That is not reasonable.
The git tool uses large digests (sha-1) to identify all commits. This is a 160-bit hash and apparently you'd need 2^80 random messages to get a 50% chance of a collision.
I recall reading that git hasn't any safeguards or double-checks to see if two commits hash to the same value. So Linus was just trusting the maths and making a calculated risk. For his purposes, at least, the chance of collision was so vanishingly small that's it's a risk that he was willing to take (for the sake of simplifying the code, one assumes).
I get what you're saying about the pigeonhole principle, it's not always unreasonable to assume that collisions won't happen.