#220 fix: the replication poll yields to the operator's click on both screens

[gregoryfoster]

Closes #220.

What was wrong

The issue had the ordering inverted. A poll landing after Replicate now is harmless: its target was already swapped out, so htmx drops it. The bug is the poll landing first. The action's target (resolved at click time) is then detached, and its swap, focus script and HX-Trigger toast are all discarded. If that poll read before the commit and no other command was open, its render stopped polling, and the section stayed on pre-issuance state until reload - with a fresh, enabled Replicate now button and no toast, inviting a second occasion. hx-confirm makes that ordering the likely one: the poll timer runs overdue behind the dialog and fires just after the POST. The RepSpec detail page had the identical defect.

The fix (both screens)

• Wrapper, inside the poll block: hx-sync="this:abort" hx-disinherit="hx-sync".
• Replicate now and Deactivate: hx-sync="closest #<section>:drop".

htmx takes the lock before confirm(), so an in-flight poll is aborted at the click, and a tick during the action is dropped; every needs no re-arming. hx-disinherit because hx-sync inherits: without it, anything inside the section - boosted links included - takes this:abort and is dropped mid-poll.

Tests

• poll_sync_violations (tests/dashboard/conftest.py) checks the rule on each screen's poll fragment. test_poll_sync.py holds the checker to the markups the issue weighed, so an empty list means safe, not unexamined. Red at 1d64ec5, green at ab4584c.
• tests/js/htmx-poll-sync.test.js drives the vendored htmx 2.0.8 through a stubbed XHR held in flight. The drop-aborts-an-abortable-poll rule is htmx's implementation, not its documented contract, so an htmx upgrade could undo Replication poll and Replicate now race for the same swap target #220 with every Python test green. Setting the actions to abort or the wrapper to drop fails it.

Browser verification (#220 acceptance)

Run in Chromium against the dev server's real templates (8001, dev DB, bus off). Instead of shortening the poll interval, latency was injected in the browser, so each ordering happens every time rather than by chance.

Scenario	Before	After
Click mid-poll, poll lands first	POST render, toast, focus lost	Poll aborted; POST render, toast, heading focus
2.6s on the confirm, open command closes meanwhile	Stale and idle until reload	Correct, still polling
Confirm cancelled	–	Ticks continue, no re-arming
Link inside the section mid-poll	Navigates	Navigates

Review rounds

Two CR rounds, six findings, all in their own commits. One correction worth reading: the #220 comment called queue all "deferred and hidden". Measured, queue parks the second action on the wrapper, where the first action's swap wipes it; only when the first fails without swapping does the queued one fire, late, behind a second confirm. replace aborts the first action's request: the server commits it, but its render and toast are lost. drop stands.

Also in this PR

• 0085e96 chore: update skills submodules - the SessionStart auto-refresh bump, which was on local main when this branch was cut. It rides along as ec800b2 did in #213 docs: index REGISTER.md; bring BUS.md, PAGES.md and AGENTS.md under budget #221.
• UI.md's "Partial fragment swaps" is removed: it described HX-Reswap/HX-Retarget headers no first-party code sends, citing the per-row swap Dashboard replication state + streams table: retire the hand-edited public_url #171 CR API Keys - updates to workflow and list #37 retired.

No contract-visible paths change, so no CHANGELOG entry. After merge: sudo systemctl restart archiver.

🤖 Generated with Claude Code