refactor: establish reusable runtime lifecycle

[domcyrus]

Runtime lifecycle foundation for #602 and #606, related to #9.

• Prepares privileged capture before sandboxed workers start, requires successful UID/GID dropping, retains worker ownership through shutdown, and finalizes outputs after producers stop.
• Reuses packet-worker buffers and reduces work inside ordered commits without changing capture order or the 10,000-packet queue bound. Matched-capacity benchmarks improve selected TCP/HTTP workloads; DNS results are mixed and overload can still drop packets.
• Retains idle-capture regressions, secure output ownership, and cross-platform workspace CI. Existing Windows API features cover file security; the unused regex dependency is removed. Adds panic/waiter and exact PCAPNG worker regressions.

[domcyrus]

Updated in 0045b9e, preserving ordering and the existing queue bound.

• Linux verification: 1,058 default and 1,080 all-feature workspace tests passed, plus clippy and formatting. New tests cover waiter/panic recovery and exact PCAPNG bytes, order, and attribution.
• Five interleaved matched-capacity trials improved four-worker TCP/HTTP throughput by 9.1%/15.4%. DNS was mixed, including a 6.7% lower four-worker median with overlapping ranges.
• Live queue drops decreased in all four before/after pairs but remain nonzero. All eight independent PCAPs contained every generated datagram exactly once. TUI capture-drop fields were unavailable, so residuals remain unclassified. These virtualized results do not establish universal capacity or recovery of every cost relative to main.

[domcyrus]

Updated in 883bcd0. Follow-up cb1ffc3 fixes a new Rust 1.98 Clippy lint in a test fixture; 93c038b adds the privileged readiness regression to CI under its own watchdog. Runtime code and benchmark results are unchanged. Local tests and clippy passed again, and CI is rerunning.

• Restore Linux capture readiness wakeups, streamline tracker lookups, group ordered commits and avoid single-worker DPI staging. Retain the 10,000-packet queue with a bounded 5 ms full-queue wait; add 23 regression tests and update all three language guides.
• On one Linux ARM64 VM, three interleaved 5-million-packet repetitions versus main 747f265d show TCP/4-worker throughput +92.8%, HTTP/4-worker +79.5%, and UDP/50,000-flow -2.5%. Single-worker DNS remains -2.0% at the median with overlapping ranges. The original 1-million-packet HTTP repeat now shows +6.7% versus main. Native queue sizes differ; these are synthetic-workload results, not universal guarantees.
• Final live stress trials: complete displayed coverage at 678k to 699k packets/s, and 99.962% to 100% at 50,000 flows. All six final-candidate runs had zero reported capture-backend drops; two retained 100 and 300 queue drops. Independent PCAPs were validated packet-by-packet. All-feature tests (1,102), three privileged checks, formatting, clippy and dependency audit passed locally. Fresh CI is running.