MD-compliant hashes don't really accept arbitary length input, do they?

When people talk about hash functions, they usually say that they accept arbitrary-length input, but if you actually look at the padding (eg MD-strengthening padding), you see it's like M100.000||length where length is the binary representation in bytes or bits. But what if the length is so large that you can't represent it in a single block?

Obviously you can get around this problem by just using the most significant bits, but doesn't this destroy the MD-compliant padding property (which you need in proofs of one-wayness and collision resistance etc, assuming properties of the compression function)?

MD5 will literally take an arbitrary length; that's because the value placed in the padding is actually computed modulo $2^{64}$.
For SHA-1 and the SHA-2 hashes, yes, you are correct; there is an upper bound on the length of the preimages that could potentially be hashed; for SHA-1, SHA-224 and SHA-256, it's $2^{64}-1$ bits; for SHA-384, SHA-512, SHA-512/224 and SHA-512/256, it's $2^{128}-1$ bits.
On the other hand, for all practical purposes, these limits are effectively infinite. For example, with SHA-1, the limit of $2^{64}-1$ bits, that's $2^{55}$ blocks; even if we were able to the compression function in 10nsec, that means it'd take us over 10 years to hash a single value that size. It's unlikely that anyone would actually attempt that.