fix(qwen3): wait for async overlap prefill while idle

docs/models/qwen3/green-ctx-sm-partition.md

@@ -63,11 +63,13 @@ Power climbs with QPS: at QPS 8 the board still has headroom (draw oscillates ~5
 
 ## Pitfalls
 
-- **The Xid 31/43 hit during bring-up was a cross-stream buffer use-after-free; the supported success and checked-`Result` paths are fixed — not an open driver risk.** Prefill temporaries are allocated on `ctx.stream` but consumed by override-stream kernels; the fix is a producer fence (`green_ctx::fence_producers_before_override`, ordering each `ctx.stream` alloc/H2D ahead of its override consumer) plus RAII stream drains that hold each stream until its kernels finish before any buffer frees or KV page is released — `PrefillTempBin` (prefill) and `DecodeStreamGuard` (decode), each fail-closed (a sync failure aborts). An arbitrary panic that unwinds past a not-yet-parked prefill buffer is out of scope.
+- **The Xid 31/43 hit during bring-up was a cross-stream buffer use-after-free; the supported success and checked-`Result` paths are fixed — not an open driver risk.** Prefill temporaries are allocated on `ctx.stream` but consumed by override-stream kernels; the fix is a producer fence (`green_ctx::fence_producers_before_override`, ordering each `ctx.stream` alloc/H2D ahead of its override consumer) plus fail-closed RAII ownership. `LaunchedStreamGuard` drains launched prefill/decode streams before any pre-handoff return can recycle KV, `DeferredDrop` parks GPU-backed prefill temporaries until prefill-stream quiescence is proven, and `AsyncPrefillEvent` owns the launched prefill across the worker-to-executor handoff. An arbitrary panic that unwinds past an unarmed launch guard is out of scope.
 - **The "`cuGreenCtxStreamCreate` itself faults on driver 590 + >16 GB resident" theory did not reproduce.** This 5090 (driver 590.48.01, ~27 GB resident — 19.3 GB of it KV) ran green-ctx clean across both 2026-06-19 sweeps and the 2026-06-20 three-mode smoke (split path fired, 0 Xid).
 
 - **`gemm_lt` is still disabled under the stream override** (`5af4fd5`, to avoid the cuBLASLt workspace Xid-31 path). So split-path decode keeps its CUDA graph but loses the per-shape Lt tuning — a remaining decode-side lever, not yet re-measured under the partition.
 - **A single request never exercises the split path** — smoke-test with concurrent load or you are only testing the full-SM graph.
+- **Record async completion with the retained Green Context handle.** A Green Context stream does not reliably identify its owner through `cuStreamGetGreenCtx` when the derived context was never made current. The split command therefore carries `gctx_prefill` from `OverlapStreams` and records its blocking, timing-disabled event with `cuGreenCtxRecordEvent`; the shared-primary-context path still uses `cuEventRecord`. This keeps idle scheduler waits blocking and prevents an invalid or unrecorded event from bypassing prefill completion.
+- **GPU launch ownership starts before a fallible forward returns.** `SplitConcurrent` arms stream-quiescence guards before both prefill and decode launches because either call may submit kernels before reporting an error. GPU-backed prefill temporaries move to a thread-local deferred-drop queue on both success and error, including errors raised inside the forward; only a successful prefill-stream synchronization drains that queue. The prefill guard remains armed through decode synchronization, decode-result construction, and completion-event setup; it disarms only after `LocalQwen3Lane::inflight_prefill` owns the launched work. The completion event then remains an armed RAII owner across worker aggregation and executor-side decode application, so any pre-handoff return synchronizes before the scheduler can recycle KV pages. A synchronization failure is fail-stop. Once query/synchronize has already proven completion, event-destroy failure is logged but intentionally non-fatal: it may leak a driver resource, but cannot race KV or temporary-buffer reuse.
 - **`pkill` from an ssh one-liner matches its own command line** — use `pkill -f "[t]arget/release/openinfer"`, and kill/launch in separate ssh invocations.
 - Build on the 5090 with `CUDA_HOME=/usr/local/cuda-13.1` (stale `/usr/local/cuda` → cuBLAS 12.9 N=1025 cliff; see `serving-perf-5090.md`). Verify `ldd target/release/openinfer | grep cublas` shows `.so.13`.