Caches provider file uploads so the same bytes are not uploaded twice.

Uploading is slow and, on some providers, billed. When the same asset is attached across several turns of a conversation - or across conversations - the second and later sends should reuse the first upload's ExAgent.FileRef.

Entries are keyed by {scope, sha256(bytes)}, where the scope digests the provider module, base URL, and API key. Scoping by API key matters: two provider structs for different accounts must never share a file_id, because one account's file reference is invalid - or worse, readable - for the other. The API key itself is never stored, only its digest.

Expiry

fetch/2 treats an entry whose ExAgent.FileRef has expired as a miss and evicts it, so the caller uploads afresh.

Growth

There is no eviction beyond expiry. Entries are small (a FileRef struct and two digests), but an application uploading a large number of distinct files will grow the table without bound, since OpenAI file references never expire. Call clear/0 if that matters for your workload.

The table is public and named so that concurrent agents read it directly rather than serializing behind this process; the GenServer exists only to own the table across crashes.