fix(backend/tty): add delayed rescan for connectors missing EDID on hotplug

[coleleavitt]

Problem

USB-C docks with DP MST / alt-mode (e.g. Lenovo ThinkPad USB-C Dock Gen 2) can report connectors as Connected to the kernel DRM subsystem before EDID data has been read. When this happens:

1. connector.modes() returns an empty list
2. pick_mode() returns None
3. connector_connected() logs "no mode" and skips activation

Smithay's ConnectorScanner treats (Connected, Connected) as a no-op — it does not re-emit events for already-connected connectors. This means the output gets stuck in a permanent "connected but never activated" dead state.

Whether activation succeeds depends entirely on timing: if a second UdevEvent::Changed fires after EDID completes, the connector recovers. This makes the bug intermittent — monitors sometimes come up and sometimes don't on the same hardware.

Root Cause

The EDID race is in the kernel/dock firmware timing, but niri has no retry path for connectors that were connected but could not be activated due to missing modes.

Fix

Add a bounded retry mechanism in Tty:

• After device_changed() processes connectors, schedule_rescan_if_needed() checks for connected connectors that have no matching surface (not yet activated) and are not non-desktop connectors
• If found, schedules a calloop::Timer (2 s delay) that re-invokes device_changed(), giving the kernel time to complete EDID reads
• Retries are capped at MAX_RESCAN_RETRIES (3) to prevent infinite rescheduling
• The timer self-clears when all connectors are successfully activated
• Existing timers are cancelled before scheduling new ones (cancel-and-reschedule)
• Timers are cleaned up on device removal

New fields on OutputDevice

• rescan_timer_token: Option<RegistrationToken> — handle to cancel pending timers
• rescan_retry_count: u8 — bounded counter, reset on successful activation

Testing

• Hardware: ThinkPad P16 Gen 3, NVIDIA RTX PRO 4000 (nvidia-drm), Intel iGPU (i915), Lenovo USB-C Dock Gen 2, two LEN S27q-10 monitors (HDMI + DP MST)
• Builds cleanly (cargo check, cargo clippy, cargo build --release)
• Follows existing calloop timer patterns used elsewhere in tty.rs (e.g. VRR timers, redraw timers)

Alternatives Considered

• Fixing in smithay's ConnectorScanner: Would require changing the (Connected, Connected) => {} behavior to re-emit events, which could have broader implications for all smithay consumers
• Polling loop: Rejected in favor of bounded async timer to avoid blocking the event loop
• Infinite retries: Rejected — capped at 3 to avoid pathological cases

[coleleavitt]

Upstream Root Cause & Fix

After deeper analysis, the underlying issue is in smithay's ConnectorScanner in smithay-drm-extras. The (Connected, Connected) arm is a no-op, which means mode-list changes on already-connected connectors are silently ignored. When a USB-C dock connector reports as Connected before EDID is ready (empty mode list), and a later rescan finds modes populated, no Connected event is re-emitted.

I've filed an issue and PR upstream:

• Smithay issue: ConnectorScanner does not re-emit Connected event when modes change on already-connected connector Smithay/smithay#1922 — ConnectorScanner does not re-emit Connected event when modes change on already-connected connector
• Smithay PR: fix(drm-extras): re-emit Connected event when modes change on already-connected connector Smithay/smithay#1923 — Fix: compare mode lists on (Connected, Connected) and re-emit when they differ
• niri issue: Intermittent monitor activation failure on USB-C dock due to EDID race condition #3410 — Tracking issue for this bug with cross-compositor analysis

Why this PR is still valuable even with the smithay fix

The smithay fix ensures that when a rescan happens and modes have appeared, the compositor gets notified. However, this niri-level rescan timer is still needed because:

1. No guaranteed second udev event: The kernel may only fire one Changed event before EDID is ready — without a timer-based rescan, there's nothing to trigger a second ConnectorScanner::scan()
2. Defense in depth: The timer provides a bounded retry window (3 × 2s = 6s) regardless of udev event timing
3. Complementary: The smithay fix makes each rescan more effective (mode changes are detected), while this PR ensures rescans actually happen

[YaLTeR]

I suppose this needs updating to use the new Changed event instead?

[coleleavitt]

I suppose this needs updating to use the new Changed event instead?

yes I'll update it today if I get time; thanks