The intake feed test fails under -race on CI, on an assertion that was never due

[jorgemanrubia]

Main's Race Detection job is red, and one of the reasons is this test: the
end-to-end intake feed test fails under -race on CI with a bare assertion
that never became true — on main's own push run at 66b1a3a
(run 35349470544)
and on at least one pull request built from it. The race detector reported no
data race. There is none.

The test was reading across an order it had not waited for. It serves a live
event on the socket, waits for that event to reach the ledger, and then reads
the durable feed position. A fresh ledger has no position to resume, so the
entry is present-class: the feed holds the page's position through the walk
and saves it only after the drain has accepted every event buffered during the
walk. The live event is one of those. Its row in the ledger says the drain
reached it, not that the save has run.

Instrumented to report what was true at the instant of that read, under twenty
competing processes on two cores, the position was absent and arrived 3.8ms,
33.2ms and 34.8ms later — with the walk's one page delivered, both pointer
lines written and both ids handed over. Everything the two events owed was
done; only the walk's position was outstanding. Under -race on a loaded
runner that window is wide enough to lose, which is exactly what CI kept
losing.

So the assertion was wrong, not the code: nothing guarantees the position is
durable at the moment the live event's row appears, and the feed's ordering —
deliver, drain, then save what the entry held — is deliberate. The test now
waits for the position to be saved before reading it, and waits for the
pointer lines and the queue hand-off too, since the ledger row is written
before both. Waiting for the condition rather than widening a timeout is the
same fix shape as
The recovery harness reads a worker's word once the worker has said it,
and the restart test next door already worked this way.

Under load on two pinned cores, over nine batches: 15 of 245 runs failed
before, always at that line and never with a data race; 0 of 90 after, the
after batches alternating with before batches in the same sessions. Unloaded,
both are clean. The test introduced in
Take the account event feed into a durable ledger
is the only thing that changes here.

[Copilot]

🟢 Approved

The waits correctly reflect the asynchronous ordering and preserve exact result assertions.

Pull request overview

Updates the intake feed test to wait for asynchronous side effects before asserting results.

Changes:

• Waits for checkpoint persistence.
• Waits for pointer output and queue handoff.

[!TIP]
If you aren't ready for review, convert to a draft PR.
Click "Convert to draft" or run gh pr ready --undo.
Click "Ready for review" or run gh pr ready to reengage.

File summaries

File	Description
internal/connector/intake_feed_test.go	Synchronizes assertions with feed processing.

Review details

• Files reviewed: 1/1 changed files
• Comments generated: 0
• Review effort level: Balanced

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.