fix(agent-server): bound accumulators + /health diagnostics (#333 hardening)

[dolho]

Summary

Speculative-but-defensible hardening for the long-uptime futex-spin
symptom reported in #333. Does not claim to fix the symptom (root
cause in #333 is still unproven, and verification needs a multi-day
soak). It locks in the remaining accumulator-leak surface reductions and
adds the observability hook that turns the next reproduction into a
one-curl diagnosis instead of an strace expedition.

What changed

• gemini_runtime.py: replace per-call ThreadPoolExecutor(max_workers=1)
  at execute() and execute_task() with a module-level singleton, mirroring
  claude_code.py:63. Per-call executors rely on CPython's weakref-callback
  cleanup of worker threads, which is not deterministic under load.
• state.AgentState.add_message: FIFO-trim conversation_history once
  it exceeds history_limit (default 1000, overridable via
  AGENT_HISTORY_LIMIT). Persistent history is in the backend DB; the
  in-memory list is only used by /api/chat/history + session-info counts.
• /health: add a diagnostics block exposing thread_count,
  asyncio_task_count, running_executions, conversation_history_size,
  and conversation_history_limit. A future repro can now be triaged with
  curl /health instead of strace -c -p.

What this does NOT do

• Prove the original symptom is gone. That requires a 2–6 day soak on a
  reproducing host.
• Identify the root cause of bug: agent-server.py develops futex spin loop after days of uptime #333. Adjacent subprocess teardown fixes
  shipped since the report (bug: drain_reader_threads closes stdout pipe before reader can drain backlog — silently loses final result line on long agentic tasks #531, bug(agent-runtime): npx MCP server subprocesses outside claude pgid hold stdout pipe open — reader stuck after #531 grace period #618, bug: drain_reader_threads post-kill timeout not enforced; leaked thread blocks capacity slot indefinitely #649, bug: drain_reader_threads t.join(30) deadlocks asyncio event loop permanently when reader thread is stuck #657, bug: agent-server.py spins at 90% CPU on OAuth token auth failure, blocking CB recovery #728/fix(agent-server): cap drain executor thread at 90 s to fix CPU spin on auth failure (#728) #730) may have
  already addressed the most likely culprits — but that remains
  unverified.

#333 stays open until reproduced on current main.

Test plan

• tests/unit/test_agent_server_hardening.py — 13 cases covering
  history bound, gemini executor singleton, /health diagnostics.
• Multi-day soak with the new /health diagnostics block scraped at
  regular intervals — looking for thread_count, asyncio_task_count,
  or running_executions trending up over time.

Relates to #333.

🤖 Generated with Claude Code

[vybe]

Clean, focused hardening. Three targeted accumulator-leak reductions with good test coverage. Executor singleton correctly mirrors claude_code.py:63; history FIFO trim is well-guarded; diagnostics block is graceful under sync/async contexts. Leaving #333 open is the right call — root cause still needs a multi-day soak to verify.