What

Adds a sensitivity classification layer to Distill's memory store. When context is stored or retrieved, Distill classifies it and returns sensitivity metadata alongside the content.

Why

Distill sits at the right point in the pipeline to detect sensitive content before it reaches the model. When an agent retrieves memory and then calls an external tool, there's no signal about what the context contains. This PR adds that signal — callers can use max_sensitivity to make authorization decisions before dispatching tool calls.

Changes

New package: pkg/sensitivity

• Level enum: None (0), PII (1), InternalIP (2), Credentials (3)
• Pattern-based Classifier with built-in detection for:
  • Credentials: AWS keys (AKIA...), OpenAI keys (sk-...), GitHub tokens (ghp_...), Slack tokens (xox...), generic password=/secret= patterns
  • PII: email addresses, phone numbers, credit card numbers, SSNs
  • Internal: configurable domain suffixes (.internal, .corp, .local)
• Classify(text) and ClassifyBatch(texts) methods
• All classification is synchronous, no LLM calls

Memory integration

• StoreEntry.Sensitivity — explicit tag at write time
• StoreEntry.AutoClassify — triggers pattern-based classification on write
• RecallResult.MaxSensitivity — highest level across all returned memories
• RecallResult.SensitiveChunks — which memories triggered it
• RecalledMemory.Sensitivity — per-entry sensitivity level
• Sensitivity metadata does not affect deduplication or retrieval ranking

Tests

• 17 classifier unit tests (patterns, false positives, batch, string representation)
• 4 benchmarks — all well under 1ms:
  • Short text: ~10µs
  • Text with matches: ~14µs
  • Long text (2000 chars): ~413µs
  • Batch of 10: ~191µs
• 7 memory integration tests (explicit, auto-classify, no-match, override, multi-entry, ranking unaffected)

Usage

# Store with explicit sensitivity
curl -X POST localhost:8080/v1/memory/store -d '{
  "entries": [{
    "text": "Q3 pricing: customer A at $120k",
    "sensitivity": 2
  }]
}'

# Store with auto-classification
curl -X POST localhost:8080/v1/memory/store -d '{
  "entries": [{
    "text": "API key: sk-proj-abc123...",
    "auto_classify": true
  }]
}'

# Recall — sensitivity metadata in response
curl -X POST localhost:8080/v1/memory/recall -d '{"query": "pricing"}'
# Response includes:
#   "max_sensitivity": 2,
#   "sensitive_chunks": [{"chunk_id": "...", "sensitivity": 2}]

Closes #82