Keep the tunnel loss estimate moving when the tunnel stalls

[TaJirax]

What changed

• retransmits now close the loss window on their own once there is a window's worth of them
• a window is still a window of fresh data in the normal case, so the measured ratio is unchanged for healthy traffic

Why

tunnelLossMeter closed its window only on fresh STREAM_DATA. A tunnel losing badly does the opposite of that: new data slows to a trickle while STREAM_RESEND piles up, so the window never completed and lastPerMille stayed frozen at whatever the last healthy period measured — usually zero.

adaptiveDuplicationCount derives the copy count from that estimate. With the floor at one copy, which is what the speed preset configures, duplication could not rise during exactly the loss it exists to answer: the tunnel sat at a single copy per packet while retransmitting nearly everything.

Validation

• new test drives a healthy window, then a stall of mostly retransmits, and asserts the estimate moves and that duplicationForLoss lifts the count off its floor
• the existing TestTunnelLossMeter_UploadRetransmitRate still passes unchanged, which is the check that the healthy-traffic ratio was not redefined
• go build ./..., go vet ./..., and go test ./... pass

🤖 Generated with Claude Code