test(e2e): wait for the loaded channel list in the #1468 tests

[adminopenclaw8-sketch]

Relates to #166.

What

This is a cherry-pick (git cherry-pick -x) of 403075e8 from #166, with no other changes. Only test-e2e-playwright.js changes (+28, test-only). #166 can only merge after #153, so the fix gets its own PR against master.

The race

The two #1468 E2E tests ("live WS CHAN message with no payload.channel is dropped" and its control) snapshot window._channelsGetStateForTest() as soon as _channelsProcessWSBatchForTest exists. channels.js sets these WS test hooks synchronously in init(), before loadChannels() has resolved /api/channels. So:

• Orphan test: before can read 0 channels. If the response lands between the before and after evaluates, after reads the full list and the test fails with before=0, after=5, although the orphan message was dropped.
• Control test: a WS-born channel pushed while the request is in flight is not in the list loadChannels() installs. mergeClientChannelState() only enriches rows that the snapshot has.

The flake has been seen on #166 and on #178 (which does not touch channels), both with before=0, after=5.

waitForChannelListSettled() waits until #chList has rendered the list (no .ch-loading) and the channel count is > 0 and unchanged for 300 ms. Both tests take their before snapshot only after that.

Evidence

The setup is a local Go server built from this branch on port 13581. It uses CI's fixture prep from deploy.yml on a copy of test-fixtures/e2e-fixture.db: freshen, the Kpa-clawbot#1486/Kpa-clawbot#1791 seed rows, migrate, and seed-2073-route-adverts.sql. The server serves plain public/, not the instrumented copy. Playwright 1.58.2 runs headless Chromium.

The two test blocks were taken verbatim from each version of test-e2e-playwright.js and run in a loop. Each iteration starts on #/home so #/channels is a hash navigation, as in the full suite. /api/channels was delayed with page.route.

/api/channels delay	master (old test): orphan fails / control fails	this branch: orphan / control fails
none, 20 runs	0 / 0	0 / 0
400 ms, 20 runs	0 / 0	0 / 0
5 ms, 10 runs	2 / 0	0 / 0
10 ms, 10 runs	0 / 1	0 / 0
15–60 ms (6 × 10 runs)	0 / 0	0 / 0
random 0–60 ms, 50 runs	2 / 3 (5 of 50 runs failed: 10 %)	0 / 0

Master's failures reproduce the CI message exactly: channel count unchanged after orphan WS msg — before=0, after=5.

A fixed 400 ms delay does not trigger it. With 400 ms, both evaluates run before the response arrives, so before=0, after=0. The failure needs the response to land in the few milliseconds between the two evaluates. A probe confirms the window: the hook exists ~18 ms after navigation with 0 channels, and the list holds 5 channels once the response is in.

On this branch, the two tests passed 180 runs in a row across all delays.

Mutants (from the cloud review)

Each mutant was served as a mutated channels.js via page.route. No file was edited.

Mutant	Orphan test	Control test
if (!payload.channel) payload.channel = 'unknown';	fails 10/10 (5 without delay, 5 with 400 ms)	passes
Drop everything (continue;)	passes	fails 10/10

Merge note

403075e8 stays on #166's branch, and #166's history is unchanged. Once this PR is merged, the same change arrives in master twice: here, and via #166's own commit. Merging #166 afterwards is a no-op for this hunk, because both sides made identical changes, so there is no conflict and no diff.

deploy.yml is unchanged (9 × github.repository == 'Kpa-clawbot/CoreScope').

🤖 Generated with Claude Code