Phase 2 + 3 pause-window: NVMe + io_uring + pre-emptive background snapshot

[WaylandYang]

Item 6 of meta #112. 3 weeks; technical depth.

Phase 2: NVMe + io_uring snapshot writer (1 week)

Targets the underlying memory.bin write that still bottlenecks total BRANCH API latency on SSD even with Diff mode.

• Replace blocking std::fs::copy in branch_sandbox with an io_uring-driven async copy (ring-buffered, configurable depth)
• Daemon flag --snapshot-writer={blocking,io_uring}, default blocking
• Measure on NVMe + SSD; expect ~10× on SSD's full-copy path
• Land RESULTS-v0.3.2.md or extend RESULTS-v0.3.md

Phase 3: Pre-emptive background snapshot (1-2 weeks)

Bounds pause regardless of source size, including for non-first BRANCHes.

• Background thread flushes dirty pages on a 1 s tick to a staging memory.bin
• At BRANCH time, only flush the dirty-since-last-tick window
• Pause becomes O(tick) instead of O(source memory)
• Measure; expect bounded pause ≤ 200 ms across all dirty footprints

Anomaly to profile

• BRANCH 3-5 pause jumps from 280 ms to 1.3-1.5 s in the multi-BRANCH sweep without diff size growing. Hypothesis: KVM control-plane accumulating state. See RESULTS-v0.3.md.

Acceptance

New RESULTS section in bench/pause-window/ with phase 2 + 3 numbers. Daemon flag(s) shipped. Anomaly root-caused or escalated upstream.