[gdevenyi]

Stacked on #64. Fixes #65.

Bug: the disk PLE table's wait-sync mode captures a WAIT on _flag in every decode graph, but the host set the flag only after cuGraphLaunch returned. A large graph that does not fit the launch queue makes cuGraphLaunch block until the GPU drains, and the GPU was parked on that WAIT, so it deadlocked. The hybrid MoE decode graph hit this on its first step with --moe-collect-stats, and intermittently without it. #65 has the evidence (py-spy stacks, flag watcher, nsys).

Fix (6ebfa18): all PleStore calls run FIFO on one ple-filler thread, so the store stays single-threaded.

• A graph fill is submitted before the launch, with the request contexts snapshotted on the engine thread.
• Eager fills are submitted and awaited, as before.
• A failed launch cancels its fill and clears a signal that no WAIT consumed.

Test: tests/models/qwen4_exp/test_ple_disk_fill_order.py fakes a launch that blocks until the flag is set.

• Two tests fail on the old code ("deadlock: the fill only runs after the launch returns").
• All 4 pass with the fix.

GPU validation (RTX 5080, Qwen3.8-Flash-Next-NVFP4, disk PLE):

• The deterministic repro (hybrid + --moe-collect-stats) used to hang on the first request. It now completes 3/3 300-token requests (8.5–8.8 s each).
• Hybrid without stats: quality 14/15 (same as baseline); needle 32K and 128K 3/3; decode 37–42 tok/s from 2K to 128K; a 255K context fits with peak VRAM 15672 MiB.
• Offload (production): a back-to-back A/B against the pre-fix build shows no decode regression (38.8 vs 38.95 tok/s).

Hybrid gives no decode gain over offload on this box (~38.2 vs ~38.8 tok/s), so production stays on offload. The fix still matters because offload uses the same wait-sync path.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Bu6LgoxLR4wETqb7RPR2vt