Divergence notes

Undertow departs from its references only on purpose, and every departure is written down — here and in tests/fixtures/wire/README.md. Silent divergence is treated as a bug.

IClient echo is verbatim, not key-sorted

The Gleam fixtures show the client’s IClient re-serialized with term-sorted keys — an Erlang map cannot preserve insertion order. Undertow preserves the supplied key order instead.

This matters because the Fluid container loader seeds its own audience entry with the object it sent (original key order), and its assert 0x4b2 requires byte-identity with any later add for the same client id. In live browser flows the Gleam reference’s re-sorting tripped exactly this: the loader received its own presence join back in a different key order and closed the container (todo-list e2e: 6/9 before, 9/9 after).

The shared fix, landed in both servers, pins initialSignals in IConnected to [] — matching Levee — so the loader never receives its own join back through a byte-reordering path. Undertow keeps the verbatim echo anyway: every remaining echo path (join op data, initialClients, peer presence signals) then matches the client’s own bytes too, which a sort can never do. The wire differ flags the key order inside those payloads as the one expected difference against the recorded fixtures.

Phoenix bad-vsn rejection is 403

A Phoenix join with vsn=1.0.0 is rejected 403 before the WebSocket upgrade. An earlier plan document assumed 400; the captured fixture (phoenix-bad-vsn.txt) pins 403, and the fixture wins.

Drop-in parity: one intentional 401

Against the Levee integration suite, Undertow passes 53 of 54 cases. The remaining case intentionally returns 401, recorded as ADR-009 — a deliberate authorization-behavior decision, not an accident of porting.

Baseline skips

The Routerlicious conformance baseline includes 3 skipped tests and 1 todo. These are the reference suite’s own state at capture time — Floodgate posts the same counts.