First, the graceless error handling is fixed in
204d34ff961d6dabf18b255decc29712e03afef0.

Second, here’s a preliminary patch that almost works with
<https://hydra-mirror.marusich.info>.

The problem is that sometimes the server closes the connection
unexpectedly, leading to an obscure backtrace like this:

--8<---------------cut here---------------start------------->8---
substitute:  629: 6 [lookup-narinfos "https://hydra-mirror.marusich.info" #]
substitute:  585: 5 [fetch-narinfos "https://hydra-mirror.marusich.info" #]
substitute:  510: 4 [http-multiple-get # ...]
substitute: In web/response.scm:
substitute:  197: 3 [read-response #<input-output: gnutls-session-port 38f59c0>]
substitute: In web/http.scm:
substitute: 1157: 2 [read-response-line #<input-output: gnutls-session-port 38f59c0>]
substitute:  151: 1 [read-header-line #<input-output: gnutls-session-port 38f59c0>]
substitute: In unknown file:
substitute:    ?: 0 [%read-line #<input-output: gnutls-session-port 38f59c0>]
substitute: 
substitute: ERROR: In procedure %read-line:
substitute: ERROR: Throw to key `gnutls-error' with args `(#<gnutls-error-enum Error in the pull function.> fill_session_record_port_input)'.
--8<---------------cut here---------------end--------------->8---

The “error in the pull function” is because ‘gnutls_record_recv’ got
ECONNRESET while reading.

I wonder whether this could be due to the particular configuration of
nginx at Cloudfront, so I’ll try with another server (I’ve set up Let’s
Encrypt on that server but it’s not accessible yet via port 443.)

To be continued!

Ludo’.