• Assuming that “cosmic variance limited” means the noise variance is 1/10 of the signal variance, I find PICO to be cosmic variance limited in EE out to ell = 2314. If I make that more conservative, requiring the noise variance to be 1/100 of the signal variance, I find ell = 1698.

• I've quickly rerun the forecast using 60% of the sky to confirm I didn't make any mistakes and I get \sigma(\tau) to be 2e-3, as before (for Version 4.1):

  components: sync+dust
  delensing option: iterative delensing CMBxCMB 
  σ(τ=0.066)=1.95e-03
  σ(h=67.74)=8.41e-02
  σ(As=2.142e-09)=7.43e-12
  σ(r=0.0)=9.41e-05
  σ(Ωch2=0.1188)=2.11e-04
  σ(ns=0.9667)=1.08e-03
  σ(Ωbh2=0.0223)=1.86e-05
  effective level of foreground residuals, reff = 8.74e-07
  degradation of the noise after comp. sep. = 1.10e+00
  noise in the cleaned CMB map [uK-arcmin] = 9.02e-01

we're also assuming that the spectral indices are constant over N_side = 4 patches (which is optimistic)

These forecasts were run using the CMB4cast Fisher matrix code (http://portal.nersc.gov/project/mp107/index.html, Errard & Feeney et al.), assuming access to T, E, B and d information, with the deflection estimated using the iterative EB estimator. The code assumes Planck-2015-level synchrotron and dust foregrounds, forecasting the experiment's ability to clean these foregrounds using a parametric maximum-likelihood approach, assuming the foreground spectral indices are constant on patches of size ~15 degrees across (N_side = 4). This is all probably a little out-of-date, being based on the Planck 2015 results and cosmology, but it doesn't seem to give a significantly different answer to Raphael's code (and I can rerun with a different tau if you'd like).