VCP and display topology fixes (from #513), plus opt-in live read

[mgth]

First extraction from #513: the VCP/topology commit (2764856), reviewed fix by fix, cherry-picked with two adaptations, plus a follow-up commit restoring OSD monitoring as an explicit bounded mode.

From #513 (adapted)

• dxva2 handle leak: VcpControl now disposes its IVcpTransport; the probe task is awaited before the transport goes away, and disposed controls are evicted from the cache.
• Registry key name over-read: GetHKeyName copied 2× the buffer size (Marshal.Copy counts chars, the buffer counts bytes); replaced by a bounded decode in MonitorDeviceHelper, with a unit test. (The buggy original still lives in HLab.Core — tracked separately.)
• Clone mode targeted the wrong panel: DxVa2VcpTransport picked _physicalMonitors[0] for every device sharing an HMONITOR; it now matches the physical monitor by description. When two identical descriptions make the choice ambiguous, VCP is disabled for that device rather than guessed — behavioral change to be aware of.
• DDC polling loop: CommandWorker/MonitorLevel no longer re-enqueue forever; levels go quiet once converged. The hand-tuned slow-monitor behavior is preserved verbatim: read-before-write, convergence-based Moving, 10 retries / 100 ms backoff / auto-disable, latest-value coalescing. Closing one monitor's panel no longer kills the shared worker.
• Topology application is transactional: WindowsDisplayController stages, commits, and rolls back via ResaveCurrentConfiguration instead of leaving the registry half-written when an AttachToDesktop fails mid-batch.
• ActiveConnection replaces order-dependent Connections[0] lookups (also removes potential IndexOutOfRange).
• Topology failures are no longer silent: SetPrimary/Attach/Detach/SetLocations failures show a dialog. Adapted from the PR version: OS-agnostic wording (the controller also targets Linux/kscreen) and Escape kept working (IsCancel moved to the visible OK button).

New: opt-in live read

The perpetual polling removed above was also what kept sliders in sync with changes made on the monitor's own OSD during white-balance sessions. That comes back as an explicit mode: a Live read toggle on the Image card re-reads every level once per second and auto-expires after two minutes, so an abandoned panel never polls the DDC bus forever. The _queued CAS guard coalesces ticks, so a slow bus cannot pile up requests. Interval/duration are constants on VcpControl (LiveReadInterval, LiveReadDuration).

Tested

• DisplayLayout.Tests: 72/72 (includes new WindowsMonitorSafetyTests)
• Plugin.Vcp.Avalonia.Tests: 60/60 (includes new VcpSafetyTests + 2 live-read tests, stable over 5 runs)
• Full UI build on Linux

Needs testing on Windows

• VCP sliders on a slow monitor: drag feel should be identical or slightly better (immediate wake instead of waiting for the polling pass)
• Live read toggle: OSD changes appear within ~1 s, button unchecks itself after 2 min
• Clone mode: right panel is driven; ambiguous twin monitors now show no VCP instead of driving panel 0
• A failing attach/detach/make-primary shows the failure dialog and rolls back

🤖 Generated with Claude Code