feat: rebuild as enterprise multi-agent portfolio management platform

Clean-replace the three overlapping legacy HELIOS roots (helios/, backend/,
src/research_assistant/) with a single DDD / Clean / Hexagonal / event-driven
platform at src/portfolio_platform/. Advisory-only: every recommendation is
explainable, auditable, and passes a mandatory human-in-the-loop approval gate.

Layers (dependency direction points inward; enforced by import-linter):
- domains/        pure core — shared kernel + 8 bounded contexts
                  (users, portfolios, market_data, risk, research,
                  recommendations, compliance, execution). No framework imports.
- contracts/      frozen Pydantic v2 agent I/O contracts + workflow state
- agents/         9 single-responsibility agents on a BaseAgent harness
                  (retries, timeout, telemetry, audit); quant decisions are
                  deterministic, LLM smooths prose only (NullLLMClient default)
- workflows/      9-stage pipeline with human-approval gate, checkpoint,
                  rollback (LangGraph when present, deterministic fallback)
- application/    use-case services + DI container (composition root)
- infrastructure/ 16-table SQLAlchemy schema, in-memory + SQLAlchemy adapters,
                  Kafka/in-memory bus, Redis/in-memory cache, OpenAI/Null LLM,
                  JSON logging, Prometheus/OTel, JWT+RBAC/ABAC+PII+injection
- apps/           FastAPI api, event-projector worker, monitoring scheduler

Runs fully in-memory with no external services or API key. 67 tests pass across
unit/agent/contract/workflow/integration/api/security/performance suites.

Deployment: multi-stage Dockerfile, docker-compose (postgres/redis/kafka),
k8s (Deployment+HPA+CronJob), GitHub Actions CI (ruff, mypy, import-linter,
pytest 95% coverage gate). Docs: docs/ARCHITECTURE.md (24 sections), README,
scripts/demo.py. Original spec preserved at prompt/PLATFORM_SPEC.md.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
