feat(observability): evaluation pipeline and validation/risk analytics

**Context:** ROADMAP.md → Evaluation pipeline + Validation and risk analytics
**Related:** #412, #415, #237, pr-outcome-tracking draft

---

## Component

API or orchestration

## Describe the feature

Backend **evaluation pipeline**: failure categorization, memory effectiveness metrics (merge rate, revision cycles, CI pass rate), and dashboards for PR risk labels, validation outcomes, and trends by repo, user, and `prompt_version`.

## Use case

Operators adopting background agents need outcome-oriented metrics, not task counts. Product and platform teams need analytics to tune prompts, memory, and validation tiers.

## Proposed solution

1. Aggregate from Tasks table, `TaskEvents`, PR outcome tracking, and validation tier outputs.
2. CloudWatch metrics + operator dashboard widgets (extend existing dashboard construct).
3. Optional `GET /analytics/...` or export API (align with #237 governance planes).
4. Eventually feed learned memory rules into Tier 2 validation when tiered pipeline ships.
5. Document metric definitions in `docs/design/EVALUATION.md`.

## Other information

- #412 covers CLI `bgagent stats`; this issue covers **backend aggregation and dashboards**.
- Design context: `docs/design/EVALUATION.md`, `docs/design/OBSERVABILITY.md`.

- [ ] This might be a breaking change

Start by reading ROADMAP.md and the design context in docs/design/EVALUATION.md and docs/design/OBSERVABILITY.md, then inspect the existing dashboard construct and the Tasks, TaskEvents, PR outcome, and validation-tier data sources. Done means the agreed backend evaluation aggregates and CloudWatch/operator dashboard coverage are implemented, with metric definitions documented in docs/design/EVALUATION.md.

Written by the indexing model from the issue text.