Draft: experimental Bazzite support with locally signed drivers

Runtime regression: do not install app release 4.bazzite yet. On 2026-10-31 the local build and all six selected component tests passed, but the tester reported black video and frozen/very slow input when connecting with Artemis after installation. Rolling back to app release 3.bazzite restored streaming. The retained-inhibitor KDE workaround remains the qualified sleep solution. The saved log shows encoder probes and saved-topology restore failures during startup, but no client connection attempt, so the cause is not yet established. Source/Force Close changes remain unqualified; investigation is in progress. No new build is requested.

This draft adds an opt-in Fedora Atomic/Bazzite RPM profile and consolidates fixes used to get Vibepollo streaming successfully on a Bazzite NVIDIA Open machine. The ordinary native Linux package lifecycle remains the default.

Changes

• Package the app without mutable-host service/DKMS installer scriptlets; preserve the trusted KWin bridge mode in RPM metadata and require a separate driver package.
• Build and sign kernel-specific DRM/DS5 modules locally, keep the signing key root-only, verify signed payload hashes after packaging, and regenerate module indexes through RPM scriptlets.
• Replace the boot-time DRM rebuild/status path with checks of the packaged driver.
• Preserve identical extended attributes without rewriting a read-only SELinux label; continue to fail on unsuccessful required metadata changes.
• Recognize Fedora Atomic's root-owned /home -> /var/home alias while retaining private-path checks; accept existing mixed-case desktop account names in setup and the controller.
• Remove the pinned NVIDIA runtime's unused build-machine RPATH from its installed copy and supply the missing versioned RPM capability for its bundled hashed CUDA runtime.
• Include local build/sign/package helpers and a qualification/setup document. No hosted Bazzite workflow is added.

Hardware qualification (tester-reported, 2026-10-29)

Bazzite NVIDIA Open 44 / deployment 44.20261026, KDE Wayland, kernel 7.2.11-ogc4.1.fc44.x86_64, NVIDIA 615.78.08, RTX 5080, Secure Boot enabled and SELinux enforcing. Client: Artemis on an M1 MacBook Air.

Passed: pairing; desktop video/audio/keyboard/mouse; disconnect/reconnect; physical monitor restoration after disconnect; streaming after a fresh reboot and desktop login; logout/login; suspend/resume. Both signed modules loaded, and the tray and mixed-case account fixes were exercised on the machine.

These results came from efd8d7341dba879174529f7319a89d8d0332cb74 plus the local fixes consolidated here. This new consolidated checkout has not yet had a clean end-to-end build/install rerun. Startup logs verified H.264 NVENC initialization; HEVC was the requested client setting, but negotiated stream codec was not independently recorded.

Source validation

• 6 focused runtime/packaging regression tests passed.
• 3 CMake package-profile tests passed.
• 27 existing desktop-tray tests passed.
• Shell syntax and git whitespace checks passed.
• Broader installer/controller suites were attempted but could not fully pass in the editing environment: Unix sockets/UID transitions are restricted, and bsdtar/libcap development headers are unavailable. No full-suite pass is claimed.
• Full app compilation and driver compilation/signing/installation were performed on the tester's PC during qualification.

Why this remains a draft

The driver helper intentionally targets the qualified 7.2.11 kernel. Automatic kernel-update rebuild/sign/deployment handling, clean reinstall and uninstall/rollback recovery, other machines/images, HDR/AV1, controller input, performance measurements, remote-network access and pre-login capture still need work. The Linux tray Force Close implementation is now included; installed runtime qualification remains pending.

The code was AI-assisted and tested interactively by the contributor. Maintainer feedback on the packaging approach and whether to split the general Linux fixes into separate PRs would be welcome.

CI intentionally skipped: the tester requested local builds only. The head commit carries [skip ci]; please do not dispatch hosted builds for this draft without coordination. This is not a request to merge without normal review/qualification.

Build steps, exact tested environment and remaining work: Bazzite README.

Kernel-update preparation prototype

The follow-up commit adds --target-image for a digest-pinned Bazzite NVIDIA Open image: inspect its kernel/headers, compile in an offline rootless target container, reuse the enrolled local key via the trusted host signing tool, and generate an RPM with a distinct kernel/source release plus a signed-image rebase command. It does not execute that command, alter boot selection or reboot. Missing headers, compiler, key or a failed build stop preparation. Eight new policy tests pass. No actual future-kernel build/staging/boot has been qualified yet, and this is not automatic updater integration.

Additional tester report: Wake-on-WLAN from Artemis successfully woke the sleeping host after enabling magic-packet wake for the Wi-Fi connection. This is host network configuration, not a driver/package code change.

Same-kernel update rehearsal passed

Tester report on 2026-10-30: prepared modules in the digest-pinned Bazzite image fecc39b4b278dabd27092d0b0cad01a6ab4ce7eb07b669561150fa77dea684c7, reused the original enrolled certificate (SHA-256 8e6958635553132de3238e36492359bdd643381da7ef9508088c295a0d19ee08), pinned the working deployment, and successfully staged the signed image plus local driver replacement in one rpm-ostree rebase transaction. After reboot, vibepollo-kmod-2.0.2-3.bazzite.k7.2.11_ogc4.1.fc44.x86_64.g32bcc958f899 was installed, both DRM/DS5 modules were loaded, and Artemis streaming/video/audio/input worked without manual startup.

This validates the preparation/package-replacement/reboot path on the same 7.2.11 kernel, not compatibility with a new kernel. Different-kernel qualification and automatic updater integration remain pending. The rehearsal rebased to an immutable image digest; it does not retain a moving stable origin, so subsequent updates require an explicit prepared target until channel/update integration is implemented.

Automatic suspend and Linux tray Force Close

Tester report on 2026-10-31: automatic suspend failed with only Vibepollo / Retained remote display left in PowerDevil. Debug logging showed that it also added InterruptSession; KDE's addInhibitionTypeHelper explicitly implies that policy for ChangeScreenSettings. Rejecting only this retained-display inhibitor through KDE restored automatic sleep, Artemis Wake PC and reconnect. Active-stream inhibition was kept enabled.

Commit f5a0e05dfb5f2d5a847b93cec4091f2cea3d86b2 removes the retained power hold, makes legacy retain-helper modes non-inhibiting, and keeps capture-owned inhibition and reconnect wake. It also adds Force close client to the Linux companion, using the existing pinned/token-scoped apps-close endpoint with an explicit force flag to drain RTSP/WebRTC sessions before native application cleanup. The host remains running and clients stay paired. Existing graceful API requests remain graceful.

Validation for this follow-up: 29 tray tests, 3 package-profile tests, 6 existing runtime-fix tests, machine-service source contracts, shell/Python syntax and whitespace checks passed. A compiled harness executed the actual close handler and checked authorization/content-type denial, malformed and mistyped request rejection, legacy graceful bodies, and force transport-drain ordering. The full app and modified GIO power-helper tests still require the local Fedora build; update-runtime.sh reuses the original qualification checkout/cache, runs those focused checks and packages app release 4.bazzite without installing it. No driver rebuild or key enrollment is needed.

Pending hardware qualification: install this source revision, remove the diagnostic KDE exception, retest automatic sleep/wake/reconnect and the tray action, then test idle display-off/reconnect. The former retained hold was intended to avoid an NVIDIA GSP hang with a virtual-only idle topology; that longer DPMS regression is not established by the two-minute sleep workaround test. This remains a draft.