test(registry-feed): filter assertions by actor to fix OR-filter flake (closes #4413)

[bokelley]

Summary

Fixes the flake in `server/tests/integration/registry-feed.test.ts > Registry Feed Integration Tests > type glob filtering > combines multiple type filters with OR` that's been failing on `main` and every PR (example failing run on main, example on PR #4412).

Root cause

The four assertions in the `type glob filtering` describe block were checking `feed.events` length directly. `queryFeed` has no actor filter — by design, since actor isn't part of the public feed contract — so when other test files (the background crawler in particular, which the server boots at integration test start) write `property.` or `agent.` events with non-`test%` actors, those events end up in the result and skew the count.

The file already filters by actor in two earlier tests (line 57: `feed.events.filter(e => e.actor.startsWith('test'))`; line 78: `feed.events.filter(e => e.actor === 'test')`) precisely because the cleanup hook (`DELETE FROM catalog_events WHERE actor LIKE 'test%'`) is intentionally scoped — see the comments at lines 19-23 explaining sibling test files writing trigger-generated events with `actor LIKE 'trigger:%'`.

The OR-filter assertion has the lowest noise tolerance because its expected count (4) is closest to what crawler-written events can push the unfiltered total to (consistently observed as 9 = 4 seeded + 5 crawler-leaked). The narrower filter assertions in the same block (expected 1 and 3) happen to pass because their type filters don't catch the crawler's event types as aggressively.

Fix

Add an `ours` helper that filters by actor before counting, applied to all four assertions in the describe block:

```ts
const ours = (events: { actor: string }[]) =>
events.filter(e => e.actor.startsWith('test'));
```

This makes all four assertions robust to whatever the leak source is, without changing what they test — they were always testing "my seeded events + this filter", which is what the new assertions check explicitly.

Test plan

• Local run against a fresh Postgres: `10/10 tests pass in registry-feed.test.ts` (was 9/10 with the OR test failing).
• Typecheck clean, no other test files touched.
• Pattern matches what the file's other tests already use, so future readers won't be surprised.

Closes #4413.

🤖 Generated with Claude Code