Renderd v1.0.0

First stable release. Streaming now holds up end to end: low latency, clean quality across screen changes, and sessions that no longer stall after ~10 minutes. This release adds loss recovery by retransmit, resolution that follows the link, and a still screen that always ends on the current picture.

Added

• Nack control message (renderd.proto): the viewer asks the host to resend specific lost fragments. Backward-compatible, so a minor bump.

Performance

• Lost fragments are resent, not paid for with a keyframe. A lost fragment used to go
  unnoticed until four incomplete frames had piled up (on a mostly clean link, never) while the
  frames after it were decoded against a reference the decoder never got, smearing the picture
  until the next keyframe, which on a 3 Mbps link is a quarter of a second of link time. The
  viewer's new ReceiveWindow hands frames to the decoder strictly in order, notices a missing
  fragment as soon as anything after it arrives, and asks for exactly those fragments in a new
  Nack message; the host answers from the last second of datagrams it keeps, and resends the
  newest frame's last datagram after 100 ms of quiet so a lost tail is noticed too. Only a frame
  still missing two round trips plus 120 ms later costs a keyframe, and nothing that depends on
  it is decoded meanwhile. Over loopback QUIC with 5% loss every frame arrives in order with no
  keyframe request. (renderd-frame::receive,
  host::network::retransmit)
• Bitrate steers on real packet loss. The ABR loop used the viewer's frame-loss rate per
  100 ms report (one lost frame among six read as 14%, two as 25%), which walked the bitrate
  down to the floor on any Wi-Fi link. The host now reads QUIC's own packet-loss counters over the
  last second; loss above abr.loss_threshold only stops probing, and the bitrate is cut at five
  times that (or on queuing delay, as before). A 4 Mbps link with 1-3% random loss now holds
  3 Mbps in simulation instead of sinking to 1.5.
• Encoded resolution follows the link. With the default 10 Mbps ceiling a 3 Mbps link still
  got 1080p, and at 2 Mbps a full-screen change of 1080p text is an 83 KB frame (a third of a
  second) against 25 KB at 544p, and no VideoToolbox setting bounds it with low-latency rate
  control. The host now steps the encoder down to 896, 720 or 544 lines when the bitrate falls
  below 75% of what the size needs for 2 s, and back up a rung after 8 s of room, without
  touching capture (VideoToolbox scales). The viewer follows from the stream. On by default;
  host.adaptive_resolution = false turns it off. (host::scale)
• A still screen ends on the current picture. After a big change the low-latency rate
  controller drops the frames that follow it (six of eight in a 2 Mbps measurement), so a scroll
  that stopped inside that window left the viewer on a stale frame until the next change. The
  host now re-encodes the newest captured surface after 60 ms of stillness until two passes come
  out of the encoder. (host::refine)
• HEVC on Windows where the GPU decodes it. --codec auto now offers HEVC first when the GPU
  has the HEVC Main profile and the Media Foundation HEVC decoder runs on it, H.264 otherwise.
• Frames are presented on arrival. The viewer presented through request_redraw, which on
  Windows is a WM_PAINT, the lowest-priority message. It now presents from the wake itself, and
  connects at network.quic_mtu instead of QUIC's 1200-byte minimum.

Fixed

• Keyframe loop on slow links. Frames held up on the wire behind a large one arrive together
  and span more than the 150 ms backlog threshold; the viewer took that for decode falling behind,
  skipped them and asked for another keyframe, which bunched the next frames the same way. A pass
  now only counts as a backlog if the previous one spent 50 ms or more decoding.
• Capture rate at session start. The first vsync report rebuilt the ScreenCaptureKit
  pipeline at every session start and replaced the negotiated frame rate with the viewer's
  refresh rate, lifting a 30 fps session to 60. Vsync now only slows capture, and the rebuild runs
  off the async workers.

Documentation

• README: refreshed status banner, Windows viewer description (DXVA decode, D3D11 presenter, Nack recovery), roadmap and release history.
• release.yml: the release-notes extraction used an awk regex that failed on the tag name, so v0.9.2 and v0.10.0 built but never published. It now matches the heading as a literal string.