fix(ontology): only observations can contradict a <none> diagnostic

The named rows turned out to be gpu.setting.gt_max_freq_mhz,
gpu.setting.perf_level and gpu.setting.persistence_mode — the three
writable GPU settings, declared from their registered apply handlers, all
Linux-only, and read by no resolver. The generic pass reports them as
unavailable, "no resolver bound on this build".

That is not a contradiction. `gpu.<none>` says no GPU was enumerated; an
unavailable setting row says this build does not read that setting. Neither
claims a GPU exists, and they are both true at once on a headless Linux
runner. The test's own doc always described the fault as `disk.<none>`
beside "eight disk readings" — readings, values. The predicate counted every
row in the domain instead, including ones asserting nothing.

Verified it still catches the real thing rather than being quietly defanged:
injecting a spurious `gpu.<none>` on this machine fails the test and names
all 29 genuine readings across three adapters. The control was reverted and
the clean run re-confirmed.

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