the delay is not an optimal solution, a barrier would be much more
appropriate, but I have not found the right place

The VFP trunk now requires Varnish master since the commit that
changed OF_CHGGZIP to OF_CHGCE.

See the discussion in Varnish issue #2910. A 304 IMS beresp could
(weirdly) set Content-Encoding to something other than the result
of previous VFP filtering, in this case compressing the response
with gzip. Nevertheless, Varnish does the right thing when the
OF_CHGGZIP flag is set.

Set the OF_CHGGZIP flag to indicate that:
- The ETag in a client response must be weakened if the response
  body differs from the fetched response.
- Any Content-Encoding header in the fetched 304 response should
  be removed.

... otherwise vcl_init fails with varnish-cache master and also
we would leak on vcl destroy.

to libbrotlicommon because it will not be found implicitly

See also https://github.com/google/brotli/pull/740

The same information is in COPYRIGHT.

As do the g(un)zip VFPs for Varnish.