Latest version frequently crashes on remote 256MB VPS with large auth directory

[wsyqhh]

Summary

After upgrading to the latest CLIProxyAPI version, the service frequently crashes on a remote 256MB low-memory VPS (not the reporter's local machine), where it used to run fine before.

Environment

• CLIProxyAPI version: latest available version at the time of report
• OS/Arch: Linux arm64
• Deployment type: remote single-binary service
• Memory limit: 256MB RAM VPS
• Auth storage: file-based auth directory
• Auth file scale reported by user: 4000+ auth files

What changed

Older versions were stable enough in the same kind of low-memory environment.
After upgrading to the latest version, the remote instance now crashes repeatedly.

Question

Has there been any recent change in:

• auth file indexing/loading,
• startup scanning of the auth directory,
• in-memory caching strategy,
• watcher behavior,
• or token/auth metadata retention

that could explain worse memory behavior on large auth-file sets?

Expected behavior

Even with a large auth directory, the service should avoid loading too much into memory at once and should remain stable on low-memory deployments that were previously working.

Additional note

If there is already a recommended config or migration path for large auth directories on small-memory machines, guidance would be appreciated.