly refused.

The single authorized original invocation ran published 70e85792e, module 9099bb3f694d5ec6dac2886b70c1920fa33f39ad9258b1e662d1ebe05dc2da0a. The unchanged original EXE hash was verified before the 600-second native clock and independently prearmed 620-second watchdog. Exact published source inheritance and matching qualification were recorded before startup.

The complete genuine Welcome to Steam image was personally inspected. Before binding, the pinned launching shell process disappeared. The outer supervisor observed its process pidfd become readable at native elapsed 80.296 seconds and correctly sent scoped SIGTERM. Both watchdogs were cancelled after actual outer-child wait and outer pipe EOF; neither deadline fired. The reason the shell disappeared is unknown. This is a host lifecycle boundary, not a proven emulator, native input, storage or installer failure.

The adapter had imported the shared runtime without running its __main__ signal-handler installation. SIGTERM therefore killed it before its native wait, pipe drains and resource-finally receipts could be published. Its buffered 64 KiB reads also left both small native log files empty. The original native process published a matching SIGTERM cancellation receipt. There is no original native actual wait or native EOF proof; process absence and outer EOF do not substitute for those missing facts. No input request, DOWN, held input, native Quit, successful installation or installed-game invocation occurred. The endpoint permanently closed on the local identity failure and was never queried again. The invocation allowance is spent.

The necessary adapter repair explicitly installs the imported helper's shutdown handlers and reads available pipe chunks incrementally. Its native process lifecycle is shared by original startup and authored tests. Actual authored CLI SIGTERM and SIGINT tests exercise run() through real visible window creation, verify incremental logs before signalling, then require an actual native wait, both EOFs, compactor join and descriptor release. These tests prove cleanup behavior; they do not explain or waive the lost external owner. The original endpoint is not revived for testing.

Installed physical footprint is unknown; the installer declaration of 3995 MB is not storage admission. Continuing normal installation requires another explicitly authorized original invocation from a persistent launching owner, using the published cleanup repair and a new personal complete-scene review.