* fix(capture): map the DMA-BUF frames niri actually sends
`uur run` reported "No Wayland frames arrived within 120 seconds" on a
session where the portal handshake succeeded. The session log showed the
stream never leaving negotiation:
portal session ready (video node 73)
stream state: connecting
stream state: paused
stream state: error - no more input formats
stream state: unconnected
Three separate defects in `uur-pw-capture`:
1. The format filter omitted `modifier`. niri marks that property
MANDATORY (SPA flags 0x18), and SPA filtering fails unless both sides
carry a mandatory property -- which surfaces as "no more input
formats". The filter now advertises LINEAR, the layout the frame
protocol already assumes.
2. `on_process` only handled `data != NULL` and `SPA_DATA_MemFd`. niri
delivers SPA_DATA_DmaBuf planes, which carry an fd and no pointer, so
every frame was dropped in silence. Since the session negotiates
LINEAR, that fd is directly mappable without GBM/EGL; the helper now
mmaps it. A non-linear modifier is refused instead of publishing
garbage.
3. DMA-BUF producers leave `chunk->size` and `maxsize` at 1. Both the
mapping length and the row count were derived from those, so the mmap
covered one byte and `rows` came out zero -- publishing an all-black
frame. Both now use the negotiated stride and height.
Verified on niri 26.04 with Wine 11.19: the stream reaches `streaming`,
~60 fps (1195 frames in 20 s), and the published frame decodes to the
real desktop rather than black.
* fix(capture): validate DMA-BUFs and recover interrupted screen streams
Use immutable transport files across geometry changes and remap Wine readers safely. Bound stream recovery and preserve portal stop-sharing intent. Refs #29.
* fix(capture): watch Mutter layout changes and support older build headers
* fix(capture): keep display changes outside the failure retry budget
* fix(capture): honor portal stop signals during GNOME recovery
Subscribe before starting on the portal-owned D-Bus connection and accept the real Session.Closed dictionary body. Correlate Mutter's close-before-MonitorsChanged ordering without restarting an explicit Stop Sharing session.
Verified in a private GNOME 51.1 namespace with real GBM DMA-BUF, 100/125 percent scaling, portrait and landscape, and native Stop Sharing. Existing 48 tests and strict clippy pass. Real Wine GDI reader captured 120 final-run frames across four layout changes with no hook fallbacks. No test scripts added.
---------
Co-authored-by: Xuc Pan <[email protected]>
* fix(wine): select the client root that actually holds bin/
`uur run` aborted with "WebView2 bootstrapper missing from the client
directory" on a complete install.
`find_client_dir` searched depth-first for `GameViewerServer.exe` and
returned the first hit, special-casing only a parent named `bin`. NetEase
ships the payload twice: as `<root>/bin/` and as a sibling
`<root>/<version>/`. When readdir order reached `<root>/<version>/` first,
that directory was returned as the client root. It has no `bin/`
subdirectory, so every `join("bin")` lookup missed -- the WebView2
bootstrapper and the terminal proxies -- and the session bailed out.
`deploy_terminal_proxy` then created a spurious `bin/` inside the version
directory and wrote proxies into it.
Now every directory holding `GameViewer.exe` is collected and ranked:
`bin/GameViewerServer.exe` scores highest because that is the layout uur
drives, then `bin/MicrosoftEdgeWebview2Setup.exe`, then a top-level
server. Ties break on path depth, then lexically, so the choice no longer
depends on readdir order. A versioned copy is still used when it is the
only candidate.
Adds four tests covering both layouts, readdir stability, and the
fallback.
* fix(wine): require complete payloads and support flat client layouts
---------
Co-authored-by: Xuc Pan <[email protected]>