Assessment: Batch Execution & Response Refinements #1200

Is your feature request related to a problem? Please describe.

• BATCH assessments finish, cost provider spend, and leave the client with nothing.
• Callback inlines every row, gets 413, the failure is swallowed, and there is no poll endpoint.
• Envelope error and metadata are hardcoded null, so a failed run says only FAILED.
• Provider dumps and OpenAI's error file are never kept: one run reported errors=0 against 389 failures.
• Legacy cron crashes on API-created runs, so they never finalise or fire a callback.
• Gemini runs with a json_output_schema fail 100%; Anthropic effort is silently dropped.
• Submission rows (3-6MB) sit inline in Postgres, in a table shared with evaluations.

Describe the solution you'd like

• Persist every dump plus an errors.jsonl on the assessment, and deliver presigned URLs and the failure reason in the callback.
• Give assessment its own submission table and a per-assessment storage prefix.
• Stop the cron touching API runs; fix the Gemini and Anthropic mappers.