[tabcat]

Description

floodsub picked a random 64-bit sequence number for each message and encoded it with leading zero bytes removed, so about 1 in 256 messages went out with a sequence number shorter than 8 bytes. gossipsub peers using fallbackToFloodsub reject any sequence number that is not exactly 8 bytes, so those messages were dropped silently. This is also why the gossipsub fallbacks to floodsub publish tests time out intermittently.

The pubsub spec requires a linearly increasing 64-bit big-endian sequence number. floodsub now uses a counter seeded from the wall clock in nanoseconds, the same as gossipsub, go-libp2p and rust-libp2p. The seed is above 2^56, so the sequence number always encodes to 8 bytes, and it keeps increasing across restarts for validators that drop non-increasing sequence numbers.

Notes & open questions

• Messages already sent by older floodsub versions keep their old encoding. Encoding and signature verification are unchanged, so nothing new is rejected.
• randomSeqno is renamed to nextSeqno. It is internal (not in the package exports).

Change checklist

• I have performed a self-review of my own code
• I have made corresponding changes to the documentation if necessary (this includes comments as well)
• I have added tests that prove my fix is effective or that my feature works