fix: real async embeddings for goal alignment; unblock event loop in planner

[haasonsaas]

Summary

• Kill the silent fake-embedding fallback. generate_embedding_sync could never call the async OpenAI embeddings API from inside a running event loop (i.e. always, under FastAPI), so it fell back to a deterministic SHA-1 hash pseudo-embedding. Goal alignment was computing cosine similarity over hash digests — every "related initiative" and Slack notification was based on garbage similarity, with no externally visible failure. The stub path and sync wrapper are deleted; embeddings are now generated via the async API end-to-end.
• Failure behavior: when embeddings can't be generated (API error, or dry-run mode with no client), goal alignment is skipped with a logged warning — never faked. generate_embedding returns None in dry-run mode and raises on API failure; the planner treats both as "skip alignment".
• Unblock the event loop. Runner.run_sync(...) → await Runner.run(...) in agent_sdk.py (verified against installed openai-agents 0.3.3: Runner.run is a coroutine function with the same kwargs). generate_plan / generate_plan_for_idea / _collect_related_goals / _notify_alignment are now async, awaited from the /plan endpoints and the procedure runner (which no longer needs asyncio.to_thread). The alignment Slack notification is awaited directly instead of the asyncio.run/create_task branching.

Test plan

• uv run ruff check . — pass
• uv run ruff format --check . — pass
• uv run mypy agent_pm — pass
• uv run pytest -q — 135 passed (was 134)
• Tests updated to mock the async APIs (_RUNNER.run, generate_embedding, async planner/critic fakes); dry-run embedding test now asserts None instead of a 1536-dim stub.
• New test test_goal_alignment_skipped_when_embeddings_unavailable: embedding failure → related_initiatives == [], no "Related Initiatives" section in the PRD, no notification dispatched, no goal_alignment trace event — proving no hash-fallback path remains.