[Eljees]

A minimal 100 Continue response must be forwarded before the client sends its body. Read the HTTP status prefix before inspecting the currently buffered response, then keep the adaptive peek of up to 56 buffered bytes. A partial first read containing only H must not cause the interim response to be treated as the final response.

The existing final-response test now reads its input one byte at a time. On the preceding implementation it returns status 100 instead of 200; with the minimum-prefix read it forwards Continue and returns the final response. The separate fixture removed during review remains removed.

Validation of the latest change:

• Windows Go 1.27.3 and Linux Go 1.27.1: all tests in ./proxy/http ./proxy/socks ./common/protocol/http (11 leaf tests), plus seven HTTP scenarios selected by ^Test(Http|HTTP) in ./testing/scenarios, pass. Linux runs both scopes with -race; the upstream scenario helper builds ordinary child Xray binaries.
• Four actual CLI cases against a synthetic loopback origin pass, including a status prefix split across writes. The preceding code loses the final response in that case, while the three controls pass. The test-suite regression uses a deterministic OneByteReader without timing.
• Vet for the three packages, main builds, repository format and git diff --check pass.
• The preceding full Windows CI run failed in TestSudokuE2ETemp/hysteria2/prefer_ascii with a Sudoku decode error. The HTTP tests passed in that run; this change does not claim to repair the Sudoku failure.

AI-assisted implementation and tests, separately reviewed and checked against the reproduced behavior.

[RPRX]

@Fangliding 先执行 r.Peek(1) 是有啥说法吗

[Fangliding]

先读一次再取出缓存中字节 应该不会有这个header都要拆开发的请求
改peek12就12吧 随便了