XHTTP transport: Reduce memory allocations ~88% (68MB→~8MB per 10MB traffic)

[rufsieus]

Summary

• Pre-allocate body read buffer in hub.go using known Content-Length instead of io.ReadAll grow-and-copy pattern — eliminates ~77% of allocations (1 alloc instead of ~7)
• Skip slices.Concat when payload comes from a single source (90%+ of cases) — zero-copy fast path
• Pre-size bytes.Buffer for HTTP/1.1 request serialization in client.go — eliminates buffer growth steps
• Sweep dead entries from globalDialerMap in dialer.go + new HasActiveClients() in mux.go — prevents unbounded map growth over long sessions

Context

Relates to #5344. XHTTP transport allocates ~7x more memory than data volume (10MB traffic → ~68MB allocations). On iOS, the system kills VPN extensions at >50MB. These 4 targeted fixes bring
allocations down to ~8MB with a minimal 39-line diff.

Test plan

• go build ./... passes
• go vet ./transport/internet/splithttp/... — no new warnings
• go test ./transport/internet/splithttp/... — all 13 tests pass

[Fangliding]

dup

[rufsieus]

Could you point to the existing PR? #5797 and #5801 were closed. This PR takes a different, minimal approach (39 lines vs larger diffs) focusing on pre-allocation with known Content-Length rather than pool-based rewriting.

[RPRX]

c1b67a9 加了 co-author