Record being wrong about which change broke a test

A UI test went red in the same run that adopted three new APIs. The one
touching that screen was reverted; the test still failed. A modifier
order was corrected; still failed. Simulator data was wiped; still
failed. Running it against a clean checkout — which should have been the
first move, not the fourth — it failed there too, with none of the work
present.

The scrolling entry had already been rewritten to explain the fault in
confident detail. That write-up was fiction, so it now says what actually
happened: the API was convicted on circumstance and remains unproven
here, and the one claim that survives is about its two-way binding
costing invalidations.

xcode-practice gets the general lesson, since it will outlive this
particular test: stash and rerun before reasoning about which edit did
it, and a note recording a false cause is worse than no note, because
the next person believes it.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

.claude-plugin/marketplace.json

@@ -12,7 +12,7 @@
       "name": "pdat-swift-practice",
       "source": "./pdat-swift-practice",
       "description": "Curated Apple platform findings with sources, plus practices learned from production incidents.",
-      "version": "0.2.0",
+      "version": "0.3.0",
       "author": {
         "name": "PDAT s.r.o."
       },

pdat-swift-practice/.claude-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "pdat-swift-practice",
-  "version": "0.2.0",
+  "version": "0.3.0",
   "description": "Curated Apple platform findings with sources, plus practices learned from production incidents.",
   "author": {
     "name": "PDAT s.r.o.",

pdat-swift-practice/skills/apple-findings/references/scrolling.md

@@ -25,9 +25,29 @@ following the offset while the user drags.
 to reach `proxy.scrollTo` from a `.task`. In our card list that is the
 filter chip strip, which scrolls the active chip into view.
 
-**Step forward?** Modest but real. One layer of nesting disappears and the
+**Step forward? Unproven — we backed it out.** Swapping `ScrollViewReader`
-scroll position stops being write-only. Not urgent — the existing code
+for `.scrollPosition(id:)` coincided with the UI test for that strip
-works — so worth doing when that view is touched for another reason.
+failing, so it looked like the culprit and was reverted. It was not:
+the same test fails on a completely clean tree, with none of the change
+present. Something in the environment broke that test, and the swap was
+convicted on circumstance.
+
+The revert stands anyway, because it was never worth much here — this
+strip only ever *writes* a scroll target, which is exactly what the proxy
+already does. But the reasoning above is worth keeping, because being
+wrong about it taught the real lesson: **when a change and a failure
+arrive together, check the failure against an unchanged tree before
+concluding.** Two rounds of edits went into explaining a fault that was
+not there.
+
+One thing that does hold, independent of all that:
+
+- **`scrollPosition` is two-way, so it costs invalidations.** It writes to
+  its binding on every scroll, where the proxy only takes commands. It
+  therefore belongs in a small view. Ours would have sat in a screen whose
+  body derives a sorted, grouped card list — every drag of the strip would
+  have recomputed all of it. Extracting the strip into its own view was
+  worth keeping regardless of which scrolling API ends up inside it.
 
 **Available since.** iOS 17, so nothing blocks it at our deployment
 target. `onScrollGeometryChange` needs iOS 18.

pdat-swift-practice/skills/xcode-practice/SKILL.md

@@ -136,6 +136,27 @@ dead when it is passing. Also: do not pipe a test run through `tail` or a
 narrow `grep` and then diagnose from what survived — filter the saved log
 afterwards instead.
 
+## Check a failure against an unchanged tree before blaming the change
+
+A UI test started failing in the same run that introduced three API
+changes. The obvious suspect — the one touching that exact screen — was
+reverted, and the test still failed; a modifier order was corrected, still
+failed; the simulator's app data was wiped, still failed. Only then was
+the test run against a clean checkout with none of the work present, and
+it failed there too. The change had been innocent all along, and two
+rounds of edits went into explaining a fault that did not exist.
+
+`git stash` and one run is cheaper than any amount of reasoning about
+which edit did it. Do it first, not fourth. The same goes for what gets
+written down afterwards: a note recording a false cause is worse than no
+note, because the next person believes it.
+
+Corollary for UI tests: when one goes red without a code change, suspect
+what the test depends on but does not control — accumulated seed data,
+`@AppStorage` surviving between launches, a simulator location and the
+"nearby" sort order it drives, or an element scrolled out of a lazy
+container and therefore missing from the accessibility tree entirely.
+
 ## Core Data in tests
 
 Build objects through the context's own model rather than