Verify desktop monitoring across GNOME and Plasma logins

[Staphylococcus]

Parent: #89. MVP slice 2 of 3. Target: v1.8.0. Validates the source runtime in #216.

Outcome

One installation completes GNOME Wayland -> logout -> Plasma Wayland -> logout -> GNOME Wayland with the same configuration and no manual LG Buddy restart. Available interfaces determine adapter contributions; shared policy does not select a desktop/session profile.

Scope

• Verify existing graphical-session service stop/start wiring, including a user manager kept alive by SSH/console logins or lingering. Fix demonstrated activation/environment gaps.
• Keep interface connections and any Wayland/logind object selection inside adapters. Retain logind filtering of remote and non-graphical sessions and existing eligibility checks; diagnose ambiguous graphical matches.
• Verify shutdown cancels subscriptions and pending checks. New monitoring cannot consume old observations/replies. Reconcile TV ownership through existing restore policy rather than discarding markers.
• Exercise late interface startup, owner replacement and reconnection for activity and inhibition capabilities independently, using Compose independent activity and inhibition capabilities for GNOME and Plasma #216's contract. Each capability owns its push/pull handling and recovery inside the adapter; sharing a transport does not merge the paths. Inhibition meets activity eligibility only at the Boolean blanking gate. A quiet healthy push capability is not a failure.
• Verify useful connection diagnostics while preserving independent gamepad monitoring and the separate system sleep/wake service. Record the available source set for each leg; an absent optional inhibition source is ordinary reduced coverage. Expected source availability belongs to the test setup, separately from the runtime decision.

Acceptance

• The GNOME -> Plasma -> GNOME round trip requires no configuration edits, desktop selection or manual service restart.
• Concurrent SSH/console logins do not change desktop integration; the round trip works with a retained user manager.
• Old connections, observations and check replies cannot affect the next login.
• Late/restarted interfaces recover automatically. Missing inhibition support does not disable a source's activity capability; startup, absence and source failure contribute no inhibitor. Quiet healthy push sources and pending refreshes retain their last observed state; failed reads drop the source contribution without manufacturing a release. Cancelled checks return no verdict and obsolete replies cannot authorize a later attempt.
• Idle/restore, gamepad, lock/unlock and suspend/resume work on each leg, including recovery with an existing ownership marker.
• Disabled built-in monitoring preserves passive notifications and unrelated GUI/CLI operation.

Boundary

The MVP supports one local graphical session per account alongside non-graphical logins. Simultaneous graphical desktops and fast user switching are separate scope; existing desktop support must not be deliberately removed.

This is lifecycle verification and fixes. #216 owns the inhibition coordinator, reconciliation semantics and adapter protocol behavior. The implementation, fixture and lifecycle evidence are delivered together in #231; final MVP validation combines the three slices.

Verification — 2026-09-12

All six acceptance items are verified. The complete implementation, concurrent TV fixture and evidence are consolidated in #231 on feat/217-desktop-session-validation.

• The combined candidate (5e3522986672e177233c18102605830376c1c863) completed GNOME → Plasma → GNOME with unchanged configuration, user manager PID 865 retained, and SSH session 177 present through both switches. Monitor PIDs changed automatically from 2319 → 11680 → 22011; lifecycle PID 2315 survived throughout.
• Every leg passed idle/restore, overlapping inhibitors and final-release delay, gamepad restore/reset, lock/unlock before and after actual guest s2idle, and joint TV power-off/WoL/HDMI_3 recovery. Both services retained their identities across suspend with zero restarts. Existing ownership markers survived logout and were restored on new input.
• Live PowerDevil restart recovery passed without restarting LG Buddy. Existing adapter/private-bus/runtime tests verify late appearance, owner replacement, capability independence, retained quiet/pending state, failed-read semantics, cancellation and rejection of obsolete replies.
• Separate disabled-monitoring checks passed on both GNOME and Plasma: real desktop Notify forwarding, installed GUI launch, CLI TV reads and no idle blanking. The original configuration was restored after these checks.
• The earlier VM stalls were reproduced with both LG Buddy services stopped: a virtio GPU command-queue wait blocked desktop switching, and an Incus-agent 9p mount blocked suspend. The test VM's QXL display and masked unused agent allowed the full matrix to complete. No production workaround was needed.

The validation report records exact hashes, service identities, test controls and limits, including the corrected unlock harness and the earlier Hearth hardware observations:
Complete #217 validation record.

This completed lifecycle matrix used PowerDevil for Plasma inhibition and does not claim native KWin inhibition coverage. The current MVP plan adds that independent source through #233, including prebuilt, local-build and absent-source validation. #233 owns the additional bridge lifecycle/integration checks; this completed #217 evidence remains valid for the source combinations actually tested. Activity and inhibition remain independent and meet only at can_blank().

All automated checks passed for the combined code, including 1,023 unit tests, 151 Cucumber scenarios, integration tests, Clippy, the real-CLI fixture smoke and all jobs in CI run 34711926512. One physical gamepad test remains ignored; virtual gamepad input was exercised live.

No remaining #217 implementation or validation blocker. Acceptance is complete and #231 merged into dev on 2026-09-12 as f1160d57d60a72d6fa63cbade6ea7858d0bbd50a. The PowerDevil-only MVP-boundary statements in the earlier evidence/comment describe the earlier plan; the current source/fallback plan is #233. This does not promote a prerelease or complete the remaining #89 MVP/reporter-validation work.

[Staphylococcus]

Hearth hardware observations — 2026-09-12

Tested candidate c50f3eab67ef9214e3173284189dcb5cd0979ad7 from #231 with both the screen and lifecycle services running against the physical LG TV on HDMI_3. Hearth runs NixOS 26.05, kernel 7.2.3 and GNOME 50.4. This was a local debug build temporarily substituted through runtime service overrides.

• Joint suspend/resume worked. Lifecycle completed TV power-off at 16:27:45 UTC before deep/S3 suspend. Resume reached lifecycle at 16:28:19; network recovery completed and WoL began at 16:28:25. HDMI_3 restoration succeeded at 16:28:39 after two failed attempts. The user confirmed the picture returned and unlocked normally.
• The two services respected their boundary. The monitor deferred the sleep lock and early post-resume activity to lifecycle. Monitor PID 1716243 and lifecycle PID 1716211 survived with zero restarts. Lifecycle reacquired its delay inhibitor, and ownership markers cleared.
• Lock/unlock worked before and after suspend, with physical input needed to show the login screen. Both lock events blanked the TV; timed CLI unblank commands succeeded. In the first check, the user saw “no signal”: GNOME's HDMI output was disabled (PowerSaveMode=3) although the TV had unblanked. Mouse movement returned the login screen. The final check also needed input, and normal unlock worked. No compositor stall was observed in this cycle.

Test controls: an RTC wake was armed; a helper requested HDMI-output wake after resume. The user's response did not separately establish whether manual input helped wake the PC. A delayed emergency TV-wake helper failed before executing a TV command because of its transient-unit PATH; lifecycle had already restored the TV before it launched. The successful recovery therefore came through the normal lifecycle path.

The original installed runtime is restored in both services, both are active, the session is unlocked, a TV read succeeds, and all test timers/overrides are removed. The configuration hash is unchanged. Inhibitor honoring remained disabled in the existing hardware configuration, so this pass makes no inhibition claim.

This adds real-hardware evidence for GNOME's joint lifecycle path; it does not close the full GNOME/Plasma matrix. PowerDevil remains the agreed Plasma inhibition boundary; native-only KWin coverage is outside the MVP gate.

The fixture is still unchanged. The next extension should allow concurrent clients and controlled TV power/wake/readiness transitions, and keep TV blanking separate from whether the desktop supplies an HDMI signal. Exact journals, helper scripts, candidate identity and user observations are retained under ~/.local/state/lg-buddy-validation/217-2026-09-12/hearth/; the local validation document has the detailed timeline.