Fill remaining AGENTS.md gaps in flutter, web, and store-locator skills

[mattpodwysocki]

Fourth pass over the drop-in AGENTS.md files (after #83 and #84), covering the milder gaps that survived measurement at --repeats=5.

mapbox-web-integration-patterns — the silent one

import 'mapbox-gl/dist/mapbox-gl.css' appears 11 times across its references and zero times in AGENTS.md. Omit it and the map still renders, nothing throws, no console error — controls, popups and markers just lose their styles. Added as rule 0 of Critical Integration Rules since it applies to every framework in the file.

mapbox-flutter-patterns

• manager.tapEvents(...) was shown without the Cancelable it returns, so the listener outlives the widget. Added the stored subscription and .cancel() in dispose, plus a note that longPressEvents, dragEvents and the other annotation managers follow the same pattern.
• Flagged the iOS deployment target as the single most common iOS build failure (platform is lower than deployment target), and that CocoaPods vs SPM is not a decision the user has to make.

mapbox-store-locator-patterns

highlightListing was missing, so selecting a marker highlighted a sidebar listing that could be off-screen. Added it with the scrollIntoView call from SKILL.md — and gave the listing element the listing-${id} id the lookup depends on, which the existing snippet never set.

mapbox-maplibre-migration

The note read "Most Mapbox plugins work directly, no alternatives needed" — reassurance, when the point is the opposite. Every @maplibre/* package generally needs a Mapbox counterpart and the table is not exhaustive.

Measured

Agents surface, n=5 per eval:

eval	before	after
flutter #3	57%	97%
flutter #1	79%	96%
web-integration #1	80%	100% (sd 0.000)
store-locator #1	70%	88%
store-locator #2	85%	97%

Two honest caveats. maplibre #3 moved 75% → 82%, but its spread widened to sd 0.171, so that gain is not cleanly attributable. And store-locator #3 reads 93% → 83% — it covers GeolocateControl and geolocation, code this commit does not touch, and the run ranges overlap (83–100 before, 75–100 after), so it is noise rather than a regression.

Deliberately excluded: web-performance #2

It scores 80% on both surfaces, so it is not AGENTS.md drift. Its expectation states that 75,000 points do not require clustering because "clustering is for 100,000+", while SKILL.md:175 says clustering is recommended from 10,000+. The eval and the skill contradict each other, and one of them is wrong. That needs a domain decision — not an edit to make a number move.

Ordering

Branches from main; independent of #80, #83 and #84. The --surface and --repeats flags used to measure this live in #80.

🤖 Generated with Claude Code