Problem

Scheduled automation (setup, automation, and config-driven reconcile) only
works on macOS. The composition root hard-codes LaunchdSchedulerAdapter, and
the automation domain model leaks launchd vocabulary (plist_path). On Linux,
setup/automation fail because there is no scheduler adapter and launchctl
is absent.

What this does

Adds Linux support behind the existing SchedulerAdapter port, using systemd
--user timers — the closest analogue to launchd LaunchAgents (per-user,
persistent, and rich enough to fill the existing ScheduledJobStatus model).

• Select the scheduler by platform. default_scheduler_adapter() returns
  the systemd adapter on Linux and the launchd adapter elsewhere, wired into the
  composition root (app.py). The explicit adapters.scheduler injection path
  is unchanged, so tests still inject fakes.
• De-launchd the domain model. plist_path becomes manifest_path, and the
  manifest location is computed per platform: a launchd plist under
  ~/Library/LaunchAgents on macOS, a systemd .timer under the user unit
  directory on Linux (honoring XDG_CONFIG_HOME when it is an absolute path).
  The macOS-only PATH fallbacks no longer leak onto Linux.
• SystemdSchedulerAdapter. Writes a .service + .timer unit pair,
  enables and (re)starts the timer through systemctl --user, and reads status
  back from systemctl show. OnActiveSec=0 mirrors launchd RunAtLoad;
  OnUnitActiveSec carries the interval. is_loaded gates on ActiveState, so
  a stopped or disabled timer is not reported as loaded. The systemctl runner
  is injectable so the adapter is unit-testable without touching the real user
  manager.
• Shared helper. The process-error helper moves to scheduler/process.py so
  neither adapter depends on the other.
• Provider-neutral automation status output and updated docs (README support
  matrix, automation architecture notes).

Rejected alternative: cron

cron was considered and rejected. crontab exposes no per-job status, run
count, exit code, or PID, so ScheduledJobStatus would be half-empty and behave
differently from macOS. systemd show maps cleanly onto the existing model.

Behavior / contract change

automation status (human output and --json) now reports manifest_path
instead of the launchd-specific plist_path; the human labels are manifest
and scheduler loaded. This is a deliberate, provider-neutral rename.

Two systemd choices are intentional and documented in the plan:

• Persistent=true is omitted — per systemd.timer(5) it only affects
  OnCalendar= timers, so it would be inert config on a monotonic timer.
• run_count is not reported on Linux — NRestarts counts Restart= restarts,
  always 0 for a oneshot service, so reporting it would be misleading. macOS
  keeps launchd's runs.

Testing

• uv run pytest — 477 passed. uv run ruff check . — clean. uv build — ok.
• Real end-to-end run on Ubuntu (systemd user manager running): enabling an
  automation task through the CLI installed a real timer; automation status
  reported it installed, scheduler loaded: yes, with the correct interval
  and state; the scheduled run fired and exited 0; is_loaded correctly
  returned false after the timer was stopped; uninstall removed both unit files
  and left no residual units in the user manager.