Context

The ProviderHealthTracker (added in #773) uses an append-only in-memory list for ProviderHealthRecord entries. The get_summary() and get_all_summaries() methods filter to a 24-hour window, but old records are never evicted from the backing list.

This mirrors the existing CostTracker pattern, which has the same unbounded growth characteristic.

Problem

Over extended runtime (days/weeks), the _records list grows without bound. For a health tracker recording every provider call, this could accumulate millions of records and cause:

• Gradual memory growth
• Increasingly slow _snapshot() calls (copies the entire list)
• Eventually OOM

Proposed Solution

Add periodic eviction of records older than _HEALTH_WINDOW_HOURS during _snapshot() when the list exceeds a configurable threshold (e.g., 100,000 records). Apply the same pattern to CostTracker as well.

Found By

Pre-PR review agent (silent-failure-hunter) during #773 review.