Video player audio playback and sync

• Type: feature
• Status: done — audio implemented, green on v37
• Owner: —
• Opened: 2026-06-21
• Relation: Follow-up from validating feature-demo-video-player.

Problem

The console video player compiles and its video chain works with a generated MP4 on pinned v32: ffmpeg spawns, raw RGB frames stream through the pipe, ANSI rendering emits frames, and the child is waited. The original player spec also listed sibling audio playback (aplay or similar), but examples/player/player.pas does not spawn or synchronize audio.

Direction

• Decide the host-only audio helper (ffmpeg audio pipe, aplay, or another tiny external player).
• Spawn the audio process through the PAL process API.
• Tie pause/quit behavior to both video and audio children, or explicitly document the first supported limitation.

Acceptance

• A short generated video with audio starts both children and exits cleanly.
• Quit/EOF closes pipes and waits for all spawned children.
• The no-audio-video path still works.

Log

• 2026-06-21 — Opened after Track B validated the existing video-only chain.

Blocker found + fixed (2026-06-22, Track B)

Investigated "what's holding us back" — the concrete blocker for a second concurrent child (the audio process alongside video) was an fd leak / EOF deadlock, not exec itself:

ExecutePipeline created its pipes with flags=0 (no O_CLOEXEC). The audio child spawned second inherited the video child's stdin write-end, so closing the parent's copy never delivered EOF to the first child and wait() hung forever. Reproduced with two /bin/cat children (first child's wait deadlocked).

Fixed in lib/rtl/sysutils.pas (commit ab71066): pipes are now O_CLOEXEC; dup2 in the child clears CLOEXEC on the wired 0/1 fds so they survive exec, while all other inherited pipe fds auto-close on exec. Regression test test/lib_process_multi.pas (two concurrent children) wired into make lib-test.

So concurrent children now work. Remaining for THIS ticket = the actual audio implementation: spawn an audio sibling (e.g. ffmpeg ... | aplay or a second ffmpeg → aplay) through ExecutePipeline, feed/sync it against the monotonic clock, and tie pause/quit/EOF to both children. (Secondary, non-blocking: the vfork child still runs Pascal with locals on the shared stack — see feature-sys-process-spawning hardening note.)

DONE (2026-06-22, Track B)

Implemented audio as a single ffmpeg → ALSA sibling (-vn -f alsa default): no aplay, no shell → one child with a known pid, no grandchildren to orphan. Added PalKill(pid, sig) to the PAL (sys_kill, all 4 posix arches; esp stub).

• Pause (space) SIGSTOP/SIGCONTs the audio child in step with the video timeline.
• Quit/EOF SIGTERMs + reaps both children; pipes closed; clean exit, no orphans.
• No-audio-stream / no-device path stays graceful (video unaffected). Verified headless on a generated 1s clip (with + without audio): exit 0, no orphan ffmpeg. Real sound needs a device + TTY (manual, like the original player). Acceptance met. Commit in the player-audio batch.