TSA-140: [BE] Keep structured output in sync with corrected text

[Yehor-Makarenko]

Closes #140

Description

Fixes an issue where human corrections in the verification UI were not propagated to the structured representation of a page. Previously, fixing plain text left edited_structured as null, allowing the original misread entities from structured to leak into the lexicon and poison the context for subsequent pages.

Key Changes

• Re-derivation on Correction: When a page is verified with action: "correct", edited_structured is now re-derived from the corrected text via a secondary text-only model call using the preset's output_schema.
• Schema Validation & Fallback: Re-derived JSON is validated against the preset's schema. Presets without an output_schema bypass re-derivation gracefully.
• Fallback Resolution (edited ?? original): Updated queries and data readers across the codebase to prioritize edited fields via COALESCE / fallback logic (edited_text ?? text and edited_structured ?? structured).
• Budget Tracking: Model calls for re-derivation are charged directly against the document budget.
• Immutability: Preserved original text and structured outputs to maintain auditability of the model's initial reading.

---

Acceptance Criteria

• Correcting a surname in the text puts the corrected form into the lexicon, not the original
• The next pages receive the corrected form in their context
• A page confirmed unchanged behaves exactly as before — no extra call, no extra cost
• The re-derivation cost is charged against the document budget like any other model call
• A preset without an output schema still verifies and exports normally
• text and structured are never overwritten — the model's original reading stays available

---

Notes

• Initially, I planned to implement the lexicon update directly within this scope, but upon investigation, it turned out to be covered by task [BE] Update the lexicon when a page is verified #296. This PR focuses on providing the re-derived edited_structured payload.