Route bulk prediction through the versioned serving snapshot

[pavsoss]

Summary

Bulk prediction now scores an uploaded file against a single, coherent serving snapshot (model + vectorizer + label encoder) captured at the start of the request, so a concurrent model reload can never score part of a file with a mismatched set. Results carry the serving model version.

Behaviour

• One bad row can no longer fail the whole upload. Each row is validated and scored in isolation; rows that are empty, over the per-row length cap, or otherwise un-transformable come back in a structured skipped list with a stable typed reason, while every valid row still returns its prediction.
• Two env-configurable limits guard resource use: a total row cap (exceeding it returns a typed error through the standard error
  envelope) and a per-row length cap (over-length rows are skipped, not fatal).
• Responses are stamped with the serving model version — a model_version field on the JSON body and a matching column plus header on the CSV export — so consumers can tie predictions to the exact model that produced them.
• Missing serving objects now surface as a typed, retryable error instead of an opaque failure.

Existing response fields are preserved; the new fields are additive.

Test plan

Regression coverage for mixed valid/invalid uploads (valid rows returned with the skip list populated), row-cap enforcement as a typed error, and version stamping on both the JSON and CSV responses.

Part 1/2 #1021

[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.