server: control plane shares port 443 with the admin UI

Captive portals, hotel wifi and corporate firewalls routinely permit only
80 and 443 — exactly the networks this tool exists to diagnose. A control
plane on 8443 is unreachable precisely when it matters most, and it fails
as "cannot reach server", which tells the user nothing about why.

The two cannot share a certificate, so sharing the port needs two names.
The control plane is trusted by SPKI pin and uses a long-lived
self-signed certificate; a browser needs one a CA vouches for. One name
on one port is one certificate. Pinning the Let's Encrypt key instead was
considered and rejected: it survives renewal only while key reuse holds,
so a routine key rotation would brick the fleet.

One listener now picks the certificate by SNI and the handler by Host.
Both have to agree, or a client gets the pinned certificate with the
admin UI behind it.

8443 stays open. Devices enrolled before this carry that URL in their
settings, and closing it for the sake of a port number would strand every
one of them; it can go once nothing points at it.

Verified per SNI on 443: fmr.echo-lot.app serves the Let's Encrypt cert,
fmr-1.echo-lot.app serves the self-signed one whose pin is unchanged, and
/v1/profile answers 401 on the control name against 303 to the login page
on the UI name.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

docs/build-status.md

@@ -1189,3 +1189,30 @@ beside an admin session.
 
 The one thing that changes on the device side is that the POST becomes HTTPS. That is a real
 certificate rather than a self-signed one, so it costs a URL scheme rather than any trust plumbing.
+
+## The control plane shares port 443 (2026-08-01)
+
+`fmr-1.echo-lot.app:443` is the control plane, `fmr.echo-lot.app:443` the admin UI, both on
+`.150`/`::150`, one listener, selected by SNI for the certificate and by `Host` for the handler.
+
+The reason is not tidiness, it is reachability. Captive portals, hotel wifi and corporate firewalls
+routinely permit only 80 and 443 — which is exactly the population of networks this tool exists to
+diagnose. A control plane on 8443 is unreachable precisely when it matters most, and it fails as
+"cannot reach server", which tells the user nothing.
+
+They cannot share a certificate, which is why this needs two names. The control plane is trusted by
+SPKI pin and so uses a long-lived self-signed certificate; a browser needs one a CA vouches for.
+One name on one port is one certificate, so the port can only be shared by splitting the names.
+Pinning the Let's Encrypt key instead was considered and rejected: it survives renewal only while
+key reuse holds, so a routine key rotation would brick the whole fleet.
+
+Verified per SNI on 443: `fmr.echo-lot.app` serves `issuer=Let's Encrypt`, `fmr-1.echo-lot.app`
+serves the self-signed cert whose pin is unchanged (`zRV9…Xlg=`), `/v1/profile` answers 401 on the
+control name and 303 to the login page on the UI name.
+
+**8443 stays open.** Devices enrolled before this carry that URL in their settings, and closing it
+for the sake of a port number would strand every one of them. It can go once no enrolled device
+still points at it — not before.
+
+The rule from the naming change still binds: `fmr` may be a CNAME to exactly one host and never a
+multi-address record, because a pinned client that reaches a different key does not fail over.