The protocol is a specification, and several independent programs implement it. That redundancy is deliberate, and its value depends entirely on how the validator set is distributed across them.
A bug in a dominant client becomes a network bug
If one implementation contains a flaw that makes it accept an invalid block, every validator running it will follow the same wrong chain.
Where that implementation runs most of the network, the wrong chain has the majority and correct clients are the ones that appear to be forked off.
Diversity converts a catastrophic failure into a partial one, because a minority following a bad chain simply loses the argument.
The thresholds are specific
Finality requires a supermajority of stake in agreement, so a client running more than a third of validators can prevent finalisation by failing.
Above two thirds, the affected group can finalise a chain by itself, which turns a bug into a chain split with real stake behind both sides.
Below a third, a failing client causes its own operators to miss duties and incur penalties while the network continues finalising without them.
Correlated failure is penalised deliberately
Penalties scale with how much stake commits the same fault simultaneously, so a bug affecting many validators at once is punished far more severely than an isolated one.
This is intentional: it makes running a majority client financially riskier for the operator, not only for the network.
The incentive is aimed at large staking operations, whose choices move the distribution more than individual validators do.
Two layers each need diversity
Validators run an execution client and a consensus client, and these are separate programs from separate teams.
Diversity is required on both sides, since a dominant client on either layer creates the same concentration risk.
The two distributions have historically differed, and improvement on one layer does not address a concentration on the other.
Defaults determine the outcome
Most operators use whichever client their setup tool installs by default, so distribution is shaped by tooling rather than by deliberate choice.
Installers that select a minority client automatically, or that ask the question explicitly, have moved the numbers more than advocacy has.
Switching also carries operational risk, since a migration performed carelessly can produce exactly the double-signing fault the operator is trying to avoid.