Port FIX adapter to wingfoil-next

[0-jake-0]

Ports the classic wingfoil::adapters::fix (Financial Information eXchange) adapter onto the wingfoil-next Op model — Phase 4 item 6 of the port plan. Faithful superset port of the synchronous, poll-based FIX session engine; like classic (and unlike the async adapters) it uses no async/tokio.

What's ported

• Codec + session state machine — the hand-written FIX 4.4 tag-value codec (encode/decode/build/find/drain), the FixSession logon/logout/heartbeat machine, and handle_initiator/handle_acceptor — a verbatim port.
• Both poll modes —
  • FixPollMode::Threaded rides source_at_start (a background session thread over the channel layer, with initiator reconnect-after-drop, acceptor re-accept, and the lock-free kanal inject channel drained per loop);
  • FixPollMode::AlwaysSpin rides a busy-spin custom_node (non-blocking socket reads on the graph thread), with the socket connect/bind deferred to graph start() via compose_spawn_at_start and a best-effort Logout at teardown.
• Full public surface — fix_connect / fix_accept (both modes), fix_connect_tls / fix_connect_tls_logon, FixConnection + fix_sub, FixOperators::fix_send, FixSender / SendError, and the pluggable FixLogon auth seam (None / Password / custom Ed25519-signer over the LogonContext).

Engine change (custom_node honours ALWAYS)

A busy-spin custom_node must set the engine's has_always flag so the realtime kernel doesn't park between cycles. custom_node accepted an Activation but ignored the always bit's kernel implication (only poll set it), so an ALWAYS custom node — a socket-polling source — was never driven each cycle. custom_node now sets has_always when activation.always (mirroring poll). Surfaced by the FIX AlwaysSpin source; guarded by fix_same_process_spin. Recorded as register A7. No effect on NONE/SCHEDULES custom nodes (existing custom_node/custom_op tests still pass).

Deviations from classic (all capabilities preserved)

1. The source factories take a &GraphBuilder + RunMode and reject RunMode::HistoricalFrom at wiring (classic checked real-time-ness at run start(); the message is the same, "real-time"). Every next source wires onto a builder, and the Threaded mode's channel receiver would block-collect a never-closing live stream and deadlock at start.
2. The sources return Streams (not Rc<dyn Stream>); fix_send returns Result<Stream<()>> and fix_sub a Stream<()>. Connect + logon still happen at start(), Logout at teardown.
3. The Threaded teardown drops classic's Arc<Mutex<Option<TcpStream>>> socket-shutdown handle for a stop-flag-against-the-200 ms-read-timeout exit (the zmq pattern — no lock on the graph path; teardown costs up to one read-timeout longer).

The canonical deviation list is the adapter's # Deviations from classic module-doc block plus next/docs/deviation-register.md (A7).

Tests & example

• Unit (inline, no network): encode_decode_roundtrip, custom_logon_fields_are_sent, password_logon_sends_username_and_password, fix_sender_queue_full, fix_sender_closed.
• Wiring (tests/fix_adapter.rs, feature fix): historical rejection for every source factory + fix_send's realtime-only start check.
• Integration (tests/fix_integration.rs, feature fix-integration-test): same-process acceptor+initiator round-trip in both poll modes, connection-refused, and initiator-reconnect-after-drop — real loopback TCP, no container.
• Example: classic fix_loopback ported to examples/fix_adapter.rs. The other four classic fix examples are peer-/credential-dependent (fix_client/fix_echo_server need a counterparty; lmax_demo/lmax_instruments need an LMAX account) and are not ported.

The credentialed LMAX-demo integration tests are not ported (external credentials).

Bookkeeping

next/docs/port-plan.md (fix marked ✅), next/docs/deviation-register.md (A7), the /new-adapter-next skill (busy-spin custom_node subsection), and CI (.github/workflows/fix-next-integration.yml + hub registration).

Verification

cargo fmt --all --check, cargo clippy -p wingfoil-next (default) and --all-features clean under -D warnings, all wingfoil-next test suites green with --features fix, the 4 integration tests pass with --features fix-integration-test -- --test-threads=1, and the example runs. (Substituted the scoped cargo clippy -p wingfoil-next --all-features for the workspace lint-all per the sandbox caveat — the classic aeron C toolchain is unavailable here; CI runs the full workspace lint.)

🤖 Generated with Claude Code

https://claude.ai/code/session_01HA2a5JgbvUJCqrRAswbS7f

---

Generated by Claude Code