The demo shows the refusal it claims

[onatozmenn]

demo/ is what this project leads with — "One clause. Two worlds." It is the
one-clause page and the sentence the
capability model is sold on.

Both it and the site say:

A host that provides Io.read but not Io.save can run read_only without
restriction. It refuses read_write because the module declared a need the
host cannot meet.

The refusal happens before any code runs.

The test under that sentence ran the module with no host at all and
asserted only that the answer began deed:io:

let stopped = call(&module, "process", &[Value::I64(0), Value::I64(0)])
    .expect_err("read_write needs a host to provide save");
let Trap::NeedsAHost(what) = stopped else { … };
assert!(what.starts_with("deed:io"));

What it actually said

Asked it to name the operation it is named for — what == "deed:io.save":

FAIL  running_read_write_without_save_names_what_it_wanted
      it should name an Io operation: deed:io.read

process reads before it saves, so with no host it stops on read. A test
called running_read_write_without_save_names_what_it_wanted never mentioned
save, and the operation it did name is the one both modules share — the
opposite of the difference the demo exists to show. The two doc comments in the
file disagreed with each other about this; the second one already said "or
deed:io.save depending on execution order".

A host that behaved the way this test does would have read the file before
noticing it could not save. That is not "before any code runs".

What it says now

Host::link is the mechanism the prose describes. It reads the whole import
section and answers before an instruction runs:

let mut host = Host::new();
host.offer("deed:io", "read", |_| Ok(Some(Value::I64(0))));

let refused = host
    .link(&module_for("read_write.deed"))
    .expect_err("a host that cannot save should refuse the module that saves");

assert_eq!(refused.name, "save");

Three tests where there were two:

• a_host_without_save_refuses_read_write_and_names_save
• the_same_host_links_read_only_without_restriction — the half that makes the
  first half mean something. A host that refused both would be a host with no
  filesystem.
• running_without_a_host_stops_at_the_first_import_either_way — the weaker
  guarantee, kept and renamed, asserting both modules stop at deed:io.read so
  the two shapes can be told apart.

The other listing on the page

The README prints this file's test names. It was typed, said running 5 tests,
and named a test that no longer exists. It is now read back and compared to the
tests the file declares — the same treatment the import listing on that page
has had since #783.

That ratchet found something the moment it ran: it counted itself, because
the marker it splits the source on is spelled in the source it reads. Fixed by
spelling it in two pieces, which is the sort of thing only a self-reading test
runs into.

Also

demo/README.md said the host is "the runner in crates/deed-rt". That crate
is hashing, http, reach and sandbox; Host, link and call are all
in crates/deed-codegen.

The site carries the same snippet, the same transcript and the same claim, so
it needs the matching change: deed-lang/deed-lang.github.io#44.