CR round 3 on #212: the detail page never started polling

[gregoryfoster]

Round-3 review of 88d3dfd..a8ab532 (the #212 work). Six findings applied, one held.

#	Finding	Action
12	🔴 The InfoItem detail page never passed poll — a full page load with an open command was inert	Fixed 508711b
13	{% if poll is defined %} turned the missing key into silent non-polling	Fixed 620a54c
14	Both partials' contract comments omitted poll / poll_interval_seconds	Fixed 5f63556
15	No Cache-Control on a fragment fetched every 2s	Fixed 66723c6
16	The stalled notice was duplicated verbatim across two templates	Fixed df85c72
17	Mapping[object, ReplicationCommand] where keys are ULIDs	Fixed 1aea43b
18	Poll/click swap race	Held — the fix is speculative

Finding 12: #212's own defect survived on the most ordinary path

The section reaches a browser by four routes, not three: the deactivate swap, the Replicate now swap, the poll — and the detail page itself. Only the three swaps went through the renderer that supplies poll, so a full page load rendered the section with no polling context at all.

Click Replicate now, press F5 before it settles, and the outcome is back to requiring a manual reload. Arriving at the item from the list while a command is in flight behaves the same. The RepSpec twin routed all four of its sites from the start, so the two screens disagreed — and the broken one is the screen the operator actually used.

Two tests pin it: the page polls with a command open, and does not with nothing open.

Finding 13 is why it shipped green

{% if poll is defined and poll.active %} — under Jinja's default undefined, a missing poll is falsy. Every test exercised a path that passed the key, and the one path without it degraded silently instead of failing. A defensive guard that converts a missing contract into the original defect is worse than no guard.

Finding 18, held

A poll firing while a Replicate now POST is in flight gives two swaps racing for one target; a poll response landing after the POST's would briefly show pre-issuance state. It self-corrects within one interval and both renders are server-authoritative, so the impact is a flicker. hx-sync is the obvious lever, but which element declares it and which strategy gives the right precedence — the poll must lose to the click, never the reverse — needs verifying in a browser rather than reasoning about. Not applied under a blanket directive.

Verification

ruff check and ruff format --check clean over 440 files; 1791 passed, 2 skipped, 3 deselected. Five net-new tests.

🤖 Generated with Claude Code