[chhoumann]

Fixes #46.

Problem

capture record --backend screencast rejected any take that ended on a still screen. A real take with 1s of changes followed by 3s of stillness failed with Encoded duration 7.7s does not match wall-clock 4.51s.

Root cause

The screencast backend builds an ffconcat list in which every frame gets its real on-screen duration. It then repeats the last frame so ffmpeg honors that frame's duration. On ffmpeg 7 and later, the frame after the last entry is timed like the gap before it, so the repeat adds the final still stretch a second time. ffmpeg 4.4 and 6.1 held the repeat for one 40 ms tick instead. Measured with ffmpeg 8.0.1:

ffconcat list	encoded duration
a 0.5s, b 3s, repeat b	6.5s
a 0.5s, b 3s, no repeat	1.0s
a 3s, b 0.5s, repeat b	4.0s
a 0.4999s, b 3s, repeat b, input -t 3.5	6.5s
a 0.4999s, b 3s, repeat b, output -t 3.5	3.5s

Takes that kept changing until the end only went over by one short frame gap, which the check tolerates. A take whose screen never changed was also affected, because its single fallback screenshot was doubled the same way.

Fix

Trim the encoded output to max(stop - start, 1/fps) with an output-side -t. The repeat still makes ffmpeg hold the last frame, and the trim ends the video where the take ended. An input-side -t is not enough. It becomes a trim before the fps filter and is defeated when the per-frame durations, each rounded to 4 decimals, add up to one tick less than the cut (the 0.4999 row above). The duration check is unchanged. CapturedSource.inputArgs is renamed sourceArgs because it now carries that output option.

CI and the release job now install ffmpeg. The runner images ship none, and the new test runs a real encode.

Verification

Real screencast takes against a live Obsidian 1.13.7 window (capture launch plus capture record), using the packed tarball with the final fix:

still 1-5.webm (1s of changes, 3s still)  4.4/4.43, 4.5/4.46, 4.4/4.44, 4.5/4.47, 4.4/4.40 (duration/wall)
still.mp4      (same)                     4.4/4.43
busy.webm      (changes until the end)    1.4/1.42
idle.webm      (no change at all)         2.0/2.01
short.webm     (instant take)             0.1/0.00

The new test drives startRecording and stop with real JPEG frames and a still ending, then runs the real ffmpeg encode and duration check. Before the fix it fails with Encoded duration 3.8s does not match wall-clock 2.10s.