feat(profile): make applied settings reversible; rewrite the handoff

`simon tune --apply` could change a machine and had nothing that could
change it back. `ApplyHandler` could only write, and `ApplyOutcome`
recorded only what was requested -- so the value being overwritten was
gone the moment it was overwritten. An autonomous tuner that moves a
machine in one direction only is not something to leave running.

Adds `ApplyHandler::read_current`, `ApplyOutcome.previous`,
`profile::apply::revert_setting` and `tuning::serve::revert_cycle`. A
revert goes back through `apply_setting`, so it is confirmed and
audit-logged on the same terms as the write it undoes -- an autonomous
loop with an unconfirmed way back would be a write path with no
confirmation.

Where no prior value was recorded, revert refuses rather than writing a
default. Putting a machine into a state it was never in is worse than
leaving it where the caller put it, and it is the same rule the tuner
already follows for values: never a number this crate made up.

`read_current` is implemented for the Windows active power scheme and
verified on hardware -- it returns the GUID that `profile explain` reports
through unrelated code in profile::cpu. The Linux cpufreq reader is
written by inspection and has never run; the code and HANDOFF.md both say
so.

`ApplyOutcome.previous` is `#[serde(default)]`, so audit records written
before this still deserialise; a test pins that.

HANDOFF.md is rewritten for 5.0.0: the 4.0.x rows are marked withdrawn,
the Dewey item records why the port was reversed and what would have to be
true to try again, the two Dewey bugs are kept as findings for that repo
rather than as live guidance here, and the new open item names what is
left to close the tuning loop -- and warns that a loop inventing its own
success criterion is the optimiser equivalent of a model picking a power
limit.

Also records the session's actual lesson: 839 tests stayed green through
four releases of a GUI that opened at 800x600, clipped every value off the
edge, responded to no clicks and reported no GPUs on a three-GPU machine.
Nothing failed because nothing was asked.

854 tests, 0 failures. clippy -D warnings clean.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>