AnyTLS inbound + Hysteria2 outbound: severe and erratic throughput degradation (83-808 Mbps swing), not CPU-bound

Environment

• sing-box version: 1.13.16 (official .deb, tags: with_quic, with_utls, with_acme, ...)
• OS: Debian 13 (trixie), kernel 6.12.43+deb13-amd64
• Hardware: KVM VPS, 1 vCPU / 1 GB RAM
• sysctl tuning: BBR + fq, tcp_wmem max 16 MB, tcp_mtu_probing=1, tcp_slow_start_after_idle=0

Configuration (minimal repro)

Server A (US West, 1 vCPU):

{
  "inbounds": [
    {
      "type": "anytls",
      "tag": "anytls-2",
      "listen": "::",
      "listen_port": 8444,
      "users": [{ "name": "u", "password": "***" }],
      "tls": { "enabled": true, "acme": { "domain": ["example.com"], "email": "a@b.c", "provider": "letsencrypt", "disable_tls_alpn_challenge": true } }
    }
  ],
  "outbounds": [
    {
      "type": "hysteria2",
      "tag": "hy2-landing",
      "server": "landing.example.com",
      "server_port": 34639,
      "password": "***",
      "tls": { "enabled": true, "server_name": "landing.example.com" }
    },
    { "type": "direct", "tag": "direct" }
  ],
  "route": {
    "rules": [{ "inbound": "anytls-2", "action": "route", "outbound": "hy2-landing" }],
    "final": "direct"
  }
}

Server B (landing, US West, same metro area as A) runs a hysteria2 server.

Problem

Throughput through the chain client -> AnyTLS inbound -> Hysteria2 outbound -> landing swings wildly between ~83 Mbps and ~808 Mbps across consecutive iperf3 runs minutes apart, while neither CPU nor the network link is saturated.

Controlled experiments (all single-stream iperf3, 10 s each, measured end-to-end through the full chain back to an iperf3 server):

Expected behavior

Chain throughput should be close to the slowest single layer (~800 Mbps) and stable across runs, as it is when the same inbound is paired with a direct outbound, or when the same hy2 outbound is fed from a local client.

Actual behavior

Throughput collapses unpredictably (down to ~10% of capacity) with plenty of CPU headroom. Suspected interaction between AnyTLS's padding-generated bursty traffic and the QUIC (BBR/Brutal) pacing in the hysteria2 outbound - the padded stream arrives in bursts, which may break the pacing/QPacing assumptions of the QUIC stack.

Note: replacing the hy2 outbound with a TCP-based chained outbound (VLESS+REALITY to the same landing) yields a stable ~800 Mbps through the identical AnyTLS inbound, so the issue is specific to the AnyTLS -> Hysteria2 combination.

Logs

Nothing relevant at warn level (no errors); happy to re-run with debug/trace logging and pprof if you tell me what you need.

Start by reproducing the minimal AnyTLS inbound to Hysteria2 outbound chain with consecutive single-stream iperf3 runs, then compare it with the direct and Hysteria2-only controls described in the issue. Use the requested debug or trace logging and pprof if needed; done means explaining the throughput swings and obtaining stable performance close to the individual components.