All Addresses field. The listener has always bound every interface at once; the UI only ever named one of them. The new read-only field lists a working URL for every IPv4 address the host answers on, best first, so a VPN, LAN or hotspot address can be copied without guessing which one the viewing device can route to.
Fixed
Source Camera now persists with the scene. Three faults compounded: the chooser was never passed to RegisterStringChooser, so nothing was written to the scene at all; the value stored was the popup label, which embeds the render texture size and so stopped matching whenever that changed; and Init scanned once, before the atom owning the camera necessarily existed. The chooser now stores a stable atomUid/cameraName key while the popup keeps showing the descriptive label, and a selection naming a camera that has not loaded yet is retried once a second for twenty seconds instead of being discarded as invalid.
Access Key and Port were impossible to edit. Both were built with CreateTextField, which renders a value but accepts no keystrokes, so the access key could never be set and the port was stuck at 8088. Both now graft an InputField onto the field and hand it to the storable, which is what routes typing into val.
Port now persists with the scene. It was the one control never passed to RegisterString, so a saved scene always came back on the default port.
OBS URL no longer prefers a VPN address over the LAN one. Host selection returned the first address in the 100.64.0.0/10 carrier-grade NAT range, where overlay VPNs hand out node addresses, so a machine running one advertised its overlay address even when an ordinary LAN address was available. Candidates are now ranked rather than taken first-match: 192.168/16, then 10/8, then 172.16/12, then everything else, with link-local 169.254/16 last. Ranking 172.16/12 below the others keeps a Hyper-V, WSL or Docker virtual switch from being advertised in place of the real LAN address. This affects which URL is displayed only — the listener has always bound every interface at once.
Threaded JPEG encoding. A built-in baseline encoder (4:2:0, Annex K tables, AAN DCT) runs on a worker thread, so the main thread pays for a buffer copy rather than a full encode. Measured in VaM at a 600x1400 source this gives 30 fps against 20 fps for Texture2D.EncodeToJPG on the main thread. It does not match the throughput of the earlier sync-capture-plus-worker configuration, which was removed because it exhausted the heap.
Flip Output Vertically toggle, on by default. Readback data starts at the bottom row while JPEG scanlines run top-down; EncodeToJPG handled this implicitly, the threaded encoder needs it stated. Persists with the scene.
Width, Height, JPEG Quality and Target FPS persist with the scene, and nothing overwrites them. Selecting a camera used to assign Width and Height from the source resolution, which discarded any downscale on every scene load since restoring a scene re-selects the camera. Width is now only clamped down to the source when it exceeds it. Any resolution and any aspect ratio are accepted.
GPU downscale. Width and Height now do something for a camera that already has a targetTexture: the source is blitted into the output RenderTexture before readback. Every downstream cost is linear in pixel count, so halving each axis quarters the readback bandwidth, the copy and the encode together. Any resolution and any aspect ratio are valid: the largest centred region of the source that already matches the target's aspect is sampled, so the output always fills the frame and is never stretched. A target that does not match the source aspect crops rather than distorts.
Removed
Sync Capture. It existed to beat the single-readback throughput ceiling by stalling the GPU, paying game framerate for stream framerate. Queued readbacks reach the same throughput without the stall, and after the allocation fix the toggle could only cost framerate: it forced EncodeToJPG back onto the main thread. With it goes the last main-thread encode, the intermediate Texture2D, and EncodeToJPG itself.
Changed
Up to three readbacks are queued rather than one at a time. A single request takes two to three frames to return, which capped throughput near game fps / 3 regardless of Target FPS. Requests complete in submission order, so the queue drains from the head, and each carries the width, height and colour space it was taken at.
Capture rate is clamped to the game's own framerate cap, discovered at runtime from QualitySettings.vSyncCount and the display refresh rate, falling back to Application.targetFrameRate. Nothing is hardcoded, and the cap is re-read once a second so changing it mid-session takes effect. Every capture costs a GPU readback and, on the manual-render path, a second scene render, so capturing faster than the game was allowed to run spent GPU the user had asked not to spend. The Status line now reports the detected cap beside the game framerate.
Capture is skipped entirely with no clients connected. The readback, gamma pass and encode previously ran regardless, so enabling the plugin cost framerate for output nobody was receiving. A readback already in flight is still retired.
The Status line reports main-thread and worker time separately — main and jpeg replace the single encode figure — and appends a dropped-frame count when the worker cannot keep up with the capture rate.
The server copies frames in and out under its lock, so a frame buffer is never read while it is being replaced, and nothing on the capture path allocates per frame.
Readback dimensions are taken from the source RenderTexture rather than assumed to match the configured Width/Height, so a mismatch no longer garbles the stream.
Fixed
RenderTextures are released only once their readback has returned. RebuildPipeline() and OnDestroy() freed _outputRT with a request possibly still pointing at it, which is an access violation rather than an exception, on the two paths exercised most during development: resolution change and plugin reload. At teardown anything still pending is left alone; leaking a few MB beats killing the host.
Sync Capture no longer feeds the worker thread. Doing so called Texture2D.GetRawTextureData() every frame, which returns a freshly allocated array on this Unity build -- there is no NativeArray overload before 2018.2. At a 600x1400 source that is 3.4 MB per frame onto the large object heap, which Mono does not compact; sustained, it exhausted the address space and killed the host with VirtualAllocRemap failed. Sync Capture encodes on the main thread again, as it did before. The async path is unaffected: it copies into pooled buffers and allocates nothing per frame.
Capture runs even with zero clients connected is resolved; removed from the known limitations.
