fix: clean lock-surface teardown + roundtrip on unlock (Hyprland multi-output)

[josephdunn]

Problem

On a multi-output Hyprland setup, unlocking rustlock consistently triggers Hyprland's "lockscreen died" failsafe screen instead of unlocking, even though rustlock exits cleanly with status 0. Single-output setups are unaffected (which is presumably why this hasn't shown up before).

Reproduced on Hyprland with three outputs (two rotated 2560x1440 + one 3840x2160).

Root cause

Two issues in the unlock path, both tolerated by Hyprland on single-output but not on multi-output:

1. Wrong teardown order. ext-session-lock-v1 recommends destroying each ext_session_lock_surface_v1 before issuing unlock_and_destroy on ext_session_lock_v1. rustlock did the opposite: it called session_lock.unlock() first inside handle_auth_result, then cleared lock_surfaces from the auth-feedback callback afterwards.

2. No event drain before disconnect. After unlock_and_destroy, the compositor sends back wl_keyboard.leave, wl_pointer.leave, and wl_display.delete_id events. Without dispatching those, rustlock disconnects mid-conversation; Hyprland treats that as an unclean exit.

A WAYLAND_DEBUG=client trace of the baseline showed unlock_and_destroy sent first, then lock_surface.destroy() / wl_surface.destroy() pairs, then disconnect — no wl_display.error from the compositor, just the failsafe afterwards. A trace of the fix shows the spec-recommended order followed by a brief drain of leave/delete_id events before exit, and Hyprland unlocks normally.

Fix

• Move self.lock_surfaces.clear() into handle_auth_result so it runs before session_lock.unlock(). Remove the now-redundant clear from the auth-feedback callback.
• Add state.conn.roundtrip() after the main event loop exits to drain the post-unlock events before disconnect.

Total change: +11/−3 in src/main.rs.

Test plan

• Multi-output Hyprland (3 outputs): unlock works, no failsafe.
• Single-output: tested on single-screen laptop — both changes should be no-ops there.
• WAYLAND_DEBUG=client trace confirms teardown order matches the spec recommendation and leave/delete_id events are drained before disconnect.

Note

AI (Claude) helped me diagnose and write the fix.