Merge pull request 'Issue 059 — the broker credential names the hub, not the broker's node' (#50) from issue/059-broker-address-review into main

2026-09-17 22:37:26 +02:00
parent 5395b6a47f d1b5210d83
commit 101848242f
1 changed files with 58 additions and 0 deletions
@@ -0,0 +1,58 @@
+---
+status: located
+opened: 2026-09-17
+located-in:
+  - mesh-controller
+fixed-by:
+amended-design:
+---
+
+# 059 — The broker credential names the hub, not the node holding the broker
+
+## Symptom
+
+An adversarial review of the issue-055 fix (`brokerReachableAt`) found the cross-node broker
+address is computed from the wrong facts. No live failure yet — every current mesh has the broker
+on the hub — but each is a wrong invariant that fails exactly when the topology stops being the
+simple one:
+
+1. **The helper returns the hub's overlay name for any on-overlay node, but nothing ties the
+   broker to the hub.** The broker lives wherever the module holding the `mesh-broker` seat is
+   assigned; hub-ness is a separate operator declaration — and the overlay design actively pushes
+   them apart (the hub must be publicly dialable; a natural topology is hub-on-a-relay,
+   foundation at home). In that mesh every issued credential dials a relay where no broker
+   listens.
+2. **"On the overlay" is tested as "has an overlay address."** The codebase itself defines
+   membership as *resolved the networking module* — "not 'has an address'… it is what resolved
+   the module" — and excludes address-only nodes from every peer list. A node placed but not yet
+   given networking is sealed an overlay URL it cannot route.
+3. **A credential issued before `overlay place` keeps the public URL for ever**, and nothing says
+   so; the remedy (re-issue + push) works but is undocumented. The sibling of issue 057's
+   ordering gap.
+4. **A portless `MESH_BROKER_ADDRESS` silently disables the fix** — `SplitHostPort` fails and the
+   helper falls back to the public address without a word. The port should default to 5671, or
+   the address should be refused at genesis.
+5. **Two hubs are never refused.** A second `overlay place --hub` leaves both flagged; the pick is
+   then order-dependent. `ErrNotOneHub` exists and nothing returns it.
+
+One review claim is recorded here as **disputed**: that the from-mesh forward rule could not have
+been the 055 failure because the broker port is opened from anywhere. The live measurements during
+055 (nftables counters; a cross-node probe of the amqps port flipping CLOSED→OPEN exactly when the
+broker module's `listens` gained the port) say the forward chain was the gate for published
+container ports. The code-reading and the wire disagree; the wire was measured.
+
+## Why it matters
+
+The 055 fix is correct for every mesh that exists, and wrong by construction for the first mesh
+whose hub is not its control-node — which the overlay design itself recommends for a home mesh
+behind NAT. Credentials are sealed and long-lived, so the failure arrives long after the decision
+that caused it, on the first module pushed to a node that joined the reshaped mesh.
+
+## Open questions
+
+- Should the broker address resolve from the `mesh-broker` seat's assignment (falling back to the
+  hub only when nothing holds the seat — genesis), and should membership use the same
+  resolved-networking test the peer graph uses?
+- Where does the stale-credential remedy live: a warning from `overlay place`, an automatic
+  re-issue at the next push, or documentation?
+- Should a second `--hub` be refused outright (`ErrNotOneHub` finally thrown)?