fix(agent): reapply volatile settings on device reconnection and retry disagreeing DPI writes

[isleofgreg]

Summary

Two related fixes to the volatile-settings re-apply path, found while instrumenting an MX Master 4 on a Bolt receiver (macOS 15.7) over several days of logged use. Standalone — applies to current master, no dependency on the smooth-scroll series.

• Reapply volatile settings when a device broadcasts its own reconnection (0x1d4b WirelessDeviceStatus). A mouse waking from a nap re-establishes its link and comes back with firmware-default volatile settings (sensor DPI, SmartShift, wheel mode). The re-apply relied on an offline→online transition between snapshots, but a nap short enough that no reconciliation observed the device offline produced no transition, and the one write a reconnect did get could race the link still stabilising — leaving the wrong DPI live until the next nap. The reconnection broadcast already reaches the inventory watcher; it now survives the settle drain and forces a volatile re-apply on the following snapshot (like a system wake, with a single link-race confirming pass instead of the boot ladder). Forced requests coalesce keeping the larger budget. Verified overnight: after a 12 h nap the DPI was rewritten and read back correct 14.5 s after display resume, with the pre-fix 8705 warnings gone.
• Retry a DPI write once when the read-back disagrees. Firmware still drowsy from a link re-establishment can ACK setSensorDpi yet keep the old value. The read-back already caught the mismatch but only warned; one re-write now separates the two silent-failure shapes — a dropped write lands on the retry, a value the device clamps deterministically comes back unchanged again and keeps the existing warning.

(An earlier revision also reclaimed a stale diverted wheel route on re-apply. It was withdrawn after review: the agent has no signal to tell an abandoned diversion from a live external consumer, and the identity-free reconnect path touches every online device. It will come back separately with an ownership check.)

Test plan

• cargo fmt --all --check; cargo clippy --workspace --all-targets -- -D warnings on aarch64-apple-darwin and x86_64-pc-windows-gnu for openlogi-device/openlogi-agent-core; RUSTDOCFLAGS=-D warnings cargo doc for those crates; cargo test -p openlogi-device -p openlogi-agent-core -p openlogi-agent — green. New unit tests cover the forced re-apply request coalescing and budget, the reconnection-broadcast path, and the single DPI retry (dropped write vs deterministic clamp). (One pre-existing tray:: test in openlogi-agent is timing-flaky under parallel execution on master — fix(agent): serialize the macOS tray tests that share NSWorkspace's center #1236 addresses it; passes serially.)
• On hardware: multi-day agent debug logs across naps and overnight sleep.

[greptile-apps]

Confidence Score: 5/5

[Medium risk] Adds device reconnection handling to volatile settings reapply logic.

The PR appears safe to merge; no actionable regression remains in the reconnect re-apply or DPI retry paths.

Findings

1.  Live Wheel Route Reclaimed ▶

Summary

This PR improves restoration of volatile device settings after wireless reconnection and retries DPI writes whose first read-back disagrees.

• Preserves WirelessDeviceStatus reconnect intent while coalescing HID++ lifecycle-event bursts.
• Forces a bounded re-apply on the next authoritative inventory snapshot and coalesces wake/reconnect confirmation budgets without shrinking them.
• Adds one DPI rewrite and verification attempt after an ACKed write leaves the old value.
• Adds focused coverage for reconnect re-application, budget behavior, and dropped DPI writes.
• The earlier wheel-route reclamation that caused the previous review finding has been removed.

Diagram

sequenceDiagram
    participant Device
    participant InventoryWatcher
    participant Lifecycle
    participant Orchestrator
    participant DPIWriter

    Device->>InventoryWatcher: WirelessDeviceStatus reconnect
    InventoryWatcher->>InventoryWatcher: Settle and coalesce event burst
    InventoryWatcher->>Lifecycle: DeviceWake
    Lifecycle->>Orchestrator: Queue forced re-apply budget
    InventoryWatcher->>Lifecycle: Authoritative snapshot
    Lifecycle->>Orchestrator: Refresh inventory
    Orchestrator->>DPIWriter: Reapply persisted DPI
    DPIWriter->>Device: Set DPI
    DPIWriter->>Device: Read back DPI
    alt Read-back disagrees
        DPIWriter->>Device: Retry set DPI once
        DPIWriter->>Device: Verify again
    end

Loading

_{Reviews (6) · Last reviewed commit: "fix(hid): retry a DPI write once when th..." · Reviewed by Greptile}

[isleofgreg]

Fair point on the reclaim — it treated every diverted route in the forced target set as abandoned, and the identity-free reconnect path makes that set every online device. There's no staleness signal available from the agent's side today (a live external consumer on another device looks identical), so rather than guess I've withdrawn that commit from this PR. It's now the two fixes with an unambiguous owner: the reconnection re-apply and the DPI write retry. I'll bring the wheel-route reclaim back separately once there's a defensible staleness check (or scope it to the agent's own startup pass), if that's of interest.

[isleofgreg]

Rebased onto master (post-#1286). The only change: the new replay_tests.rs fixture matches InventoryEvent exhaustively, so the DeviceWake variant this PR adds needed adding to its panic arm — folded into the first commit. No production changes.

[isleofgreg]

Rebased onto master (a92aa43, v0.8.6) after the orchestrator split. No functional change: the forced re-apply budget, the refresh path and the two entry points stay in orchestrator.rs; the plan_reapply rewrite and RECONNECT_REAPPLY_CONFIRM_RETRIES moved with plan_reapply into orchestrator/devices.rs; the tests into orchestrator/tests/reapply.rs. The DPI retry commit applied unchanged. Previous head for reference: 0114202.

[isleofgreg]

Rebased onto current master (1505c652). #1649 added an exhaustive InventoryEvent match in replay_tests.rs, so the new DeviceWake variant gets the same ignore arm there as SystemWake. No other changes; workspace fmt/clippy/tests are green.