A shard chain block producer can equivalently selectively withhold data until close to the boundary (would this be the finalization time, or a different parameter?),
There is no “boundary” in the eth2 context. There’s an argument that if one node sees a block as available through the data availability proof mechanism, then with very high probability >=50% of the block is out there, so “soon” (ie. within 2 * delta) an honest node will see those chunks, reconstruct the full data and rebroadcast the full data, so all future nodes will see their checks pass. This mirrors the status quo (ie. eth1, bitcoin, etc), where if an honest node sees a block it can rebroadcast it, guaranteeing that other honest nodes will see the block within time delta.
Data availability checking isn’t a “do it once, pass or fail” operation, it’s more like “choose some indices, keep looking for the data corresponding to those indices, and accept the block once you get all of the responses, even if you don’t get them at first”.
I suppose you could argue that kind of “voting” happens implicitly, in that if a node delays publication of data and too many nodes see a block as unavailable at first then that block will not be accepted into the canonical chain, even if the data is published later.