sync-skill: resolve the locale directory, and guard seams against upstream test fixtures

.claude/skills/sync-upstream/references/file-ownership.md

@@ -61,7 +61,7 @@ on every seam and feature path, either auto-resolving disjoint hunks or leaving
 Open each listed path and check it by hand against the manifest's declared `lines` for that path —
 those lines are the protected floor, not the whole file — before continuing.
 
-Upstream owns every key it defines, so the manifest leaves `src/renderer/src/locales/*.json`
+Upstream owns every key it defines, so the manifest leaves `src/renderer/src/i18n/locales/*.json`
 unclaimed and they reset to the tag through `checkout.txt` like any other upstream file. The fork's
 own keys live in per-feature bundles under the feature directories, which a feature glob claims. Keep
 that split: a fork entry duplicating a key upstream defines shadows upstream's real translation with
@@ -74,10 +74,19 @@ now fails `verify:localization-catalog` with
 day before had only en/es/ja/ko/zh. Sweep after resolution:
 
 ```sh
-comm -23 <(ls src/renderer/src/locales/*.json | xargs -n1 basename | sort) \
+comm -23 <(ls src/renderer/src/i18n/locales/*.json | xargs -n1 basename | sort) \
          <(ls <fork bundle dir> | sort)
 ```
 
+Resolve the upstream locale directory rather than pasting the path: it has moved once already, and
+a sweep pointed at a directory that no longer exists finds zero locales and reports every fork
+bundle complete. `ls` at least errors; a glob in Python or a `find` returns an empty set silently,
+which reads exactly like a pass. Confirm the set is non-empty before trusting the comparison.
+
+```sh
+dirname "$(git ls-files 'src/renderer/**/locales/en.json' | grep -v fork- )"
+```
+
 `pnpm sync:localization-catalog` will **not** fix this — `--fix` only repairs catalog registration
 and leaves the missing file missing, reporting the same error it was run to clear. Write it by hand:
 the file is exactly `{}` and a newline, matching every existing fork bundle, and
@@ -277,6 +286,31 @@ fork — the fork's direct call skipped a queued mock response and the next cons
 frame. And the fork's own tests mock whatever the fork used to call, so they have to move to the new
 module too.
 
+## When a seam meets an upstream test double
+
+Upstream's own suites drive every seam, and their fixtures are partial doubles built for upstream's
+code — they supply only what upstream's path reads. A seam that dereferences anything the fixture
+does not provide throws inside upstream's test, so one fork line fails suites the fork has no stake
+in, in files it does not own.
+
+v1.4.206 added a pending-close fixture with no `settingsRef`. The terminal-dock seam read
+`deps.settingsRef.current?.experimentalTerminalDock` — optional on `.current`, but not on the ref
+itself — and three pending-close suites threw `Cannot read properties of undefined (reading
+'current')` across three separate shards.
+
+The tell is a `TypeError` whose stack runs from an upstream test, through an upstream fixture, into
+one fork line in an upstream file. That shape is never a merge-resolution question: the seam is
+present and correct, and `--verify-seams` passes. Read the fixture before reading the seam.
+
+Write every seam so it degrades to its feature being off:
+
+- optional-chain the whole access path, not just its last hop
+- treat absent state as the feature disabled, never as a reason to throw
+- open durable state lazily, so merely importing the module in a test does not touch a store
+
+This is not the fixture's bug to fix. Upstream's fixture is upstream-owned and resets to the tag
+every sync, so hardening it there is undone at the next release; the guard belongs in the seam.
+
 ## Verifying
 
 Two manifest checks run against the new release, and the second one is where a sync goes quietly