Add build verification and isolated Minecraft puppet CI

[TeamDman]

The 1.19.2 branch lacked a reproducible build check, and a future Discord help worker needs to run the actual modded client without exposing the Windows host to each session. This PR adds Linux and Windows verification with the canonical sfm-propagate-changes CLI, uploads the built mod, and runs the existing title and GameTest puppets inside an isolated Linux container with Xvfb and Mesa software graphics. Feature-branch pushes and PR merge checkouts both run; the busy 1.19.2 checkout is untouched.

The container is prepared with network access, then starts fresh clients offline. The smoke worker uses UID 10001, no capabilities or privilege escalation, seccomp, a read-only root, no host mounts, GPU or engine socket, and CPU, memory, PID and wall-time limits. It requires the raw game completion marker, the item-transfer GameTest pass, and exactly three title plus eight orbit PNGs. Failed runs retain bounded diagnostics, and stale evidence cannot satisfy a retry. The same smoke script now supports local Podman; its empty bounded tmpfs mounts avoid Podman's default copy-up of multi-gigabyte image build caches.

The frozen Vox Java artifact exposed a late-credit race. Its repair is published in Facet PR #2 and pinned here by exact commit and JAR hash; all other dependency declarations remain unchanged. Three asynchronous test assertions now wait for the event they actually observe. A bounded Linux JUnit watchdog writes thread stacks to the existing always-uploaded diagnostics; its six snapshots identified a production Explorer deadlock. The loader now claims one terminal result under its handle lock and completes callbacks after releasing it, so session cancellation cannot invert the locks. Failed release-review surface futures are evicted on lookup as well as by completion callback, so a retry cannot reuse an already-failed result. The canonical test command and 15-minute CI limit are unchanged.

Validation:

• Hosted Linux and Docker passed at source d7e22e73e: 2,075 Java tests passed, one Windows-only skip, six existing opt-in prerequisite aborts, verified mod JAR, and both offline game puppets with 11 decoded PNGs. The cold prepared-image build took 19m12s and the two client launches together took 3m42s.
• Hosted Windows passed at the same source: 2,076 Java tests and package verification, with zero failures and the same six opt-in aborts. The new Vox JAR bytes match Linux; full mod archives differ in text line endings, manifest timestamp and compiler debug offsets.
• Local rootless Podman 6.0.2 on WSL passed at image source 01b239124: default-container DNS and HTTPS worked, while --network none blocked DNS and direct-IP egress. Xvfb/Mesa reported llvmpipe OpenGL 4.5. The offline worker logged move_1_stack_direct passed!, produced 3 title and 8 orbit PNGs, exited 0 without OOM, and left no container or anonymous volume. The settled title and final orbit images were visually inspected.
• At 73ead2d0a, the PR Linux and PR Windows workflows fully passed. Later Linux diagnostics exposed the Explorer lock cycle and a Windows PR run exposed the failed-surface cache race. The six new lock regressions all fail against the original loader and pass with the fix. In local rootless Podman, the focused loader suite passed 19/19, review-surface suite 6/6, and the full canonical suite passed 2,081 tests with zero failures, one skip and six existing optional-prerequisite aborts (2,088 found). The Podman wrapper lifecycle tests, 11 evidence-verifier regressions, JUnit watchdog lifecycle test, actionlint, and git diff --check also pass. At final head 5dbbaeae9, both Linux PR and Linux push workflows fully passed: 2,081 Java tests, verified mod JAR, graphics isolation, and both offline game puppets with 11 PNGs. Both Windows PR and Windows push workflows fully passed: 2,082 Java tests and package verification. All hosted suites had zero test failures; the six optional-prerequisite aborts remain expected.

Discord and Kubernetes wiring remain future integration work. A help bot should use a separate controller and token, bounded inputs and outputs, prebuilt trusted images, and a readiness check before selecting an orbit frame; the first frame can catch incomplete geometry. The anonymous workspace has no disk quota, and arbitrary executable uploads would need a stronger sandbox. No game process ran directly on the Windows host and neither draft PR was merged.