@ljking First of all, thanks for the valuable detailed discussion, and sorry I didn't respond earlier. I've been conducting a bit of a review of available devices, and their merits and shortcomings in the role discussed here. I've also stumbled on the email list exchange between you and Grant over the design of the Sensors board, very interesting and informative too.
What's especially noteworthy, as you highlight, is the unusually high probability that these level-translator solutions behave poorly in a "wires hanging off the board" situation, versus their desirable behavior on a board with nice short well-routed traces.
That said, a few notes:
> "All the chips I tried... have one-shots".
I don't think the PCA9306 has edge accelerator one-shots. Actually, I think the PCA9306 should be quite close to the BSS138s in behavior, so long as attention is paid to the pin connection details that set the bias on the FETs' gates (and which should give it larger logic margins).
> "I tested the following..."
From a static DC analysis point of view, the PCA9306 and the TXS series are the only ones that should have a chance of working for I2C. The TXB and similar MAX3000/1 should not, and their datasheets say they should not.
I would also point out that the TXS0108, intended for I2C or at least open drain, has a particularly aggressive edge-acceleration scheme, accelerating both edges, not just the rising edge as in the seemingly same-family TXS0102 and 0104.
> "Side note: The TXB works fine for I2C (as long as cables are short)"
Are you sure it was the TXB? That seems counterintuitive judging from the circuitry. Perhaps you meant the TXS?