Apply the preparation contract to bulk and mailbox scoring

[pavsoss]

Summary

Bulk scoring and mailbox scanning each built their own model input and skipped text preparation entirely, so a classification depended on which entry point a message arrived through. The same text could be labelled one way via the prediction endpoint and another way via a file upload or an inbox scan.

Part of #1037.

Changes

Both paths now prepare text the same way the prediction path does, so identical input produces identical results regardless of entry point. Bulk results still echo each row exactly as uploaded, so a caller can match a result back to its input line even though a canonicalised form was scored.

Test plan

• Recording doubles capture what actually reaches the vectorizer, confirming bulk rows and scanned messages are scored in prepared form.
• Obfuscated and plain spellings of the same message are shown to reduce to one scored string.
• Bulk responses are checked to echo the original row text.

Depends on

Stacked on #1038. Its commit is included here, so this diff shrinks to just the bulk and mailbox changes once #1038 merges.

[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.