fix(cortex): flip observability age check from oldest to newest write

[mcardinali-data]

Fixes #1915.

Problem

CortexHealth.ts (new in v7.40.4) warns when the oldest file under MEMORY/OBSERVABILITY/ exceeds 30 days. Nothing in the tree prunes or rotates these logs, so on any install older than a month the warning is permanently on and cannot be cleared — a health signal with no remedy, which trains operators to stop reading the health line (same class as #1910).

It also measures the wrong property. Observability holds append-only telemetry plus files that are legitimately stable; an old file is not a defect. The two real failure modes are:

• unbounded growth — already covered directly by the existing 256 MiB bytes check, untouched here
• a dead telemetry pipeline — nothing writes anymore; this is what an age signal can actually detect

Fix

Flip the polarity: track the newest mtime instead of the oldest, and warn as observability-stale when nothing has written within the threshold. That warning is actionable (check hook wiring) and self-clearing on every healthy install.

• observabilityEvidence(): oldestMs/Math.min → newestMs/Math.max
• finding id observability-age-exceeded → observability-stale, message names the likely cause
• CORTEX_OBSERVABILITY_MAX_AGE_MS env knob keeps its name and shape
• bytes check unchanged

This is remedy option 3 from the issue (scope the measurement to what active logging proves), implemented as the minimal 6-line diff rather than shipping a rotation mechanism and a retention policy every install would have to adopt.

Verification

Synthetic fixtures against assessCortexEvidence:

• dir with a 80-day-old file plus a fresh file → no observability finding (previously: permanent warn)
• dir where the newest file is >30 days old → observability-stale
• 300 MiB evidence → observability-bytes-exceeded still fires

Live install (logs written minutes ago, some files months old): overall ok, zero observability findings.