Ironbullet v0.6.2-rc.3

This release candidate fixes Xray adapter lifecycle and job-state races reported in #73 and adds a deterministic config-speed regression harness.

Fixed

Single-flight startup for identical VMess, VLESS, and Trojan proxy URIs.
Parallel startup remains available for different encrypted proxy URIs.
Application shutdown now fences in-flight Xray initialization, rejects late resolutions, terminates every managed Xray child, and removes generated configs.
Xray cleanup runs on CLI completion, confirmed GUI close, and event-loop destruction.
Failed Xray startup removes temporary configuration files.
CLI completion no longer waits forever on a hit channel whose sender was still owned by the runner.
Stopped jobs are no longer overwritten as Completed when asynchronous work drains.
Per-run generations prevent stale hit, stats, and completion tasks from mutating or detaching a restarted job.
A distinct Stopping state prevents restart, edit, or removal until old workers drain; a cancellation latch also prevents stop-before-start from being undone.
Paused Save & Resume now waits for the update acknowledgement and resumes the existing runner; stopped/completed jobs are labeled and handled as restarts.
Running/paused progress no longer displays 100% before final completion, and stale deleted-job selections are cleared.

Windows x64 native build with the pinned NASM, LLVM, and Xray runtime:

90 Rust library tests passed.
Binary, config fixture, proxy parsing, and Shadowsocks integration tests passed.
Frontend production build passed.
Packaged Xray smoke test observed an Xray child, then verified IronBullet exit code 0, zero leaked Xray processes, and zero leaked generated configs.
Direct loopback benchmark: 1,000 processed, 1,000 successful HTTP-200 outcomes, 0 fails/errors/missing/duplicates, 331.6 successful requests/second, 100 peak concurrent requests.
Local forwarding-proxy benchmark: 1,000 processed through the proxy, 1,000 successful HTTP-200 outcomes, 0 fails/errors/missing/duplicates, 326.1 successful requests/second, 100 peak concurrent proxy requests.
A separate ten-candidate public HTTP-proxy sample found two reachable and eight failed/timed-out proxies, confirming public proxy volatility is not a valid deterministic performance baseline.

ironbullet-v0.6.2-rc.3-windows-x64.zip

SHA-256: 85356f67e314a9b0818b6c46e96d2f3fbe40c2b54daba7d19cd24a5109aa8655

The ZIP contains ironbullet.exe, reqflow-sidecar.exe, the pinned xray.exe, XRAY-LICENSE.txt, and release-manifest.sha256.