PGHOST doc/code mismatch

[progman1]

the README says it will be taken as a unix socket if PGHOST has a 1st char of /.
that doesn't work because the code (describe_connection) only does the check on ~host parameter values, not both.
so, in fact, pgocaml's interception of PGHOST prevents it from being set with a socket directory since an exception is raised trying to open it as a network address.

here's my fix:

diff --git a/src/PGOCaml_generic.ml b/src/PGOCaml_generic.ml
index a1a0567..536ee79 100644
--- a/src/PGOCaml_generic.ml
+++ b/src/PGOCaml_generic.ml
@@ -291,0 +291,0 @@ type uuid = string
@@ -1022,7 +1019,11 @@ let describe_connection ?host ?port ?user ?password ?database
       `Unix_domain_socket_dir s
     | None, None ->
       try
-        `Hostname (Sys.getenv "PGHOST")
+        let h=Sys.getenv "PGHOST" in
+        if String.length h > 0 && String.get h 0 = '/' then
+          `Unix_domain_socket_dir h
+        else
+          `Hostname h
       with
         Not_found -> (* fall back on Unix domain socket. *)
           `Unix_domain_socket_dir PGOCaml_config.default_unix_domain_socket_dir

I also found a problem in my installation with the generated cfg file
PGOCaml_config.ml which confounded the above issue.
installing via opam gave this:

let default_unix_domain_socket_dir = "/tmp"

bizarre as the /var/run/postgresql directory does exist.
deleting this cfg file in the dune _build directory under opam and rebuilding/reinstalling regenerated it correctly!

on another note a source of great frustration was the unchanging behaviour of the test executable when changing PGXXX vars.
even after I grokked it I still tripped up again and again.
eventually I added this to the test dune stanza:

 (preprocessor_deps 
  (env_var PGHOST)
  (env_var PGUSER)
  (env_var PGDATABASE)
  (env_var PGPASSWORD)
  (env_var PGPORT)
  (env_var PGPROFILING)
  (env_var PGCUSTOM_CONVERTERS_CONFIG)
  (env_var PGCOMMENT_SRC_LOC)
 )

... and the ppx stage of db access worked reliably during experimentation/dev/test without further stumbling! I recommend this or a cut-down version is added to the test case and highlighted to potential users!

lastly, note the PGCOMMENT_SRC_LOC - the README.md lists it without the PG prefix incorrectly.