Zero, empty or unset: lnd's fee fields, Cosmos fee grants and Geth's recovery IDs
Three pieces of infrastructure code changed in the last two days for what is, at bottom, the same reason. Each could not distinguish a value that was meant from one that was missing, empty or out of range.
In lnd, the value was zero. The UpdateChannelPolicy call uses protobuf, and proto3 does not transmit zero values, so a request setting a channel's base fee to zero looked exactly like a request that said nothing about the base fee. The workaround had been to require every field on every call. Pull request 11528, merged on 14 November, adds an explicit marker for each field an update intends to change, and rejects a non-zero value in a field that is not marked. The developers rejected the simpler-looking alternative, proto3's optional keyword, because it would have changed generated types and broken existing clients.
In the Cosmos SDK, the value was empty. Chains that upgraded from v0.53 or earlier could hold fee grant expiry entries stored with an empty value, which the keeper cannot decode. Nothing happens until such a grant expires; then the x/feegrant end-of-block logic returns an error. Releases v0.55.1 and v0.54.5 on 14 November add a migration that rewrites those entries and retract the versions without it, with a deadline set by each chain's earliest affected grant.
In Geth, the value was out of range. A signature's recovery ID should be 0 to 3, but the pure-Go recovery path built without cgo accepted 4 to 7, because the library it used reserves those codes for compressed keys. The default cgo build rejected them. Pull request 36132, merged on 13 November, adds the missing check, after the team found that Clique block seals and peer-to-peer handshakes recovered keys from signatures one build accepted and the other refused.
In each case the failure was quiet until a specific condition arrived: an operator setting one fee, a grant reaching its expiry, a mixed network seeing an unusual signature.

Why it matters
Blockchain software multiplies the cost of these gaps because every node must agree. A value one implementation treats as zero and another as absent, or as valid and invalid, becomes a consensus question. The fixes share a method worth copying: make the intent explicit, check the range at the boundary, and test every value the type can hold, as the Geth test does for recovery IDs from 4 to 255.