Add requires_review flag to detect incomplete LLM extraction

[Arijit429]

Closes #450
Closes #40

Overview

This PR enhances the reliability and safety of the FireForm extraction pipeline by introducing a validation layer that detects incomplete or low-confidence LLM extraction results.

In the current workflow, form submissions proceed even when the language model fails to extract critical fields (e.g., returning placeholder values like "-1" or missing entries). This change introduces a structured mechanism to flag such cases for human review.

Key Changes

Extraction Validation Utility

• Added a reusable validation helper (src/utils/validation.py) to evaluate extracted JSON data against required template fields.
• The validator checks for:
  • Missing keys
  • Empty string values
  • Placeholder extraction outputs such as "-1"

requires_review Flag Integration

• Extended the extraction pipeline to compute a requires_review boolean after LLM-based extraction.
• Updated FileManipulator.fill_form() to return both:
  • Generated PDF path
  • Review requirement status

Controller and API Flow Updates

• Propagated the validation flag through the Controller layer.
• Updated /forms/fill route to persist the review state alongside form submission data.

Database Model and API Contract Extension

• Added requires_review column to the FormSubmission model.
• Extended FormFillResponse schema to expose review status to API consumers.

Motivation

LLM-driven extraction is inherently probabilistic and may produce incomplete outputs in real-world emergency scenarios. By surfacing a review indicator, this change supports safer operational workflows and enables future human-in-the-loop validation interfaces.

Impact

• Non-breaking enhancement to submission pipeline.
• Improves transparency and reliability of automated form generation.
• Establishes groundwork for advanced validation, confidence scoring, and UI review flows.

Future Work

• Field-level confidence metrics
• Structured extraction schema enforcement
• Frontend review dashboards for flagged submissions
• Retry or fallback extraction strategies