Opt-in inline async scheduler (CG_ASYNC_INLINE): skip per-node asyncio.create_task

[erez-work]

Summary

The async runner currently wraps every node downstream of any async node in asyncio.create_task, even when that node's inputs resolve synchronously. Measured at ~6.3µs/node of pure task machinery on trivial nodes — which matches the ~7.3µs/node "plumbing floor" we see on large voice agents (the floor is asyncio scheduling, not node bodies).

This adds an opt-in runner behind the env flag CG_ASYNC_INLINE=1 (off by default → zero behavior change for every existing caller) that:

• runs a node inline (a plain call, no task) when its dependencies are already resolved values;
• spends a task only on nodes that genuinely depend on a still-pending async result;
• uses eager tasks so a short-circuiting async node completes immediately and its downstream collapses to inline execution.

Only the async path is affected; the sync path is untouched.

Semantics preserved

• make_first/make_or priority option-fallback on handled raises (the deferred path mirrors node_to_input_async)
• side-effect / _profile hooks, and the sync isawaitable guard
• concurrency of independent in-flight async ops (eager tasks start them when reached, gathered at the end)
• the unhandled-exception cyclic-ref cleanup dance

Validation

• Full test suite: 69/69 with the flag on and off (Python 3.14).
• Real bon_secours_voice agent (63,912 nodes), Python 3.12, baseline vs flag-on:
  • greeting: graph-compute 870ms → 610ms (−30%), byte-identical response
  • schedule (routing turn): byte-identical response + events (ASKED_DOB_QUESTION/FLOW_STARTED/SKILL_STARTED), no crash
  • graph-compute = depth-tracked outermost-reducer wall (isolates scheduler from VIC/LLM network)

Requires Python 3.12+

Eager asyncio tasks are native only on 3.12+. The runner raises a clear error if the flag is set on an older interpreter; the default (flag off) path is unaffected on any version.

🤖 Generated with Claude Code