2 Commits

Author	SHA1	Message	Date
root	4074e68715	experimental/windivert: P2.1+P2.2 with WinDivert NETWORK+SOCKET layers ... WIP snapshot before pivot to sing-box+TUN. Reached: - TCP redirect via streamdump pattern (swap+Outbound=0+reinject) - SOCKET layer for SYN-stage flow detection (avoids FLOW Establish-too-late race) - Lazy PID→name resolution (catches Update.exe inside procscan tick) - UDP forward via SOCKS5 UDP ASSOCIATE relay + manual reinject - Result: chat works, voice times out (Discord IP discovery / RTC handshake fails) Reason for pivot: WinDivert NAT-reinject pattern has subtle layer-3 semantics issues that DLL-injection / TUN-based proxies sidestep entirely. Going with embedded sing-box + wintun as the engine — proven path for Discord voice through SOCKS5. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>	2026-05-01 22:27:54 +03:00
root	dd402d4fc4	internal/redirect: TCP NAT-loopback redirector ... Listener on 127.0.0.1 accepts NAT-rewritten Discord SYNs (rewrite done by divert layer in Task 10), looks up the original destination in a sync-protected map keyed by source port, opens a SOCKS5 CONNECT to the upstream proxy targeting that destination, and pumps bytes both directions until either side closes. 30-minute TTL sweeper handles T-6 in the edge case matrix (mapping leak when a flow never properly closes). Pump teardown: when one direction's io.Copy exits, the goroutine CloseWrite's its write side AND sets a past read deadline on the OTHER conn so the peer goroutine's blocked read unwinds promptly even when the upstream half never sends EOF (test fake-SOCKS5 hits this; the real upstream may too). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>	2026-05-01 20:00:21 +03:00