Assessment BATCH: Clean Up Submission Config #1154

Problem

For BATCH assessments, the per-row prompt template lives on the API request (BatchInput.query), separate from the input_schema it interpolates (buried in config_blob.assessment.params). Consequences:

• A typo'd {column} in the template is caught only at run time — the literal {placeholder} leaks into the prompt sent to the provider, silently corrupting the assessment. No early feedback to the config author.
• Every caller resends the full template on each request, though it belongs to the config.
• input_schema sits inside the assessment provider params even though it describes the shared input rows (it's already popped before the provider call — proof it's not a provider param) and pre-filters have to reach cross-block to validate against it.
• The duplicate_detection pre-filter in the API-client config adds confusion and duplicates a legacy-RUN concept.

Proposed change (BATCH only)

1. Move the prompt template into the config as submission — mandatory on assessment.params, optional on each pre-filter's params. Holds the {column} template.
2. Hoist input_schema to the config_blob top level (flat sibling of pre_filters/assessment). One data contract, read by both consumers. json_output_schema stays assessment-specific.
3. Validate at config-save: every {placeholder} in every submission must resolve against the top-level input_schema keys; an unknown column rejects the save. Runtime only validates the structured data rows against the schema.
4. Remove query from the request (BatchInput); server interpolates submission with each row's values.
5. Remove duplicate_detection from the API-client pipeline (legacy RUN pipeline untouched).

RESPONSE is out of scope — no proper wiring today; its query removal is cosmetic.

Target config shape

{
  "config_blob": {
    "input_schema": {
      "submission_id": { "type": "text" },
      "level": { "type": "text" },
      "artwork_image": { "type": "image", "format": "url" }
    },
    "pre_filters": {
      "topic_relevance": { "provider": "openai", "params": { "instructions": "...", "submission": "grade {submission_id} at {level}" } }
    },
    "assessment": {
      "provider": "openai",
      "params": { "instructions": "...", "submission": "assess {submission_id} {level} {artwork_image}", "json_output_schema": { } }
    }
  }
}

Acceptance criteria

• submission mandatory on assessment, optional on pre-filter.
• input_schema is a mandatory, non-empty top-level config_blob field.
• Config save rejects a submission with an undeclared {column} (assessment and pre-filter).
• query rejected on BatchInput (extra=forbid); prompt sourced from config submission.
• Runtime validates data rows against the top-level input_schema.
• duplicate_detection gone from API config/pipeline/results; legacy RUN unaffected.

Notes / impact

• Breaking: clients sending query now 422; ASSESSMENT configs now require top-level input_schema + submission.
• No DB migration (config_blob is JSONB); existing configs re-validate only on next save.
• Frontend (kaapi-frontend) config CRUD + submit form need a follow-up.

Implemented in #1153.