Track down a memory leak in a service that slowly falls over

You are a senior engineer who has chased memory leaks in production. My service's memory climbs until it restarts. Using the evidence below, help me find the cause in five sections:

1. PATTERN READ - what the memory graph shape suggests (steady leak, step jumps, load-linked growth, cache with no cap). Say which evidence supports it.
2. SUSPECTS - table: Code location | Why it could retain memory | Evidence for | Evidence against. Max 5, ranked.
3. CONFIRMING TEST - for the top two suspects, the cheapest experiment or heap snapshot comparison that proves or kills each.
4. FIX - the code change for the most likely cause, as a diff.
5. GUARDRAIL - one metric or alert so this is caught before the next restart.

Rules: no preamble. Do not suggest "add more memory" or "restart on a schedule" as the fix. If the evidence cannot separate two suspects, say so instead of picking one.

RUNTIME + VERSION: {{e.g. Node 20}}
MEMORY OVER TIME + RESTART TIMES: {{paste or describe}}
WHAT CHANGED BEFORE IT STARTED: {{deploys, traffic, config}}
SUSPECT CODE: {{paste caches, listeners, pools, globals}}