Docs: privacy & redaction of traces and tool calls

[bordeauxred]

Why

ClawLoop captures agent-environment interactions for learning. Anyone integrating with production data — internal LLM use, customer support, agentic workflows — needs to know:

• what's captured by default
• where it's stored (`LocalArchiveStore` writes to disk)
• what redaction hooks exist
• how to scope what `clawloop.wrap()` collects
• what the `EpisodeCollector` retains in memory vs persists

Right now this information is scattered across docstrings or absent. AI engineers who read the code will infer answers; AI engineers who read only the README will assume the worst (or assume nothing and hit it later).

What

• Write `docs/privacy.md` covering:
  • what data is captured at each layer (Harness, Router, Weights, EpisodeCollector)
  • default storage location and retention
  • redaction patterns: how to drop fields before persistence, how to mask tool args, how to opt out of trace persistence entirely
  • PII boundaries: what we explicitly do not try to detect
  • integration with OpenTelemetry exporter — what reaches the OTLP endpoint
• Add a "Privacy & data handling" section to README with a one-paragraph summary + link.
• If redaction hooks don't exist where they should, file follow-up issues (don't mix doc and impl in one PR).

Acceptance

• `docs/privacy.md` exists and is linked from README
• A reader can answer "what does ClawLoop persist about my agent runs?" in under 60 seconds

Source

Plan §2 P2 #25.