feat(memory): anomaly detection on memory writes

**Context:** ROADMAP.md → Memory security → Anomaly detection

---

## Component

Agent (Python runtime)

## Describe the feature

**CloudWatch metrics** on memory write patterns; alarms for burst writes, unusual content size, or suspicious keyword spikes.

## Use case

Memory poisoning attacks may manifest as rapid bulk writes or anomalous content. Operators need early warning before retrieval spreads bad data.

## Proposed solution

1. EMF metrics: write rate, bytes, entries per task/repo.
2. Alarms with sensible defaults + Blueprint overrides.
3. Optional auto-quarantine hook (see quarantine draft).
4. Dashboard panel on operator dashboard.

## Other information

- Design context: `docs/design/MEMORY.md`, `docs/design/OBSERVABILITY.md`.

- [ ] This might be a breaking change

Start by reading docs/design/MEMORY.md and docs/design/OBSERVABILITY.md, then inspect the Agent Python runtime and operator dashboard entry points. Define how EMF metrics, default and Blueprint-overridden alarms, the optional quarantine hook, and the dashboard panel fit together. Done means the proposed write-pattern signals and operator controls are implemented and documented.

Written by the indexing model from the issue text.