feat(pi): schedule the Pi credential renewal so it runs unattended

[ruby-dlee]

Stacked on #251. Review that one first; this PR's diff is only the schedule.

Why

A renewal an operator has to remember is not a renewal. #251 makes renewal possible; this makes it happen on its own, which is the actual requirement.

Machine-global, and identified by what it runs

One credential pool at ~/.pi/agent/auth.json serves every Firstmate home, and this machine has nine. So there is one label, and every fact the report depends on is read from the one plist at that label, never from the checkout that is asking.

The first version of this compared the installed job against the reporting checkout's own path. That would have let exactly one home report healthy while the other eight printed PI_AUTH_REFRESH: unavailable at every session start forever, and the remedy that line prints would have booted out the working job and repointed it at whichever home ran it. An adversarial review reproduced the thrash across two homes.

The nonce is what makes the liveness claim true

install-scheduler mints an activation nonce, bakes it into the plist environment, and record_heartbeat writes nothing unless the process carries the matching one. A nonce reaches a process only through launchd's copy of the job environment.

So a hand-typed run-once --scheduled writes nothing at all: it can neither forge proof of life nor destroy the real one. The earlier version claimed this property in four places and did not have it. bin/fm-report-retention.sh holds itself to the same rule the same way.

Proved against the real launchd: launchd's own run stamped a matching nonce and scheduler-status reported installed; a hand run afterwards left the heartbeat byte-identical.

States

absent, orphaned, foreign, unloaded, unproven, stale, failing, attention, installed.

• launchd is asked before the disk, so a job loaded with nothing describing it is orphaned rather than reading as absent.
• unproven ages against its own plist mtime: loaded six months ago and never fired is stale, not waiting for a first fire.
• attention is a working schedule whose profiles need a browser. ensure passes on it, because calling a working schedule unavailable is the alarm that gets ignored.

Install and uninstall are transactional now

The plist used to be published before the bootstrap, so a refused bootstrap left the new definition on disk with the old job running, and every report read the file. It now restores the previous definition and rebootstraps it. It also binds the loaded job rather than the file it just wrote, because launchctl can report success while launchd holds nothing.

uninstall-scheduler re-probes after its bootout and refuses to remove the plist while launchd still holds the label, rather than reporting a removal that did not happen and leaving a credential refresher running with nothing describing it.

Platform

The macOS-only refusal goes through FM_PI_REFRESH_PLATFORM, the seam bin/fm-checkout-refresh.sh uses, because CI runs this suite on Linux and the test has to exercise a macOS-only install there. Without it this file went red the moment the stack retargeted main.

Mutations

Seven that an adversarial review found surviving, or that this change introduced, checked by making the edit and running the suite:

mutation	result
uninstall-scheduler drops its bootout	red
install-scheduler ignores a refused bootstrap	red
install-scheduler drops the post-install verification	red
install-scheduler drops the --interval-seconds bounds	red
bin/fm-bootstrap.sh drops the install pi-auth-refresh target	red
record_heartbeat stamps regardless of nonce	red
the installed job is compared against this checkout's own path	red

The third needed a new case to kill: the fake launchctl always registered on a successful bootstrap, so checking the return code and checking the loaded state were the same assertion. It now models launchctl accepting a bootstrap that launchd does not end up holding.

Also

The suite exports its own state root at file scope. Only one unit set it, so the other units' isolation rested on the very property one of them asserts, and a mutation wrote a fixture heartbeat into the operator's real state root.

docs/configuration.md gets its own heading rather than sitting inside the checkout-refresh section, where the paragraphs that follow it described a different job's guarantees.

Follow-up, not in this PR

Installing it on this machine for real waits for the stack to land, so the plist points at ~/firstmate/bin/fm-pi-refresh.py rather than at a worktree that disappears.