feat: MIF v2 vendor-neutral schema with format adapters

[varun29ankuS]

Summary

Redesigns the Memory Interchange Format (MIF) as a vendor-neutral standard with pluggable adapters, fixing all known issues with the v1 implementation.

What was broken (v1)

• Entity types exported as "UNKNOWN" — graph lookup was missing
• Import destroyed UUIDs — generated new IDs, breaking all references
• O(n*k) dedup — recalled all memories per import candidate
• Merge strategies were no-ops — silently ignored
• Encryption was dead code — unused stubs bloating the module
• Export bulk-loaded all memories — OOM risk on large stores
• handlers/mif.rs was 1688 lines with search handlers mixed in

What's fixed (v2)

New module: src/mif/

• schema.rs — Vendor-neutral types: MifDocument, MifMemory, MifKnowledgeGraph, MifTodo, MifProject, MifReminder with vendor extensions
• export.rs — Streaming export with entity type resolution from GraphMemory, per-memory vendor extensions, PII redaction
• import.rs — UUID-preserving import via remember_with_id(), O(1) SHA256 content-hash dedup, graph reconstruction
• pii.rs — PII detection/redaction (email, phone, SSN, API key, credit card)

Adapter pattern: src/mif/adapters/

• MifAdapter trait with detect(), to_mif(), from_mif()
• shodh.rs — MIF v2 JSON (lossless round-trip) + v1 backward compat
• mem0.rs — mem0 format: [{memory, metadata, ...}]
• generic.rs — Generic JSON array: [{content, timestamp, tags, ...}]
• markdown.rs — Markdown with YAML frontmatter (Letta/Obsidian style)
• AdapterRegistry with auto-detection (most specific → generic fallback)

Handler cleanup

• handlers/mif.rs rewritten: 1688 → ~300 lines (thin HTTP layer)
• Multimodal/robotics search moved to handlers/search.rs
• Dead encryption code removed entirely

New API endpoint: GET /api/mif/adapters — lists available formats

Files changed

• 10 new files in src/mif/
• 5 modified files (handlers, lib.rs, memory/mod.rs)
• +2958 / -1413 lines net

Test plan

• cargo check — clean ✅
• cargo clippy — no warnings in new files ✅
• Round-trip: export MIF v2 → import → verify UUIDs, entity types preserved
• Import mem0 JSON → memories created with correct content/timestamps
• Import markdown → memory created with frontmatter metadata
• Existing multimodal/robotics search routes work after file split
• PII redaction works on export with redact_pii=true