[httpsphl]

Closes the last gap of slice 3 (spec 14.1). On Windows, a bot's Job Object dies with the daemon. On Linux under systemd, the cgroup does. On macOS, launchd ends only the daemon's own process group, so a bot's browser, and anything claude started, outlived a crashed daemon. A daemon started by hand had the same gap on Linux.

How

• Writing groups down: ProcessJob::track(&command) records each bot and browser group in <home>/run/groups (0600), one JSON line with the group id and the arguments. The runtime and the browser call it right after assign. On Windows it does nothing.
• Ending what was left: platform::track_groups(<home>/run) runs at daemon start, right after the instance lock and before any bot starts. It kills (SIGKILL to the group) every recorded group whose leader still runs all of its recorded arguments (ps -ww -o command= -p), then starts the list over.
• Why another program is never hit: the arguments carry paths and ids unique to the bot (browser profile, mcp.json, session id), so a reused pid that belongs to someone else never matches.
• Logging: only the count of ended groups is logged.

Tests

• Unit test: a group written down and left running is ended. An entry naming this test's own pid with arguments it never had is left alone. The list starts over.
• Platforms: Windows (fmt, clippy, lib) and Linux (Docker: clippy, lib, supervisor and browser tests) are green. macOS runs only in CI.

Manual test (macOS)

1. With a bot's browser open, kill -9 the daemon.
2. launchd brings it back within about 10 s.
3. pgrep -fl "user-data-dir=.*browsers" shows nothing left over.