thermal_zone: Skip zones with per-zone read errors

[neoLsH]

What this PR does

Makes the thermal_zone collector respect procfs' per-zone ReadErrors, so a thermal zone whose sysfs attributes could not be read is skipped instead of being exported as a bogus 0°C.

Background

procfs.ClassThermalZoneStats returns every globbed thermal_zone<N> and records per-zone sysfs read failures in the ReadErrors field (an errors.Join of the type/policy/temp/mode/passive read errors). This is the outcome of prometheus/procfs#712 ("skip failed zones") and the later refactor in prometheus/procfs#794.

The collector currently ignores ReadErrors and unconditionally emits float64(stats.Temp)/1000.0. When a zone's temp read fails (offline zone, transient driver error, permission issue, etc.), procfs leaves Temp at 0, so node_exporter exports node_thermal_zone_temp 0 — a silently wrong reading that can trigger false "low temperature" alerts.

The fix

for _, stats := range thermalZones {
    if stats.ReadErrors != nil {
        c.logger.Debug("Skipping thermal zone with read errors", ...)
        continue
    }
    ch <- ...Temp...
}

Only zones whose attributes were read successfully are exported.

Interaction with #3657

#3657 added syscall.EINVAL to the whole-collector error path. This PR is complementary and does not change that behavior: procfs deliberately does not record EINVAL from a temp read in ReadErrors (see the !errors.Is(err, os.ErrInvalid) guard in parseClassThermalZone), so an EINVAL zone still has ReadErrors == nil and continues to be reported with Temp == 0, exactly as #3657 intended. This PR only affects zones with genuine read failures that procfs does flag in ReadErrors.

Testing

• GOOS=linux go build ./... && go vet ./collector && go test -c ./collector are clean.
• The existing e2e fixture (thermal_zone0, a healthy zone with ReadErrors == nil) is unaffected: it still exports node_thermal_zone_temp{...} 12.376, so the e2e golden output does not change.
• Note: thermal_zone has no unit test today and procfs/sysfs is Linux-only, so a macOS-local runtime test isn't possible; verification relies on Linux CI (build + e2e). I kept the change minimal and did not add a failed-zone fixture to the shared fixtures/sys tree to avoid perturbing the e2e golden output. Happy to add a dedicated Linux unit test using a temp-dir sysfs fixture if reviewers prefer.

Signed-off-by: neoLsH 43921685+neoLsH@users.noreply.github.com

[neoLsH]

@SuperQ friendly ping — small fix so thermal zones with sysfs read errors get skipped instead of exported as a bogus 0°C. Complements #3657 (EINVAL zones are unaffected). CI is green.