Ollama batch generation

A batch of hundreds of local-model calls fails in ways a single call never shows: a timeout at item 180, a reply with prose after the JSON, a model that reloads every few requests, copy that invents facts. Build the job so any item can fail, the run can stop at any point, and nothing unvalidated reaches the output.

Shape of the job

records -> prompt per record -> cache lookup -> model call (retry) -> parse -> validate -> write output + manifest

• One record per request. Batching several records into one prompt makes one bad item fail the rest and makes validation ambiguous.
• Outputs and the manifest are written serially, in source order, even if requests run in parallel. Parallel writes to one file corrupt it.
• Inputs are never written. Generated files go to a separate output folder.