Is it a fata morgana or not: long FIR filters with low latency? The topic has been on my mind for years, but I was ignorant of techniques that were already invented decennia ago. Now I finally sorted out how it can be done in Pd.
The key points are: minimum-phase filter kernels and partitioned convolution. Pd-extended has Ben Saylor's [partconv~] object. A minimum-phase filter kernel can be computed by Hilbert Transform of the log-amplitude spectrum. All taken together, it is not quite trivial, but it works! I am happy, this will save me ~40 ms of latency in my real time performance patches.
See attached patch for a demo.