fix(agent-server): make drain_reader_threads async to prevent event-loop deadlock (#657)

[vybe]

Summary

• Converts drain_reader_threads in subprocess_pgroup.py from sync to async def
• Replaces all t.join(timeout=N) calls with await asyncio.wait_for(asyncio.to_thread(t.join, N), timeout=N+1) so deadlines are enforced at the event-loop level even under CPU starvation from orphan subprocesses
• Replaces the orphan-killer _orphan_thread.join(timeout=10) with a daemon thread + threading.Event: event.wait(10) always returns within 10s, so asyncio.run()'s shutdown_default_executor() never blocks on a slow /proc scan
• Updates 4 sync call sites in claude_code.py to use asyncio.run(drain_reader_threads(...))
• Updates all 5 test call sites in test_subprocess_pgroup.py to use asyncio.run()

Changes

• docker/base-image/agent_server/utils/subprocess_pgroup.py — drain_reader_threads → async def, asyncio-based joins
• docker/base-image/agent_server/services/claude_code.py — 4 call sites wrapped in asyncio.run()
• tests/unit/test_subprocess_pgroup.py — 5 call sites wrapped in asyncio.run()
• docs/memory/feature-flows/parallel-headless-execution.md — revision history entry

Test Plan

• All 12 test_subprocess_pgroup tests pass (3 skipped: Linux-only /proc tests)
• TestDrainOrphanKillerTimeout.test_drain_bounded_when_orphan_killer_blocks completes in ~14s instead of 100s — regression for deadlock
• test_buffered_data_preserved_after_grandchild_kill passes — data integrity regression (bug: drain_reader_threads closes stdout pipe before reader can drain backlog — silently loses final result line on long agentic tasks #531)
• test_unwinds_reader_stuck_on_grandchild_pipe passes — end-to-end drain still works

Fixes #657

Generated with Claude Code