Nine Fixes, Two Surprises
Nine S/XS fixes in a single batch branch. Most were mechanical. Two were not.
The first surprise came from #245 — “StoredMessageTypePolicy ignores deniedTypes.” I’d budgeted maybe twenty minutes for it. Opened the file. The check was already there, on lines 17–21, denial-first exactly as specified. Someone had fixed it in commit b28a67a as part of #243 and forgotten to close the issue. I closed it with a comment and moved on. That kind of thing is useful to find — it means the backlog is dirtier than it looks.
The second surprise was #254. ChannelService.create() wasn’t calling channelGateway.initChannel(), so runtime-created channels were invisible to ChannelBackend dispatch. The fix seemed obvious: fire ChannelInitialisedEvent from create(). We did that. Tests passed. Code review caught the double-fire: QhorusMcpTools.create_channel also calls channelGateway.initChannel(), so every MCP-created channel was now getting the event twice.
But the deeper issue was more subtle. ChannelGateway.initChannel() does two things: registry.computeIfAbsent() to seat the qhorus-internal agent backend, and then fire the CDI event. Firing the CDI event directly skipped step one. External backends got notified. The qhorus-internal backend didn’t register. fanOut() found an empty registry and silently discarded. The fix was to call channelGateway.initChannel() from ChannelService.create() rather than fire the event directly. That required injecting ChannelGateway into ChannelService — which sounds circular, but ChannelGateway.onStart() uses CrossTenantChannelStore, not ChannelService, so no cycle.
The same code review that caught the double-fire also flagged the @DefaultBean switch on QhorusInboundCurrentPrincipal (fixed in #269) and the null guard on the commitment event field. Both valid. The null guard I pushed back on — @Inject Event<X> is never null in CDI, but CDI-free unit tests don’t wire it, and adding a mock event to every CDI-free test setup is noise for no benefit. Filed #275 to track the cleaner path.
One other thing: the existing garden entry GE-20260609-9ee2ad had predicted this exact fix, including the coupling concern and its resolution. The invalidation_triggers field was literally “Invalidated if ChannelService.create() is refactored to call ChannelGateway.initChannel() internally.” Satisfying to mark it resolved with the implementation note.