feat: implement generated redactions across all modalities

[martsokha]

Summary

Implement the Generate redaction strategy — replacing sensitive values with realistically generated alternatives using LLM or rule-based generation.

Scope

• Text: generate realistic replacement names, addresses, phone numbers, etc. matching entity kind and format
• Image: generate replacement face/object images (e.g., synthetic faces via diffusion model)
• Audio: generate replacement speech segments (e.g., TTS with altered voice)

Context

TextStrategy::Generate was removed in the redaction pipeline rework since it produced only placeholder text ([GEN:PersonName]). This issue tracks implementing actual generation across all modalities.

Tasks

• Define GeneratedRedaction type or extend existing strategies
• Wire LLM-based text generation into RedactionApplicator
• Evaluate image generation approaches (synthetic faces, inpainting)
• Evaluate audio generation approaches (TTS re-synthesis)
• Add configuration for generation providers and parameters

🤖 Generated with Claude Code