diff --git a/learner/playwright.config.ts b/learner/playwright.config.ts
index c83524b..9c0468d 100644
--- a/learner/playwright.config.ts
+++ b/learner/playwright.config.ts
@@ -8,6 +8,9 @@ export default defineConfig({
   forbidOnly: !!process.env.CI,
   retries: 0,
   reporter: 'list',
+  // Every spec loads the whole SPA from the shared dev server, so a slow first module graph
+  // under parallel workers must not fail an assertion that the app itself would pass.
+  expect: { timeout: 15000 },
   use: { baseURL: 'http://127.0.0.1:5173', headless: true, trace: 'off' },
   projects: [{ name: 'chromium', use: { ...devices['Desktop Chrome'], channel: 'chrome' } }],
   webServer: { command: 'npm run dev -- --host 127.0.0.1', url: 'http://127.0.0.1:5173', reuseExistingServer: !process.env.CI },
diff --git a/learner/src/__tests__/phrase.spec.ts b/learner/src/__tests__/phrase.spec.ts
index 7849a9b..c063922 100644
--- a/learner/src/__tests__/phrase.spec.ts
+++ b/learner/src/__tests__/phrase.spec.ts
@@ -131,6 +131,21 @@ describe('phrase selection and panel', () => {
     expect(view.get('[data-testid="phrase-range"]').text()).toContain('已保存')
   })
 
+  it('stops claiming a stored phrase once its range is adjusted', async () => {
+    const span: PhraseSpan = { id: 7, status: 'new', wordCount: 3, startToken: 2, endToken: 7 }
+    const { view } = await open({ phrases: [span] })
+    await view.findAll('.reader-word').find(item => item.text() === 'small')!.trigger('click')
+    await flushPromises()
+    expect(view.get('[data-testid="phrase-range"]').text()).toContain('已保存')
+    await view.get('[data-testid="range-end-right"]').trigger('click')
+    await flushPromises()
+    // The adjusted range is another identity, so the panel no longer says it is stored, while
+    // the text the learner already had stays in the form.
+    expect(view.get('.lookup-panel').text()).toContain('新词条')
+    expect(view.get('.lookup-panel').text()).not.toContain('已保存')
+    expect((view.get('#term-definition').element as HTMLTextAreaElement).value).toBe('一小步')
+  })
+
   it('refuses a range longer than the phrase limit and keeps the panel closed', async () => {
     // One word more than a phrase may hold.
     const longTokens: ReaderToken[] = []
diff --git a/learner/src/composables/useReaderLookup.ts b/learner/src/composables/useReaderLookup.ts
index 659abcd..84c8530 100644
--- a/learner/src/composables/useReaderLookup.ts
+++ b/learner/src/composables/useReaderLookup.ts
@@ -289,7 +289,14 @@ export function useReaderLookup(chapter: Ref<ChapterDetail | null>) {
     const currentChapter = chapter.value
     if (!current || !currentChapter) return
     const next = adjustTokenRange(tokens.value, currentChapter.originalText ?? '', current, edge, direction)
-    if (next) range.value = next
+    if (!next) return
+    // The adjusted range is a different identity, so the panel must stop claiming that the
+    // stored entry is what is on screen: both the phrase link and the entry the form was loaded
+    // from are dropped. The typed text stays, and saving it writes the phrase that matches the
+    // current range (which may update another entry, never two).
+    rangeTermId.value = null
+    savedTermId.value = null
+    range.value = next
   }
 
   /** True when the panel is editing an entry that is already stored. */
diff --git a/server/app/lexgo/phrase_test.go b/server/app/lexgo/phrase_test.go
index bae84bd..11a78c5 100644
--- a/server/app/lexgo/phrase_test.go
+++ b/server/app/lexgo/phrase_test.go
@@ -118,6 +118,35 @@ func TestPhraseIdentityRules(t *testing.T) {
 	}
 }
 
+// TestWordKeysNeverContainSpaces locks the premise the derived kind relies on: a word token
+// never contains a space, so the identity key alone tells a word from a phrase and no column is
+// needed for it.
+func TestWordKeysNeverContainSpaces(t *testing.T) {
+	samples := []string{phraseFixtureText, "a  b\tc\r\nd", "don’t stop-BELIEVING.", "🙂 café e\u0301 42"}
+	checked := 0
+	for _, sample := range samples {
+		for _, token := range Tokenize(sample) {
+			if token.Kind != "word" {
+				continue
+			}
+			checked++
+			if strings.ContainsAny(normalizeWord(token.Text), " \t\r\n") {
+				t.Fatalf("word key %q contains whitespace", normalizeWord(token.Text))
+			}
+		}
+	}
+	if checked < 8 {
+		t.Fatalf("expected several word tokens, checked %d", checked)
+	}
+	// The derived kind therefore agrees with the shape of the key in both directions.
+	if termKind("a small step") != termKindPhrase || termWordCount("a small step") != 3 {
+		t.Fatal("a key with spaces must be a three-word phrase")
+	}
+	if termKind("curiosity") != termKindWord || termWordCount("curiosity") != 1 {
+		t.Fatal("a key without spaces must be a word")
+	}
+}
+
 func mustWords(t *testing.T, text string) []TextToken {
 	t.Helper()
 	words, err := phraseWords(phraseTokens(text), 0, len([]rune(text)))