fix(session): persist the idle inhibitor across shell restarts

[Mark2Mac]

Description

The idle inhibitor ("Keep Awake") does not survive a shell restart. SessionService holds it as a plain in-memory property and nothing ever writes it out:

property bool idleInhibited: false

Every restart silently clears it. The user's explicit choice is dropped with no message — the pill disappears from the bar and the screen starts sleeping again mid-task.

Restarting is not an edge case: dms restart is a first-class CLI command, and assets/systemd/dms.service in this repo sets Restart=on-failure with TimeoutStopSec=10, so a shell that is slow to stop is killed and brought back automatically.

Every other Control Center toggle already survives a restart. Measured on upstream master, over the state read by Modules/ControlCenter/Components/DragDropGrid.qml and its neighbours:

toggle	declared in SessionSpec.js
doNotDisturb	yes
isLightMode	yes
nightModeEnabled	yes
wallpaperCyclingEnabled	yes
idleInhibited	no

So this follows the doNotDisturb idiom exactly rather than inventing a mechanism: one spec entry, one property, one setter that saves.

• SessionSpec.js — idleInhibited: { def: false }
• SessionData.qml — the property plus setIdleInhibited(enabled)
• SessionService.qml — enableIdleInhibit() / disableIdleInhibit() write through; the state is restored at startup

Restore goes through two entry points because the singletons are lazily created and their order is not fixed: onLoaded covers "service instantiated before the session file is read", Component.onCompleted covers the opposite. enableIdleInhibit() is idempotent, so whichever runs second is a no-op.

The default stays false and SessionStore.toJson omits defaults, so an untouched session.json gains no new key.

Type of change

• Bug fix (non-breaking change that fixes an issue)
• New feature (non-breaking change that adds functionality)
• Breaking change (fix or feature that changes existing behavior)
• Refactor / internal cleanup
• Documentation
• Other

Related issues

None open for this. #1483 (configurable timer) is adjacent: the field added here has the same shape doNotDisturb uses alongside doNotDisturbUntil, so a later idleInhibitedUntil slots in beside it without another migration.

Screenshots / video

No visual change. The bar pill already reflects SessionService.idleInhibited; it simply now shows the state the user left it in. Behaviour is shown as measurements below instead.

How this was verified

In a nested niri with isolated XDG_CONFIG_HOME/XDG_STATE_HOME, so the bench could never touch the live session:

qs ipc --pid <nested> call inhibit enable
→ session.json:  "idleInhibited": true
kill <nested>                 # what systemd does at TimeoutStopSec
relaunch, then: qs ipc --pid <new> call inhibit status
→ Idle inhibit is enabled

Counter-test on the same session.json, with the patch reverted to plain upstream:

	after the shell restarts
with this patch	Idle inhibit is enabled
upstream as-is	Idle inhibit is disabled

Also exercised directly against SessionStore.js + SessionSpec.js: parse({idleInhibited: true}) → true, toJson → {"idleInhibited": true}, and with the value at its default the key is omitted from the output.

Checklist

• My code follows the conventions in CONTRIBUTING.md
• I have tested my changes locally
• New user-facing strings are wrapped in I18n.tr() — n/a, this PR adds no user-facing strings
• Go changes — n/a, QML/JS only
• QML changes: ran make lint-qml with no new warnings — could not run locally: the script requires the Quickshell tooling VFS (quickshell/.qmlls.ini + <buildDir>/qs/qmldir), and on quickshell 0.3.0 here the VFS directory only ever contains tooling.lock, so lint-qml refuses to start. The two touched QML files parse cleanly under qmlformat, and the pre-commit hooks that apply were run by hand (no trailing whitespace, single trailing newline, no console.*). Happy to re-run if you can point me at what generates the VFS.
• Documentation: companion PR opened in dlx-docs: docs(ipc): idle inhibit persists across restarts, plus the two undocumented verbs DankLinux-Docs#138

[bbedward]

/claude review

[claude]

Claude review

No issues found.

Verified the restore path both ways: SessionData declares signal loaded (emitted after Store.parse in both loadSettings and parseSettings), so onLoaded covers "service completes first" and Component.onCompleted covers "session already parsed" — enableIdleInhibit() is idempotent, so the second is a no-op. SessionService has no other Component.onCompleted to clash with, the new idleInhibited key does not collide with an existing spec/property name, and consumers (IdleService._applyMonitorEnableds, DankBarWindow IdleInhibitor, control center, OSD) read the value through bindings or recompute from current state, so an early restore is picked up correctly rather than depending on the inhibitorChanged emission. No new mutual-singleton hazard beyond the existing SessionData -> SessionService Connections.

Checked: restore ordering/idempotency, consumers of SessionService.idleInhibited, spec/JSON round-trip and default omission, no Go-side session schema mirror, no dank-qml-common contract property involved, en.json/template.json untouched, no new user-facing strings. Model: claude-opus-5.