Bring every prediction path under the hot-reload swap

[pavsoss]

Summary

A hot reload atomically swapped the text classifier, but two prediction paths sat outside it: the URL classifier was loaded once into module state, and mailbox scanning read model objects pinned on the application at startup. After a reload those paths kept serving whatever the process booted with, so a swap updated some paths and not others — the same class of defect as train-serve skew, arrived at from a different direction.

Closes #1037.

Changes

Both now travel in the versioned serving snapshot, so one reload moves every prediction path together and a request continues to see an internally consistent set of objects even if a reload lands mid-flight. Loaders that do not supply the URL pair keep working unchanged, so existing callers and lightweight test doubles are unaffected.

Test plan

• The URL pair is present in the initial snapshot and replaced by a reload, including the case where a loader omits it.
• A mailbox scan performed after a reload is shown to use the reloaded objects rather than the startup ones.
• A snapshot already handed to a reader is unaffected by a later reload.

Depends on

Stacked on #1038, #1039 and #1040. Their commits are included here, so this diff shrinks to just the reload-coverage changes once they merge.

[vercel]

@pavsoss is attempting to deploy a commit to the Aditya Sharma's projects Team on Vercel.

A member of the Team first needs to authorize it.

[Userunknown84]

@pavsoss merge conflict h fix it fastly