Some responses.. 

The proposal adds another gratuitous limit to the system: A maximum
transaction size where none existed before, yet this limit is almost
certainly too small to prevent actual DOS attacks while it is also
technically larger than any transaction that can be included today
(the largest possible transaction today is 1mb minus the block
overheads).  The maximum resource usage for maliciously crafted 1MB
transaction is enormous and permitting two of them greatly exacerbates
the existing vulnerability.

I think that limiting the maximum transaction size may not be the best possible solution to the N^2 hashing problem, yet it is not a bad start.

There are several viable soft-forking solutions to it:

1- Soft-fork to perform periodic reductions in the maximum non-segwit checksigs per input (down to 20)

2- Soft-fork to perform periodic reductions in the number of non-segwit checksigs per transaction. (down to 5K)

3- Soft-fork to perform periodic reductions in the amount of data hashed by non-segwit checksigs.

Regardless which one one picks, the soft-fork can be deployed with enough time in advance to reduce the exposure. The risk is still low. Four years have passed since I reported this vulnerability and yet nobody has exploited it. The attack is highly anti-economical, yet every discussion about the block size ends up citing this vulnerability.

, 

> Assuming the current transaction pattern is replicated in a 2 MB plain-sized block that is 100% filled with transactions, then the witness-serialized block would occupy 3.6 MB

But in a worst case the result would be 8MB, which this document fails
to mention.

I will mention this worst case in the BIP. 

Even if artificially filling the witness space up to 8 MB is anti-economical, Segwit exacerbates this problem because each witness byte costs 1/4th of a non-witness byte, so the block bloat attack gets cheaper than before. I think the guilt lies more in Segwit discount factor than in the plain block size increase.

I would remove the discount factor altogether, and add a fixed (40 bytes) discount for each input with respect to outputs (not for certain input types), to incentivize the cleaning of the UTXO set. A discount for inputs cannot be used to bloat an unlimited number of blocks, because for each input the attacker needs to first create an output (without discount). There is no need to incentivize removing the signatures from blocks, because there is already an incentive to do so to save disk space.

> Deploy a modified BIP91 to activate Segwit. The only modification is that the signal "segsignal" is replaced by "segwit2x".

This means that BIP-91 and your proposal are indistinguishable on the
network, because the string "segsignal" is merely a variable name used
in the software.

No, it is exposed to the rpc mining interface (getblocktemplate). It must be redefined, even if it's not a consensus change.