feat(lbm-agent): the reconciler — display changes and engine hooking as one state machine

[mgth]

Phase 3 of the v6 plan (docs/v6-architecture-plan.md on v6): the heart of lbm-agent, the reconciler — what the C# UI decided in DisplayChangeCoordinator, EngineController and MainService's daemon-event routing, as one state machine without I/O. The design and its deviations are written down in docs/v6-agent.md (French, like the plan).

Shape

• Input in (a display change, a hook event, the user's Start/Stop/Refresh, a timer coming back), Effects out in order (Start, Stop, SaveEnabled, SaveLayout, a timer, a process seen). The layout is read and rebuilt through a World trait.
• The runtime owns the clock and the I/O; the reconciler owns the decisions. Timers are effects, so tests run on a virtual clock, not on shrunk real delays.
• One input is handled to completion before the next, and Effect::Start carries no zones: the runtime computes them from the current layout when it sends. The stale Start of Cursor gets trapped at screen edges after disconnecting and waking from sleep #607 cannot exist, and fix: never hook a layout the desktop no longer matches (#607) #609's LatestRequestGate is not needed here.

Rules (from C#), and the deviations

The C# rules are kept: 300 ms debounce with generations, settle loop (100 ms, 8 steps), idempotence guard that re-hooks instead of rebuilding (#412), Refresh past the guard (#443), user Start/Stop recorded before (un)hooking, foreign layouts simulated never adopted, no re-hook while asleep / against the user / over a running engine / while the watchdog drives, the post-resume watchdog (500 ms steps, 3 stable, 6 restarts, 60 steps), and the "Stopped right after Connected" hand-over.

Deliberate deviations, documented:

• Cursor gets trapped at screen edges after disconnecting and waking from sleep #607 included (fix: never hook a layout the desktop no longer matches (#607) #609's behaviour): a fresh layout the user turned off sends Stop, instead of leaving the hook running the layout of the desktop that went away. Master's C# sends nothing.
• One Start per event: a Start the same event just sent is given one watchdog step before being re-asserted. C# sends it twice back to back (its watchdog checks before the hook's Running can arrive).
• A superseded watchdog no longer clears the newer one's "active" flag.

Tests

lbm-agent/tests/reconcile.rs: the 10 DisplayChangeCoordinatorTests and 19 EngineControllerTests, one Rust test per C# test, plus the hook routing. Where the C# test counted calls into its neighbour in isolation, the integrated Rust test checks what the call does, and says so.

Not wired to anything yet: the runtime (hook connection with --fake-hook, supervision per D5, Linux sources, persistence) comes next.

🤖 Generated with Claude Code