yasl.git/src, branch one-loop

End every drag with one event

2026-10-16T10:26:20+00:00

A drag that dropped on no YASL surface, such as one dropped on the desktop or canceled, ended unheard: the public drag start set no end listener. An application then waited for the drag's end forever. The public layer now listens for the end. When no surface got the drop, the origin gets DRAG with the new action YASL_DROP_CANCEL. Co-Authored-By: Claude Opus 5.5

Pace frame requests by the frame presented after them

2026-10-16T06:33:39+00:00

A frame request now only records intent. Each loop step resolves it after the callbacks that present, so a request made before the present waits for that frame. An animation that asked first got two REFRESH events per frame on X11 and with GL. With no frame in flight, the REFRESH comes one refresh interval after the last present or answer, so a request without a present cannot spin. Queued events now go before an owed REFRESH. Input reaches the application before it draws, and a queued REFRESH pays an owed one for the same surface. The loop prepare only peeks at a lost event, so deliver can report it as a final loss. A failed push from another thread wakes the loop. Signals and posts run at the end of a step, after descriptors and timers. A signal handler now sees the output that came with it. A render-thread GL swap counts as in flight on Wayland. Co-Authored-By: Claude Opus 5.5

Make the loop public and let the display drive it

2026-10-15T17:38:14+00:00

An application that has its own descriptors and timers needed a second loop next to yasl's, and could not see yasl's key-repeat timer, clipboard pipes or reconnect timer. Now there is one wait: the application owns a yasl_loop, and yasl_init can run the display on it. - Public yasl_loop_*: descriptors, timers and signals. Descriptor callbacks return void. Timer IDs are 64 bits and never reused; a zero delay runs in the next step. A nested step or run fails with EBUSY, a step off the loop's thread with EPERM, and a quit made before a run ends it at once. An interrupted wait keeps its deadline. Creating a loop fails when it cannot make its wake pipe. - The display drives the loop through prepare, wait and check hooks instead of wrapping its step. Wayland finishes its read before any callback runs, so loop callbacks may create surfaces and do roundtrips. Readiness (queued events, an owed REFRESH) is derived in the step, so the wait never blocks over pending work. A draw that cannot proceed for want of a buffer is no longer counted as work. - yasl_init takes struct yasl_options. With loop and on_event, each step delivers events to on_event and yasl_wait_event fails with EBUSY. EALREADY only for the same options, EBUSY for others. - yasl_surface_request_frame: a REFRESH once the window can show a new frame, for pacing animation. - DISPLAY_LOST reports whether the display reconnects. Pull waits still fail with ENOTCONN after a final loss. - Events pushed from another thread wake the loop. Co-Authored-By: Claude Opus 5.5

Add damage presentation and reduce CPU surface latency

2026-10-15T10:19:58+00:00

Add yasl_cpu_present_damage to submit partial damage rectangles. Skip commit timers for CPU surfaces to prevent display delay. Remove frame callback gating in surface_draw_frame when free buffers exist. Flush the display immediately on CPU present. Forward tearing option to window specifications.

Revert "Resend the text input caret when the surface scale changes"

2026-10-14T12:29:32+00:00

This reverts commit b2f963670845e7e1bcb9f3b141e12ea7ba7560a4.

Resend the text input caret when the surface scale changes

2026-10-14T12:27:42+00:00

The caret's logical rectangle depends on the surface scale. A scale change left the compositor with the old rectangle until the app sent a new caret. Send the caret again on every rescale while text input is active. Co-Authored-By: Claude Haiku 5.5

Map text purposes to Wayland content purposes

2026-10-14T12:24:11+00:00

The Wayland backend passed yasl's purpose values straight to set_content_type. The values differ: yasl TERMINAL (3) reached the compositor as the protocol's NUMBER. Map each yasl purpose to its protocol constant. Unknown values fall back to normal text. Co-Authored-By: Claude Haiku 5.5

X11: survive server loss, chain X errors, fix popups and pointer requests

2026-10-14T11:22:39+00:00

- Losing the X server queues DISPLAY_LOST instead of exiting. A process-wide IO error handler returns for YASL displays, so Xlib calls the per-display exit handler and returns; later waits fail with ENOTCONN, and dispatch and flush make no Xlib calls on a lost display. - Errors on other displays chain to the previous handler, restored when the last YASL display closes. - MIT-SHM attach is synced; a refused attach detaches and falls back to XPutImage. Shared images are used only for 32-bit TrueColor 0xAARRGGBB. - Undecorated windows set MWM_HINTS_DECORATIONS and leave functions alone. - Fixed-size windows keep their min/max hints in step with set_size. - Swap interval is stored per surface and applied when it becomes current; gl_make_current(NULL) without EGL succeeds as a no-op. - XInitThreads runs before the first connection. - Move, resize, window menu, drag start and place-under-pointer need a held pointer button (ESTALE). Move and resize also need _NET_WM_MOVERESIZE on the root (ENOTSUP) and send the held button. - _GTK_SHOW_WINDOW_MENU carries the XI2 client pointer and root coordinates. - Popups always present through CPU and grab the pointer (XI2 device grab when available); a press outside every YASL window closes the topmost popup. - Drag start while active returns EBUSY; drag carry without a drag EINVAL. Co-Authored-By: Claude Haiku 5.5

Count touch as a press, report a missing bell, and test X11 INCR

2026-10-14T11:21:13+00:00

- Wayland: a touch contact is a press for move, resize, menu and drag again, and a compositor without xdg-system-bell reports ENOTSUP. - yasl_surface_set_cursor no longer reports a stale errno. - The X11 input test expects the TIMESTAMP and MULTIPLE targets and drives an INCR transfer instead of expecting INCR to be rejected. Co-Authored-By: Claude Opus 5.5

Wayland: fix GL thread races, reconnect ordering and sizing bugs

2026-10-14T11:17:37+00:00

- GL swaps on a rendering thread no longer touch frame callbacks or presentation feedback. They set a per-surface flag and wake the loop; the owner thread arms both at its next flush and commits, so the frame callback fires without waiting for the application's next swap. - Clear the offline flag only after the reconnect rebuild succeeds, and not when the connection was lost again during it. - make_current fails with ENOTCONN for a window without a context. Detaching is a no-op when nothing is current or the EGL display is gone. - Window move, resize, window menu and drag start need a pointer button or tablet tip that is still held; otherwise they fail with ESTALE and use that press's serial. Popups fail with ESTALE without an input serial and with ENOTCONN while the display is lost. - An oversized device size shrinks to the buffer limits, keeping the aspect ratio, instead of closing the window. - FIXED_SIZE windows pin xdg min and max on every set_size. The 64px minimum applies only to compositor-suggested sizes, not to application sizes. - The keyboard-enter serial becomes the input serial. Clipboard copy with no keyboard focus fails with ESTALE. - Relative-pointer windows relock after a reconnect rebuild. - errno: EBUSY for an active drag, ENOTSUP for place_under_pointer and for GL without EGL, EINVAL for drag carry without a drag. - Default window title and app ID are "YASL". - Pointer motion and button events carry unrounded device-pixel positions in precise_x/precise_y with precise_position set. Co-Authored-By: Claude Haiku 5.5