[QTI] Backfill all non-raw AssessmentItems to QTI items

❌ This issue is not open for contribution. Visit Contributing guidelines to learn about the contributing process and how to find suitable issues.

Overview

Run the one-time global backfill converting every legacy AssessmentItem to type='QTI' + raw_data XML via #6003's conversion, then remove the API-layer dual-read.

Complexity: Medium
Target branch: unstable

Context

• #6003 provides the reusable conversion.
• #6030 wires it into the endpoint as a transitional API dual-read.
• This issue makes that change permanent in the data.
• A data-only backfill — no schema change — over a large table.

The Change

• Backfill every legacy row to type='QTI' + raw_data XML via the #6003 conversion.
• Leave perseus_question rows untouched.
• Remove the API-layer dual-read (#6030) once the backfill completes.

Acceptance Criteria

• After the backfill, every legacy row is type='QTI' with raw_data XML; only perseus_question rows remain unconverted
• Rows that fail conversion are surfaced, not silently dropped
• The backfill is resumable and safe to re-run
• With the backfill complete, the API serves QTI directly without the dual-read
• Tests cover the backfill over a mixed fixture

AI usage

Architecture decided with the maintainer across an iterative session: blanket QTI type with the item XML in raw_data; XSD-authoritative validation across all sources; a legacy→QTI global migration with an API-layer dual-read; ricecooker upload delegating to the AssessmentItem serializer; and a Perseus custom-interaction contract confirmed against the QTI 3.0 specification. Claude mapped the existing publish/validation/ricecooker code, proposed the breakdown, and drafted each issue; the maintainer steered every decision and reviewed throughout.

Start by reading the reusable conversion in #6003 and the transitional API dual-read in #6030. Design the global backfill to convert non-Perseus legacy rows, surface failures, and support safe reruns; completion means mixed-fixture tests pass and the API no longer needs the dual-read.