Orbiscreen v0.39.6

Fixed

• fix: stop the AOA proxy from leaking streams and quiet the reconnect and unplug log noise

A long USB session produced three noisy symptoms and one real leak.

The real leak: every AOA proxy stream maps one client connection to one TCP
connection to the local daemon, and that connection is HTTP keep-alive. When a
client went away the daemon kept the connection open, so the host's reader never
saw EOF and the entry stayed in the map forever. The map filled to its 48-entry
bound and then evicted one stream per new connection, logging
AOA stream limit reached; evicted the oldest stream indefinitely — the log
showed stream ids climbing past 100 with 47 entries live.

• finish_proxy_stream now records a last-activity stamp per stream, touched on inbound client data, on frames written back, and on successful reads, and open_proxy_stream reaps streams idle for 30 s before applying the bound. An idle stream is closed and its close frame is sent, so a keep-alive connection that outlives its client can no longer pin a slot.
• The eviction warning is rate-limited to the first three occurrences, then drops to debug, so a genuine storm is visible once instead of once per connection.

The remaining noise was mislabelled rather than harmful:

• Unplugging the cable produced USB bulk read error on accessory: No such device (os error 19), which reads like a fault. ENODEV and ESHUTDOWN now log accessory disconnected; releasing the AOA bridge at info, because pulling the cable is the expected path.
• A failing write logged once per packet while the device drained, so a single unplug produced ten identical warnings. Only the first failure is logged and the rest are summarised by the existing giving up warning.
• A bare Io error: Broken pipe (os error 32) came from wayland-backend, not from Orbiscreen: the crate falls back to eprintln! unless its log feature is enabled. The workspace now enables log on wayland-client and wayland-backend, installs a tracing-log bridge in init_tracing, and mutes wayland_backend=off by default because Orbiscreen already reports the same failures with the output and operation that caused them.
• The Android input dispatcher retried the input WebSocket on every input event when it was down, and each attempt plus the HTTP fallback opened a new AOA proxy stream. Retries are now spaced 1.5 s apart, which is what reduced the stream churn above at the source.
• Version bumped to 0.39.6 across the Cargo workspace, Android (versionCode 187), PKGBUILD, Debian, COPR, and documentation badges.