fix(backend): the macOS CPU path invented almost everything it reported

`MonitoringBackend` is what the `simon` CLI runs. Windows and Linux call real
readers; the macOS arm hand-rolled from `sysctl` and the load average, and
almost every field it produced was fabricated:

- Per-core figures were the system-wide number repeated across every core --
  exactly what `cpu.core.{n}.utilization` declares it is "deliberately not", and
  what the resolver's own comment calls "one number repeated".
- A fixed 60/40 user/system split, invented outright.
- `governor: "performance"` on a platform with no governors.
- `min: freq_mhz / 2`. The handoff already records this defect from a previous
  round: a minimum is a distinct property, and dividing the maximum produced "a
  plausible number that was not one".
- `"Apple CPU"` when the brand string was missing. `platform::macos::read_cpu_stats`
  rejects exactly this in its own comment: a wrong identifier is "a wrong reading
  rather than an absence".
- A failed load-average parse fell to 0.0, publishing 100% idle -- the shape of
  the zero-constructor defects 6.0.0 removed, where no threshold could ever be
  crossed.

`update_memory` had no macOS arm at all, so `simon` reported no memory on a Mac.

Both real readers have existed since 5.2.0 and were wired only into the ontology
resolver. So on one Mac, `simon snapshot` and the CLI backend gave different CPU
figures -- one measured, one invented -- and nothing compared them.

**The structural lesson: wiring a new reader into one consumer leaves the old
fabrication running in the others, and no test notices.** The ontology gained a
macOS arm in 5.2.0 and this backend kept its invented one for two releases.

Verified: both cross-target checks with `--all-targets` (the code is
`cfg(target_os = "macos")`, so the local compiler never reads it), fmt, clippy
`-D warnings`, 826 lib tests.

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