Back off and reject ambiguous samples

Double the timeout after expiration; do not feed ambiguous retransmitted-segment timings into the estimator.

• On timeout, retransmit the earliest unacknowledged segment (§5.4). The host MUST double RTO (§5.5), subject to a permitted maximum, then restart for that doubled value (§5.6). It does not resend every outstanding segment merely because this timer expired.
• The first example expires at 2400 ms: resend A, change RTO 2400→4800 ms and restart at 2400 with deadline 7200 ms. The second separate expiration for B doubles 4800→9600 ms. No maximum is chosen in these examples.
• TCP MUST use Karn’s algorithm: RTT samples MUST NOT come from retransmitted segments because the reply can refer to the original or a later instance (§3). The TCP timestamp option is the stated exception when it removes that ambiguity.
• An ACK for A arriving at 2600 ms after sends at 0 and 2400 could suggest 2600 ms or 200 ms. The data can be acknowledged, but do not admit either ambiguous value as a new RTT sample. Acknowledgment processing and sampling are distinct.
• Once new data is sent and acknowledged with an unambiguous RTT sample, recompute using §2. This can collapse a backed-off RTO. The final state retains the old estimator 800/400 ms and uses a clean 800 ms sample, producing SRTT 800/RTTVAR 300/RTO 2000 ms; it does not blindly reset to 1000 ms.
• An implementation MAY clear SRTT/RTTVAR after multiple backoffs (§5). If it clears them, the next valid sample uses first-sample initialization, not the subsequent-sample update. The drawn final case explicitly retains them. Congestion control after a timeout remains RFC 5681’s separate responsibility.