…tion.

…tion to ensure the widest possible loss detection scan window.

Loss detection code is concurrent to the application threads that are consuming from the image but also to the receiver thread that inserts new packets into the image and advances the high-water mark position counter. Reading that counter last ensures that we take into account all inserted packets up to this point.

… never zero. Also do not track padding bytes as the "sent bytes".

…-water mark on the receiver.

A single network packet might contain multiple frames (up to an MTU). When such packet contains only `HDR_TYPE_DATA` then packet length will correspond to the combined lengths of the frames it contains. However, if there is a trailing `HDR_TYPE_PAD` packet then the packet length will not be equal to the sum of the frame sizes.

For example, given a packet containing two 80 byte messages (whose aligned length is 128 bytes each) and a padding for 224 bytes. The length of the packet will be 288 bytes (256 + 32) but the "frame size" will be 512 bytes (256 + 256).

…ENGTH` and streamlining non-zero frame lookup.

…ame of data.

…will advance on next heartbeat/data frame anyway.

…fit a frame of data."

This reverts commit 9b8229c.

…s almost never zero. Also do not track padding bytes as the "sent bytes"."

This reverts commit 411f102.

…ild position to ensure the widest possible loss detection scan window."

This reverts commit 60e6047.

…TH` and streamlining non-zero frame lookup.

…e when heartbeats advance the high-water mark position, or it can decrease if messages arrive out of order. In both cases we want to adjust the NAK range of the receiver.