feat(orchestration): PR risk classifier at submission

**Context:** ROADMAP.md → Agent quality → PR risk classification
**Related:** #425 (review routing depends on this), tiered validation draft

---

## Component

API or orchestration

## Describe the feature

Rule-based **PR risk classifier** evaluated at submission (and optionally updated post-agent). Outputs structured risk label(s) that drive model selection, budget defaults, Cedar approval requirements, and human review tier (#425).

## Use case

Not all tasks deserve the same model, budget, or review depth. A docs-only typo fix should not inherit the same gates as a payment-service schema migration.

## Proposed solution

1. Classifier inputs: workflow ref, repo metadata, touched paths (when known), task description signals, Tier 3 validation output (when available).
2. Output schema: `risk_tier: low | medium | high | critical`, `signals: string[]`, `recommended_model`, `recommended_max_budget_usd`.
3. Persist on task record; expose in `bgagent status` and API.
4. Blueprint overrides per repo for path-based rules.
5. Start advisory-only; enable policy gates via Cedar in follow-up.

## Other information

- #425 implements **routing**; this issue implements the **classifier**.
- Design context: `docs/design/EVALUATION.md`, `docs/design/CEDAR_HITL_GATES.md`.

- [ ] This might be a breaking change

Commencez par docs/design/EVALUATION.md et docs/design/CEDAR_HITL_GATES.md, puis examinez #425 pour comprendre comment le routage dépend du classifier. Définissez le classifier en mode consultatif uniquement autour des entrées et du schéma de sortie indiqués, persistez son résultat dans l’enregistrement de la tâche et exposez-le via bgagent status et l’API.

Rédigé par le modèle d'indexation à partir du texte de l'issue.