fix(recording): stop wf-recorder start-race producing 0.5s fragment files

On Wayland the video start path signalled any running wf-recorder with pkill -INT and then waited a fixed 0.1s before launching a new one. Releasing the compositor capture takes longer than that, so the new wf-recorder launched into a still-held grab, died immediately, and ffmpeg finalized a ~0.5-second fragment .mkv. Seeing nothing record, I'd press the key again and burst out a pile of fragments until the capture finally freed.

I replaced the fixed wait with a poll that waits until no wf-recorder remains (capped at 2s), the same move the stop path already made when it swapped its own fixed wait for cj/recording--wait-for-exit. Two supporting changes came with it. wf-recorder stderr no longer goes to /dev/null, so a failed grab shows up in the process buffer. And the sentinel now tells a ~0.5s failed start from a real stop, saying "failed to start, try again" instead of silently clearing, so one failure doesn't snowball into a blind-retry burst.

The poll and the fail-fast timing are unit-tested. The live capture needs a manual check.

2 files changed, 64 insertions, 0 deletions

diff --git a/tests/test-video-audio-recording--build-video-command.el b/tests/test-video-audio-recording--build-video-command.el
index 4f290978..1ffce95b 100644
--- a/tests/test-video-audio-recording--build-video-command.el
+++ b/tests/test-video-audio-recording--build-video-command.el
@@ -27,6 +27,14 @@
         (should (string-match-p "-i pipe:0" cmd))
         (should (string-match-p "-c:v copy" cmd))))))
 
+(ert-deftest test-video-audio-recording--build-video-command-normal-wayland-keeps-wf-recorder-stderr ()
+  "Wayland command does not discard wf-recorder stderr, so a failed grab is diagnosable."
+  (let ((cj/recording-mic-boost 2.0)
+        (cj/recording-system-volume 1.0))
+    (cl-letf (((symbol-function 'executable-find) (lambda (_prog &rest _) t)))
+      (let ((cmd (cj/recording--build-video-command "mic" "sys" "/tmp/out.mkv" t)))
+        (should-not (string-match-p "2>/dev/null" cmd))))))
+
 (ert-deftest test-video-audio-recording--build-video-command-normal-x11-uses-x11grab ()
   "X11 command uses ffmpeg with x11grab, no wf-recorder."
   (let ((cj/recording-mic-boost 2.0)
diff --git a/tests/test-video-audio-recording--start-race.el b/tests/test-video-audio-recording--start-race.el
new file mode 100644
index 00000000..36ea8595
--- /dev/null
+++ b/tests/test-video-audio-recording--start-race.el
@@ -0,0 +1,56 @@
+;;; test-video-audio-recording--start-race.el --- start-race fix tests -*- lexical-binding: t; -*-
+
+;;; Commentary:
+;; Tests for the wf-recorder start-race fix: the poll that waits for a dying
+;; wf-recorder to release the compositor capture before launching a new one, and
+;; the fail-fast timing predicate that tells a 0.5s failed start from a real
+;; recording.  The pgrep wrapper and the live start/stop wiring are exercised in
+;; the daemon, not here; the poll is tested with an injected predicate so no real
+;; process is needed.
+
+;;; Code:
+
+(require 'ert)
+
+;; Stub dependencies before loading the module.
+(defvar cj/custom-keymap (make-sparse-keymap)
+  "Stub keymap for testing.")
+
+(require 'video-audio-recording)
+
+(declare-function cj/recording--start-failed-p "video-audio-recording-capture" (elapsed threshold))
+(declare-function cj/recording--wait-for-no-wf-recorder "video-audio-recording-capture" (timeout-secs &optional running-p))
+
+;;; ------------------------- cj/recording--start-failed-p ---------------------
+
+(ert-deftest test-recording-start-failed-p-short-exit-is-failure ()
+  "Normal: an exit well before the threshold is a failed start."
+  (should (cj/recording--start-failed-p 0.5 1.5)))
+
+(ert-deftest test-recording-start-failed-p-long-run-is-not-failure ()
+  "Normal: a long-running recording that ends is not a failed start."
+  (should-not (cj/recording--start-failed-p 30.0 1.5)))
+
+(ert-deftest test-recording-start-failed-p-at-threshold-is-not-failure ()
+  "Boundary: an exit exactly at the threshold is not counted as failed."
+  (should-not (cj/recording--start-failed-p 1.5 1.5)))
+
+;;; -------------------- cj/recording--wait-for-no-wf-recorder ------------------
+
+(ert-deftest test-recording-wait-for-no-wf-recorder-clears ()
+  "Normal: returns t once the injected predicate reports wf-recorder gone."
+  (let ((n 0))
+    (should (cj/recording--wait-for-no-wf-recorder
+             2.0
+             (lambda () (setq n (1+ n)) (< n 3))))))
+
+(ert-deftest test-recording-wait-for-no-wf-recorder-already-clear ()
+  "Boundary: an already-clear predicate returns t immediately."
+  (should (cj/recording--wait-for-no-wf-recorder 2.0 (lambda () nil))))
+
+(ert-deftest test-recording-wait-for-no-wf-recorder-times-out ()
+  "Error: a predicate that never clears returns nil at the timeout."
+  (should-not (cj/recording--wait-for-no-wf-recorder 0.15 (lambda () t))))
+
+(provide 'test-video-audio-recording--start-race)
+;;; test-video-audio-recording--start-race.el ends here