[Aschenbath]

Closes #1463.

What this does

When the overview's 3 s deadline fired, it called push(key) and then flush() on the fast layer. That also invalidated any fast key queued in the last 300 ms, before that key had finished its own settle window. Now the overview deadline invalidates only the overview. Only flushAll, used on reconnect and teardown, still sends the overview through the fast layer. That way a reconnect's broader ["graph", kbId] covers it and the overview isn't fetched twice, which is what the old comment was protecting.

The two layers moved out of the hook into an exported createKbInvalidations(invalidate, kbId), so they can be tested without React. useKbEvents calls it, and its behaviour is otherwise unchanged.

Not changed: #1462

#1462 asks for a true debounce. #1311 deliberately refreshes the overview every few seconds while a source pushes. A debounce that resets on every event would freeze the overview for the whole push. I left that issue alone so a maintainer can decide it.

How it was checked

On Windows 11 with Node and pnpm, commit 79b34901 on dev 1d94b532:

• pnpm exec vitest run: 49 files, 342 tests passed. Four of them are new, in "the overview deadline and the fast keys". With the old push(key); flush(); restored, the first new test fails, and it passes with this change.
• pnpm exec tsc -b: passed. pnpm build: passed (existing chunk-size advisory only).

🤖 Generated with Claude Code

[WaylandYang]

Thanks for this, and sorry for the overlap. I missed your PR and fixed #1463 in #1465, which is now on dev, so I'm closing this one. Your read of the bug and of #1462 was right.