Deployed to a real VPS for the first time. The stack came up, the patched
wizard served, the plugins activated and the 40-check auth suite passed over
the public internet — and then every single export failed:
cURL error 28: Failed to connect to <public ip> port 8081 after 10001 ms
Not a bug in the export. A container cannot reach the host's own published
port on most Docker hosts; there is simply no route back in. The exporter
renders each page by asking the site for it over HTTP, so the one thing it
depends on is the one thing a container cannot do. This would have hit every
user of the documented install path on their first publish, and the error
says nothing about the cause.
SIRIUS_PRESS_LOOPBACK_URL gives the exporter a second address — in the
bundled stack, the nginx service on the compose network — while the request
still carries the site's real Host header. That matters twice: WordPress
renders exactly the page a visitor gets rather than redirecting to a
canonical URL the container cannot follow, and the internal hostname never
appears in the exported HTML. Verified on the VPS: the page builds, the
content is right, no absolute self-links, no `//web/` leaking out.
Also documents the question underneath it — what WP_HOME should be when the
site's public address is a BCNR name. Not the name: whatever the server is
actually reachable at. The exported links are document-relative, so they work
under the name regardless.
The docs said nobody had clicked the buttons in a browser. Somebody has now,
and it cost two bugs, so testing.md gains a section saying to do it after
touching the auth assets — the HTTP suites cannot see that class of failure.
Also sharpens the honest gap: the Docker stack has not been run anywhere, not
just 'not on a clean Ubuntu box'.
The subtree is in place, so everything that used to fetch and patch core at
build time now just copies it.
tools/build.sh copies wordpress/ — no download, no checksum step,
because there is nothing to fetch and nothing to
trust that is not already in the repository
docker/Dockerfile COPY wordpress/ instead of curl + sha256 + patch;
the build args and the `patch` package are gone
docker-compose.yml no WP_VERSION / WP_URL / WP_SHA256 to keep in step
tools/update-wordpress.sh is rewritten around what the subtree makes
possible. It imports the pristine release onto sirius-press/wordpress-upstream
and then `git subtree merge`s that branch, which three-way merges upstream
against the fork's own commit. A patch either applies with fuzz and hopes or
fails and leaves you re-deriving the change by hand; a merge conflict is
resolved once, in the file, and the next release merges against the
resolution.
patches/ survives as documentation rather than mechanism, and is now
generated: tools/refresh-patches.sh diffs the subtree against the pristine
import and rewrites the directory, with --check for CI. It answers the
question anyone auditing a fork asks first — what exactly did you change
inside WordPress? — in a minute, which `git log wordpress/` cannot, because
that log is mostly upstream imports. Generated documentation stays true; a
hand-maintained record of a core diff drifts, and a stale one is worse than
none because people trust it.
One test change worth noting: the syntax sweep no longer walks all of
wordpress/. It lints the fork's own PHP plus every core file patches/ says
the fork touches, which keeps the suite at seven seconds instead of a minute
while still covering the only core file that can break.
Running the fork against a live WordPress found a real hole in registration,
and it is the kind that only shows up when you actually try it.
ECDSA public-key recovery always succeeds. Given any well-formed signature
and any digest it returns a key — just not the signer's, unless the digest is
the one that was signed. The auth flow leaned on that as if a wrong message
would fail. It does not; it quietly yields a stranger's address.
At sign-in this was harmless, because the wrong address matches no account
and the attempt fails. Registration and wallet-linking were another matter:
both took the recovered address and bound it to an account, so a signature
over slightly different text — a challenge copied without its blank line, a
wallet that rewrote the text, a login signature replayed at the registration
form — created an account keyed to an address nobody could sign for. The
person would see "success" and discover the truth the next time they tried to
get in. Wallet-linking was worse still: it would move an existing account onto
a dead address and lock its owner out of their own site.
Both paths now require the address the signer claims and compare it to the
recovered one, which is what verification actually means. Sign-in accepts the
claim when the page sends it and uses it to turn "no account uses that wallet"
into the more useful "that signature is not over the text we asked for".
Also from running it:
URL rewriting mangled every link on a site whose URL carries a port. The
protocol-relative pass matched inside absolute URLs and gave each one a second
scheme, and matching the host without its port left the port stranded as
`//host:8760:8760/`. Local and staging installs would have exported a site of
broken links.
Plain permalinks silently collapse an entire site onto one exported file,
because every post's URL is `/?p=N` and its path is `/`. The queue looks
healthy the whole time. The Publishing screen now says so.
Translations loaded on `plugins_loaded`, which WordPress 6.7 warns about on
every request — the kind of noise that trains people to stop reading logs.
And one deletion: an `is_email()` filter written on the assumption that
WordPress rejects `.invalid` addresses. It does not — `is_email()` validates
syntax, not whether a domain could exist — so the filter never fired. A filter
that appears to relax a rule but does not is worse than no filter, because
someone later reasons from it. The documentation made the same claim and has
been corrected.
Verification added rather than asserted: tests/live.mjs drives a real instance
over HTTP (40 checks), and tests/mock-gateway.mjs answers uploads with the
signature check transcribed from the gateway's own source, so the publishing
path can be exercised without a registered name.
