Re: [RFC] WebRTC console for VMs via QEMU's D-Bus display

[Dmitry R <rdmitry@taktcycles.com>]

Hi Alexandre, hi all,

A follow-up to the standalone/APT update: stock4.3 now includes the
reference-detail fix and a completed-action input mode for both VM and LXC.

Genuine bitrate changes could still reset hardware rate control and replace
unchanged reference detail, producing changing sharpness or horizontal
bands. The worker now keeps the encoder settings and budgets raw pictures
before encoding. Under constraint it reduces motion cadence; media timestamps
preserve playback time. An abrupt capacity drop can still cause recovery
pauses. NVENC H.264/HEVC and browser H.264 checks are included; Intel/AMD
hardware and interaction latency were not remeasured for this change.

For input, reliable delivery alone was insufficient: a delayed key release
could cross the guest's native repeat threshold. Read-only QEMU/guest
observations found complete press/release pairs but stretched holds. Exact
typing is now the default. Each press or actual client repeat is one complete
key action with its modifiers, executed locally by the worker. The first
character is immediate. Physical holds/Raw remain available for applications
which require a key to stay down.

In Firefox with a real Sway/Qt LXC desktop, a 900-ms control-delivery gap gave
8 characters from one short source press in the old mode and 1 with exact
typing. A heartbeat-only guard still gave 5 after a recent confirmation;
exact typing gave 1. Six confirmed source events produced six characters.

Stock4.3 applies the same approach to short mouse/trackpad/touch clicks,
including their coordinates. A short click is sent on release; movement of
4 CSS pixels or a 250-ms hold starts a drag/hold. Drag motion and release
stay ordered through acknowledgement, with a per-viewer coordinate filter
shared by both channels. Normal idle motion retains the fast pointer path.
Fractional trackpad deltas and horizontal scrolling are supported.

The same injected gap gave two activations of a Qt auto-repeat button with
physical edges plus watchdog, versus one with the complete click. Double
click, right click, holds, dragging and two-axis scroll passed. These are
browser-event tests, not physical trackpad or end-to-end latency benchmarks.
Native double-click timing can still change with delivery jitter. A 200-ms
confirmation timeout can end a drag early; stale snapshots never re-press it.

Source, five matching debs, guides and detailed qualification:
https://github.com/rdmitry0911/qsm-rd/releases/tag/v0.4.0-qsm-direct-20261024-stock4.3

Signed APT repository:
https://rdmitry0911.github.io/qsm-rd/

Update installed QSM profiles together, then reload PVE with cache bypass
and reopen consoles. No guest reboot or agent update is needed. This release
has 73 bridge tests, 22 CTest gates and worker/fake-QEMU E2E coverage.

The standalone release remains separate from the native shared-bus/API/UI
proposal. It does not change the proposed upstream interface or establish
maintainer agreement. Feedback on bus ownership/lifetime, configuration and
the service/API boundary would still be welcome.

Thanks,
Dmitry