diff --git a/src/renderer/src/components/cmd-j/palette-project-results.ts b/src/renderer/src/components/cmd-j/palette-project-results.ts
index e4b26441bde..d0fcc15636c 100644
--- a/src/renderer/src/components/cmd-j/palette-project-results.ts
+++ b/src/renderer/src/components/cmd-j/palette-project-results.ts
@@ -1,4 +1,9 @@
 import { isCmdJPaletteQueryTooLarge } from './palette-results'
+import {
+  cmdJPaletteTokenScore,
+  normalizeCmdJPaletteQuery,
+  uniqueNormalizedCmdJPaletteKeywords
+} from './palette-query-tokens'
 import type { Project, ProjectGroup, ProjectHostSetup, Repo } from '../../../../shared/types'
 import { translate } from '@/i18n/i18n'
 import {
@@ -39,73 +44,6 @@ type RankedProjectResult = {
 const PROJECT_GROUP_ALIASES = ['group', 'repo group']
 const PROJECT_ALIASES = ['project', 'repo']
 
-function normalizeQuery(value: string): string {
-  let normalized = ''
-  let pendingWhitespace = false
-  for (let index = 0; index < value.length; index += 1) {
-    const code = value.charCodeAt(index)
-    if (isCmdJPaletteWhitespace(code)) {
-      pendingWhitespace = normalized.length > 0
-      continue
-    }
-    if (pendingWhitespace) {
-      normalized += ' '
-      pendingWhitespace = false
-    }
-    normalized += value.charAt(index).toLowerCase()
-  }
-  return normalized
-}
-
-function isCmdJPaletteWhitespace(code: number): boolean {
-  return (
-    code === 32 ||
-    (code >= 9 && code <= 13) ||
-    code === 160 ||
-    code === 5760 ||
-    (code >= 8192 && code <= 8202) ||
-    code === 8232 ||
-    code === 8233 ||
-    code === 8239 ||
-    code === 8287 ||
-    code === 12288 ||
-    code === 65279
-  )
-}
-
-function uniqueNormalized(values: readonly string[]): string[] {
-  return [...new Set(values.map(normalizeQuery).filter(Boolean))]
-}
-
-function tokenize(value: string): string[] {
-  return normalizeQuery(value)
-    .split(/[^a-z0-9]+/)
-    .filter(Boolean)
-}
-
-function tokenScore(query: string, values: readonly string[]): number {
-  const candidateTokens = values.flatMap(tokenize)
-  if (candidateTokens.length === 0) {
-    return 0
-  }
-
-  let score = 0
-  for (const queryToken of tokenize(query)) {
-    let best = 0
-    for (const candidateToken of candidateTokens) {
-      if (candidateToken === queryToken) {
-        best = Math.max(best, 3)
-      } else if (candidateToken.startsWith(queryToken)) {
-        best = Math.max(best, 2)
-      } else if (candidateToken.includes(queryToken)) {
-        best = Math.max(best, 1)
-      }
-    }
-    score += best
-  }
-  return score
-}
-
 function buildCmdJProjectSearchCandidates({
   projectGroups,
   repos,
@@ -134,7 +72,7 @@ function buildCmdJProjectSearchCandidates({
       ),
       rowKey: getProjectGroupHeaderKey(group.id),
       order,
-      keywords: uniqueNormalized([group.name, ...PROJECT_GROUP_ALIASES])
+      keywords: uniqueNormalizedCmdJPaletteKeywords([group.name, ...PROJECT_GROUP_ALIASES])
     })
   })
 
@@ -156,7 +94,11 @@ function buildCmdJProjectSearchCandidates({
       rowKey: target.key,
       repo: target.repo,
       order: projectGroups.length + repoIndex,
-      keywords: uniqueNormalized([target.label, repo.displayName, ...PROJECT_ALIASES])
+      keywords: uniqueNormalizedCmdJPaletteKeywords([
+        target.label,
+        repo.displayName,
+        ...PROJECT_ALIASES
+      ])
     })
   })
 
@@ -191,7 +133,7 @@ function projectRankingForCandidate(
   query: string,
   candidate: CmdJProjectSearchResult
 ): RankedProjectResult | null {
-  const title = normalizeQuery(candidate.title)
+  const title = normalizeCmdJPaletteQuery(candidate.title)
   if (query === title) {
     return { result: candidate, rule: 1, score: 0 }
   }
@@ -199,13 +141,13 @@ function projectRankingForCandidate(
     return { result: candidate, rule: 2, score: 0 }
   }
   const aliasKeywords = candidate.kind === 'project-group' ? PROJECT_GROUP_ALIASES : PROJECT_ALIASES
-  if (aliasKeywords.map(normalizeQuery).includes(query)) {
+  if (aliasKeywords.map(normalizeCmdJPaletteQuery).includes(query)) {
     return { result: candidate, rule: 3, score: 0 }
   }
   if (candidate.keywords.some((keyword) => keyword.startsWith(query))) {
     return { result: candidate, rule: 4, score: 0 }
   }
-  const score = tokenScore(query, [candidate.title, ...candidate.keywords])
+  const score = cmdJPaletteTokenScore(query, [candidate.title, ...candidate.keywords])
   return score > 0 ? { result: candidate, rule: 5, score } : null
 }
 
@@ -242,7 +184,7 @@ export function searchCmdJProjectResults({
   if (isCmdJPaletteQueryTooLarge(query)) {
     return []
   }
-  const normalizedQuery = normalizeQuery(query)
+  const normalizedQuery = normalizeCmdJPaletteQuery(query)
   // Why: project/group rows sit after worktree matches, so one-character
   // searches would add broad noisy navigation targets before intent is clear.
   if (normalizedQuery.length < 2) {
diff --git a/src/renderer/src/components/cmd-j/palette-query-tokens.test.ts b/src/renderer/src/components/cmd-j/palette-query-tokens.test.ts
new file mode 100644
index 00000000000..3f829ace14a
--- /dev/null
+++ b/src/renderer/src/components/cmd-j/palette-query-tokens.test.ts
@@ -0,0 +1,15 @@
+import { describe, expect, it } from 'vitest'
+import { normalizeCmdJPaletteQuery, tokenizeCmdJPaletteQuery } from './palette-query-tokens'
+
+describe('Cmd+J palette query normalization', () => {
+  it('lowercases supplementary-plane characters as whole code points', () => {
+    // Why: code-unit iteration split '\u{10400}' into surrogate halves, which lowercase to
+    // themselves, so the query never matched its lowercase counterpart.
+    expect(normalizeCmdJPaletteQuery('\u{10400} HELLO')).toBe('\u{10428} hello')
+    expect(tokenizeCmdJPaletteQuery('\u{10400} HELLO')).toEqual(['\u{10428}', 'hello'])
+  })
+
+  it('collapses surrounding and repeated whitespace', () => {
+    expect(normalizeCmdJPaletteQuery('  New\n\tTerminal  ')).toBe('new terminal')
+  })
+})
diff --git a/src/renderer/src/components/cmd-j/palette-query-tokens.ts b/src/renderer/src/components/cmd-j/palette-query-tokens.ts
new file mode 100644
index 00000000000..3ba9068b0c1
--- /dev/null
+++ b/src/renderer/src/components/cmd-j/palette-query-tokens.ts
@@ -0,0 +1,116 @@
+// Why: the middle band and the project band scored queries with byte-identical copies of
+// this logic, so ranking fixes only ever landed in one of the two.
+
+function isCmdJPaletteWhitespace(code: number): boolean {
+  return (
+    code === 32 ||
+    (code >= 9 && code <= 13) ||
+    code === 160 ||
+    code === 5760 ||
+    (code >= 8192 && code <= 8202) ||
+    code === 8232 ||
+    code === 8233 ||
+    code === 8239 ||
+    code === 8287 ||
+    code === 12288 ||
+    code === 65279
+  )
+}
+
+// Why: iterating code units would lowercase surrogate halves separately, leaving
+// supplementary-plane characters uncased and unmatchable.
+export function normalizeCmdJPaletteQuery(value: string): string {
+  let normalized = ''
+  let pendingWhitespace = false
+  for (const character of value) {
+    if (isCmdJPaletteWhitespace(character.codePointAt(0) ?? 0)) {
+      pendingWhitespace = normalized.length > 0
+      continue
+    }
+    if (pendingWhitespace) {
+      normalized += ' '
+      pendingWhitespace = false
+    }
+    normalized += character.toLowerCase()
+  }
+  return normalized
+}
+
+export function uniqueNormalizedCmdJPaletteKeywords(values: readonly string[]): string[] {
+  return [...new Set(values.map(normalizeCmdJPaletteQuery).filter(Boolean))]
+}
+
+// Why: splitting on non-ASCII would drop CJK and accented words entirely, so a localized
+// query would silently skip the coverage rule below instead of failing it.
+export function tokenizeCmdJPaletteQuery(value: string): string[] {
+  return normalizeCmdJPaletteQuery(value)
+    .split(/[^\p{L}\p{N}]+/u)
+    .filter(Boolean)
+}
+
+// Why: navigation filler carries no intent ("open ssh settings" means "ssh settings"), so an
+// unmatched filler word must not count against a candidate's coverage.
+const CMD_J_QUERY_FILLER_TOKENS = new Set([
+  'a',
+  'an',
+  'and',
+  'change',
+  'edit',
+  'for',
+  'from',
+  'go',
+  'goto',
+  'in',
+  'into',
+  'jump',
+  'me',
+  'my',
+  'of',
+  'on',
+  'open',
+  'please',
+  'set',
+  'show',
+  'the',
+  'to',
+  'view',
+  'with'
+])
+
+export function cmdJPaletteTokenScore(query: string, values: readonly string[]): number {
+  const candidateTokens = values.flatMap(tokenizeCmdJPaletteQuery)
+  if (candidateTokens.length === 0) {
+    return 0
+  }
+
+  let score = 0
+  let meaningful = 0
+  let covered = 0
+  for (const queryToken of tokenizeCmdJPaletteQuery(query)) {
+    let best = 0
+    for (const candidateToken of candidateTokens) {
+      if (candidateToken === queryToken) {
+        best = Math.max(best, 3)
+      } else if (candidateToken.startsWith(queryToken)) {
+        best = Math.max(best, 2)
+      } else if (candidateToken.includes(queryToken)) {
+        best = Math.max(best, 1)
+      }
+    }
+    score += best
+    if (CMD_J_QUERY_FILLER_TOKENS.has(queryToken)) {
+      continue
+    }
+    meaningful += 1
+    if (best > 0) {
+      covered += 1
+    }
+  }
+
+  // Why: "linear triage" matched the Linear pane as strongly as "linear" did, so a candidate
+  // now has to cover most of what was typed, not just one word of it.
+  if (meaningful > 0 && covered * 2 <= meaningful) {
+    return 0
+  }
+  return score
+}
diff --git a/src/renderer/src/components/cmd-j/palette-results.test.ts b/src/renderer/src/components/cmd-j/palette-results.test.ts
index 73049bcc149..bf875ebe9a2 100644
--- a/src/renderer/src/components/cmd-j/palette-results.test.ts
+++ b/src/renderer/src/components/cmd-j/palette-results.test.ts
@@ -208,6 +208,66 @@ describe('Cmd+J palette middle-band ranking', () => {
     })
   })
 
+  it('drops candidates that cover a minority of the words typed', () => {
+    // Why: "linear triage" used to surface the Linear and Integrations panes on the
+    // "linear" token alone, with the unmatched word costing nothing. See screenshot report.
+    const integrationSections: SettingsNavSection[] = [
+      {
+        id: 'linear',
+        title: 'Linear',
+        description: 'How Linear works in Orca.',
+        icon: Settings,
+        searchEntries: [],
+        group: 'capabilities'
+      },
+      {
+        id: 'integrations',
+        title: 'Integrations',
+        description: 'Connect GitHub, GitLab, and Linear.',
+        icon: Settings,
+        searchEntries: [],
+        group: 'setup'
+      }
+    ]
+    const rank = (query: string): string[] =>
+      rankCmdJMiddleResults({
+        query,
+        settingsResults: buildCmdJSettingsResults(integrationSections),
+        actionResults: []
+      }).map((result) => result.id)
+
+    expect(rank('linear triage')).toEqual([])
+    expect(rank('linear')).toEqual(['settings:linear', 'settings:integrations'])
+    // Why: a majority still counts, so one stray word cannot blank an otherwise good match.
+    expect(rank('linear integrations triage')).toEqual(['settings:integrations'])
+  })
+
+  it('drops verb-prefixed settings queries whose middle words match nothing', () => {
+    // Why: the verb-plus-settings-keyword rule reads only the head and tail of the query, so
+    // "new terminal <junk> browser settings" used to win on those two ends alone.
+    expect(top('new terminal sparkles glitter browser settings')).toBeUndefined()
+    expect(top('new terminal settings')).toBe('settings:terminal')
+  })
+
+  it('keeps out-of-order partial-word matches when every query word lands somewhere', () => {
+    expect(top('font term')).toBe('settings:terminal')
+    expect(top('font sparkles')).toBeUndefined()
+  })
+
+  it('ignores navigation filler words when measuring coverage', () => {
+    // Why: "open"/"go to" carry no intent, so they must not read as unmatched words
+    // and blank the band mid-query.
+    expect(top('open terminal settings')).toBe('settings:terminal')
+    expect(top('go to ssh settings')).toBe('settings:ssh')
+    expect(top('change the terminal font')).toBe('settings:terminal')
+  })
+
+  it('counts non-latin query words instead of discarding them', () => {
+    // Why: tokenizing on ASCII only let a localized query skip the coverage rule entirely.
+    expect(top('ssh 主机 设置')).toBeUndefined()
+    expect(top('ssh 设置')).toBeUndefined()
+  })
+
   it('does not match settings on one-character or description-only queries', () => {
     expect(top('t')).toBeUndefined()
     expect(top('cookie import')).toBeUndefined()
@@ -316,6 +376,23 @@ function projectGroup(id: string, name: string, parentGroupId: string | null = n
 }
 
 describe('Cmd+J project and repo-group search', () => {
+  it('drops projects that only match the query through a generic alias word', () => {
+    // Why: every project carries the 'repo'/'project' aliases, so "repo triage" used to
+    // match all of them on that word alone — the same bug the middle band had.
+    const search = (query: string): string[] =>
+      searchCmdJProjectResults({
+        query,
+        projectGroups: [],
+        repos: [repo('repo-1', 'linear-sync'), repo('repo-2', 'billing')],
+        projects: [],
+        projectHostSetups: []
+      }).map((result) => result.title)
+
+    expect(search('repo triage')).toEqual([])
+    expect(search('linear triage')).toEqual([])
+    expect(search('linear repo')).toEqual(['linear-sync'])
+  })
+
   it('finds a Project Group by name', () => {
     const [result] = searchCmdJProjectResults({
       query: 'infra',
diff --git a/src/renderer/src/components/cmd-j/palette-results.ts b/src/renderer/src/components/cmd-j/palette-results.ts
index 9246599843c..3b643c1fe8e 100644
--- a/src/renderer/src/components/cmd-j/palette-results.ts
+++ b/src/renderer/src/components/cmd-j/palette-results.ts
@@ -1,6 +1,11 @@
 import type { SettingsNavIcon, SettingsNavSection } from '@/lib/settings-navigation-types'
 import type { CmdJQuickAction } from './quick-actions'
 import { isClipboardTextByteLengthOverLimit } from '../../../../shared/clipboard-text'
+import {
+  cmdJPaletteTokenScore,
+  normalizeCmdJPaletteQuery,
+  uniqueNormalizedCmdJPaletteKeywords
+} from './palette-query-tokens'
 
 export type CmdJSettingsResult = {
   id: string
@@ -53,40 +58,6 @@ export function isCmdJPaletteQueryTooLarge(
   return isClipboardTextByteLengthOverLimit(query, maxBytes)
 }
 
-function normalizeQuery(value: string): string {
-  let normalized = ''
-  let pendingWhitespace = false
-  for (let index = 0; index < value.length; index += 1) {
-    const code = value.charCodeAt(index)
-    if (isCmdJPaletteWhitespace(code)) {
-      pendingWhitespace = normalized.length > 0
-      continue
-    }
-    if (pendingWhitespace) {
-      normalized += ' '
-      pendingWhitespace = false
-    }
-    normalized += value.charAt(index).toLowerCase()
-  }
-  return normalized
-}
-
-function isCmdJPaletteWhitespace(code: number): boolean {
-  return (
-    code === 32 ||
-    (code >= 9 && code <= 13) ||
-    code === 160 ||
-    code === 5760 ||
-    (code >= 8192 && code <= 8202) ||
-    code === 8232 ||
-    code === 8233 ||
-    code === 8239 ||
-    code === 8287 ||
-    code === 12288 ||
-    code === 65279
-  )
-}
-
 function keywordParts(section: SettingsNavSection): string[] {
   const baseId = section.id.startsWith('repo-') ? 'repo' : section.id
   const idWords = baseId.replace(/-/g, ' ')
@@ -107,10 +78,6 @@ function targetEntryKeywordParts(entryTitle: string): string[] {
   return [entryTitle, `${entryTitle} settings`]
 }
 
-function uniqueNormalized(values: readonly string[]): string[] {
-  return [...new Set(values.map(normalizeQuery).filter(Boolean))]
-}
-
 export function buildCmdJSettingsResults(
   sections: readonly SettingsNavSection[]
 ): CmdJSettingsResult[] {
@@ -123,7 +90,7 @@ export function buildCmdJSettingsResults(
       icon: section.icon,
       sectionId: section.id,
       order,
-      configKeywords: uniqueNormalized(keywordParts(section))
+      configKeywords: uniqueNormalizedCmdJPaletteKeywords(keywordParts(section))
     }
     const targetedResults = section.searchEntries
       .filter((entry) => entry.targetSectionId)
@@ -136,7 +103,7 @@ export function buildCmdJSettingsResults(
         sectionId: section.id,
         targetSectionId: entry.targetSectionId,
         order: order + (entryIndex + 1) / 100,
-        configKeywords: uniqueNormalized([
+        configKeywords: uniqueNormalizedCmdJPaletteKeywords([
           ...targetEntryKeywordParts(entry.title),
           ...(entry.cmdJKeywords ?? entry.keywords ?? [])
         ])
@@ -154,35 +121,6 @@ function startsOrIsStartedBy(query: string, keyword: string): boolean {
   return keyword.startsWith(query) || query.startsWith(keyword)
 }
 
-function tokenize(value: string): string[] {
-  return normalizeQuery(value)
-    .split(/[^a-z0-9]+/)
-    .filter(Boolean)
-}
-
-function tokenScore(query: string, values: readonly string[]): number {
-  const candidateTokens = values.flatMap(tokenize)
-  if (candidateTokens.length === 0) {
-    return 0
-  }
-
-  let score = 0
-  for (const queryToken of tokenize(query)) {
-    let best = 0
-    for (const candidateToken of candidateTokens) {
-      if (candidateToken === queryToken) {
-        best = Math.max(best, 3)
-      } else if (candidateToken.startsWith(queryToken)) {
-        best = Math.max(best, 2)
-      } else if (candidateToken.includes(queryToken)) {
-        best = Math.max(best, 1)
-      }
-    }
-    score += best
-  }
-  return score
-}
-
 function rankingForCandidate(
   query: string,
   candidate: CmdJMiddleResult,
@@ -193,6 +131,17 @@ function rankingForCandidate(
     return null
   }
 
+  // Why: the shortcut rules below only inspect the head and tail of the query, so coverage has
+  // to gate all of them — otherwise "new terminal <junk> browser settings" still wins on rule 3.
+  const values =
+    candidate.kind === 'settings'
+      ? [candidate.title, ...candidate.configKeywords]
+      : [candidate.title, ...candidate.verbKeywords]
+  const score = cmdJPaletteTokenScore(query, values)
+  if (score === 0) {
+    return null
+  }
+
   if (candidate.kind === 'action' && candidate.verbKeywords.some((keyword) => query === keyword)) {
     return { result: candidate, rule: 1, score: 0 }
   }
@@ -227,12 +176,7 @@ function rankingForCandidate(
     return { result: candidate, rule: 5, score: 0 }
   }
 
-  const values =
-    candidate.kind === 'settings'
-      ? [candidate.title, ...candidate.configKeywords]
-      : [candidate.title, ...candidate.verbKeywords]
-  const score = tokenScore(query, values)
-  return score > 0 ? { result: candidate, rule: 6, score } : null
+  return { result: candidate, rule: 6, score }
 }
 
 function compareRanked(a: RankedResult, b: RankedResult): number {
@@ -263,7 +207,7 @@ export function rankCmdJMiddleResults({
   if (isCmdJPaletteQueryTooLarge(query)) {
     return []
   }
-  const normalizedQuery = normalizeQuery(query)
+  const normalizedQuery = normalizeCmdJPaletteQuery(query)
   if (normalizedQuery.length < 2) {
     return []
   }