core+qt: add opt-in asynchronous model registration

[Yaraslaut]

Summary

• Bridge registers a model via IBackend::registerModel, which is synchronous. QtWebSocketBackend implements that with a nested QEventLoop parked until the reply arrives. On a WASM main thread, Qt refuses to spin a nested loop at all, so the very first registerModel call aborts the page — exactly the issue's repro.
• Adds IBackend::registerModelAsync(typeId, factory, contextKey, onRegistered, onError), defaulting to "unsupported" (false, calls neither callback) so every backend that hasn't opted in is unaffected. Bridge::registerHandler() (both overloads) prefers this path when a backend offers one, falling back to the synchronous call otherwise.
• The success callback guards against a switchBackend() racing ahead of a still-pending registration, reusing the exact weak_ptr<IBackend> + liveness-token pattern installReconnectHandler already uses — a stale reply can never clobber the fresh id switchBackend's own re-registration already assigned. Verified with a dedicated regression test.
• QtWebSocketBackend overrides registerModelAsync, matching execute()'s existing callId-based reply matching (no protocol change needed — the server already echoes callId on every reply kind, register included).
• Found and fixed while implementing: making the Qt override unconditional silently broke the existing Qt test suite — measured directly, 20 of 50 test cases failed, because code that fires an action immediately after constructing a BridgeHandler (assuming synchronous-enough registration) now sees "handler not bound" instead. Gated the override behind a new QtWebSocketBackendConfig::asyncRegistrationEnabled, defaulting to false: every existing embedder's behavior is unchanged; a WASM build opts in explicitly and adapts its call sites to wait for registration before firing an action. Confirmed by running the real Qt/WebSocket test suite against Qt6 both before and after this fix.
• Scope: only the plain (non-shared) registration path — a BridgeHandler's initial construction — goes through this. registerModelShared/attachModel (shared/keyed handlers) and the re-registration switchBackend()/reconnect-handler paths remain synchronous; making those async too is a larger change to Bridge's locking model, left for a follow-up if it proves necessary.
• Design docs updated: docs/spec/core/backend.md (new "Asynchronous registration" section, IBackend/QtWebSocketBackend API reference rows) and docs/spec/core/bridge.md (registerHandler behavior).

Test plan

• tests/test_async_registration.cpp (new): async path preferred and binding starts unbound; async execute works once registration completes; onError leaves the binding unbound with no crash; a stale reply after switchBackend() does not clobber the new id; a reply arriving after ~Bridge() is a safe no-op; a backend with no async support still falls back to the synchronous path.
• tests/qt/test_qt_websocket.cpp (new case): end-to-end against a real Qt WebSocket server — registration doesn't block, and execute succeeds once the deferred registration completes.
• Full suite: ./build/tests/morph_tests — all 817 test cases / 8305 assertions pass.
• Full Qt suite built and run against real Qt6 (-DMORPH_BUILD_QT=ON): 51 test cases / 340 assertions pass — confirmed byte-for-byte matching the pre-change baseline (334 assertions / 50 cases) plus the one new case.

Closes #26

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!