FastMCP: Support dynamic annotation updates

Problem

FastMCP will support resource annotations (audience, priority, lastModified) via the @mcp.resource() decorator (see #1468). However, annotations are set once at decoration time and cannot be updated as the server runs.

This is problematic for lastModified especially. If you annotate a file resource, there's no easy way to keep the timestamp current as the file changes:

@mcp.resource("file://docs/readme.md", annotations=Annotations(
    audience=["user"],
    lastModified="2025-01-12T15:00:58Z"
))
def read_readme() -> str:
    return open("readme.md").read()

Once deployed, this annotation is stale. If the file updates tomorrow, clients see outdated metadata.

Solution

Allow annotations to be updated dynamically. Possible approaches:

• Provide a method to update resource annotations at runtime
• Allow annotation callables/lambdas that compute values on each list_resources() call
• Return fresh annotations from resource handler functions

This would enable:

• Accurate lastModified timestamps from file mtimes
• Dynamic audience/priority based on server state
• Better integration with external data sources

Issue written by Claude, reviewed by @maxisbey

Commencez par le chemin d’enregistrement FastMCP de @mcp.resource() et le comportement de list_resources() mentionné dans l’issue. Comparez les approches proposées de mise à jour à l’exécution, de callable et de retour du handler, puis définissez comment une implémentation maintiendrait à jour les annotations audience, priority et lastModified pour les clients.