fix(health): make the verdict mean what the card says it means

[iyashjayesh]

Closes #84. Two defects in the same place — the second found while fixing the first.

1. The badge contradicted its own explanation

Healthy came from Percent > 50 (a composite score) while IconMsg came from a threshold comparison. Nothing kept them in step, so one payload produced one card saying both things at once:

SYSTEM HEALTH   Degraded
host cpu        27.79% / 95.00%
host memory     83.88% / 95.00%

System usage is within limits: CPU Usage 18.06% / 95.00%, Memory Usage 81.25% / 95.00%

Every number inside its limit, the sentence saying so, and the badge saying the opposite.

It also flapped. A binary taken from a continuously drifting score oscillates near its cutoff. Sampled every 3s on an idle machine:

55.66%  healthy=True   -> Healthy
48.91%  healthy=False  -> Degraded
44.93%  healthy=False  -> Degraded
51.39%  healthy=True   -> Healthy

Two crossings in 45 seconds, while the bars beside it did not move.

The verdict now comes from the same check that writes the message, set at the same point so they cannot drift. Percent is untouched — it still drives the ring and the graded status text. It just no longer answers a binary it was never designed for, which removes the flapping without hysteresis, because thresholds do not sit near the boundary.

2. An idle resource masked a saturated one

This surfaced in the first test run. The breach test ran on the average of the usage ratios:

"within limits: CPU Usage 11.69% / 95.00%, Memory Usage 104.21% / 95.00%"

Memory at 104% of its allowance — over the limit — reported as within limits, because (11.69 + 104.21) / 2 = 57.95.

Left alone, fix #1 would have made the badge faithfully report a wrong verdict. A breach is now any single resource over its own limit. Averaging stays where it belongs, in the graded score.

Both regressions are pinned

Restoring Percent > 50:

--- FAIL: TestTheVerdictAgreesWithItsOwnExplanation
    system reports Healthy=false while its own message says the usage is within limits.
      message: System usage is within limits: CPU Usage 11.69% / 95.00%, Memory Usage 104.21% / 95.00%

Restoring the averaged breach test:

--- FAIL: TestOneBlownResourceIsABreach
    system: a fixture that exceeds a limit reports Breached=false
    system: message claims the usage is within limits while a resource is over it

Behaviour change worth knowing

Services that were quietly reporting healthy while one resource was over its limit will now report unhealthy. That is the point, but it is a change in what existing deployments display.

Not in this PR

While reading the same file I found the service-health CPU normalisation looks wrong — (cpuUsage / totalCores) * 100 on a gopsutil value that is already a percentage of one core, which inflates by 100/TotalCores (10× on a 10-core machine, 25× on a 4-core one). Cross-checked against load_statistics.service_cpu_load: 1.68% reported there, 17.66% in the health message for the same quantity. It changes every user's numbers, so it wants its own issue and its own PR rather than riding along here.

Verification

go test -race ./... green across all 12 packages, go vet clean.