diff --git a/.sync/PORTING.md b/.sync/PORTING.md
index 4f3f4a50..804dda02 100644
--- a/.sync/PORTING.md
+++ b/.sync/PORTING.md
@@ -177,13 +177,16 @@ material. Reproduce its *intent* in b24ui by editing files under `src/` only.
   re-inspected here** — treat "upstream has no such parameter" as an inference
   from b24ui's own history rather than a verified fact, and re-check before
   acting on it. What is certain is that replaying upstream's four-parameter
-  signature wholesale drops behaviour shipped in v2.8.0 while **nothing goes
-  red**: `useTokenSearch` appears nowhere under `test/`. (Reverting the signature
-  alone does fail `typecheck`, since `CommandPalette.vue` passes five arguments;
-  it is a port that replays both files that goes quiet.) Several §2 rules here
-  name a guarding spec and some — locales, playground-manifest mirroring — do
-  not; this one does not, and #363 tracks the gap. Port the intent, keep the
-  fifth argument, and prefer adding that coverage over trusting this paragraph.
+  signature wholesale drops behaviour shipped in v2.8.0. Until #369 that
+  happened while **nothing went red** — `useTokenSearch` appeared nowhere under
+  `test/`, and reverting the signature alone only failed `typecheck`, since
+  `CommandPalette.vue` passes five arguments; a port that replayed both files
+  went quiet. Now guarded by `describe('useTokenSearch')` in
+  `test/utils/search.spec.ts`, which fails against a revert to upstream's
+  four-parameter signature, and by `describe('highlighting through fuse')` in
+  `test/components/CommandPalette.spec.ts`, which fails if the fifth argument
+  stops being forwarded from `processGroupItems`. Port the intent and keep the
+  fifth argument; trust those two specs over this paragraph.
 - **`skills/` is b24ui-authored — never replay upstream skill or doc prose into
   it.** The package was seeded from nuxt/ui's skill, and every defect the #93
   audit found was an inherited upstream idiom rather than an ordinary typo:
@@ -358,3 +361,4 @@ forward, since every commit between the two would then never be judged.
 - 2026-08-11 — review of PR #347 (issue #339): added the §2 **`highlight()` takes a fifth `useTokenSearch` argument** invariant, and corrected the `.sync/nuxt-ui.json` summary for `2a172ef` that asserted "highlight signature matches 1:1". The divergence has been in the tree since v2.8.0 and was never recorded: `c502157b` added `tokens`/`minTokenLength` and `6743f793` the parameter itself, both to `src/runtime/utils/fuse.ts`, and the port in `557a5178` renamed the file to `search.ts` — so a pickaxe on the current path returns only `557a5178`, a genuine upstream port, unless you pass `--follow`. That rename, not the trailer convention, is what hid it; `Upstream:` trailers are too rare (16 of ~3200 commits) to carry an inference either way. It has no test coverage; #363 tracks that. Worth recording how the error was found: the "byte-identical with upstream" premise originated **here**, in `595923b9` (PR #338), was repeated in #339, and was inherited in good faith by the external contributor whose PR prompted the check — nuxt/ui itself has still not been inspected, so the divergence is established from b24ui's history alone. Last reviewed: 2026-08-11.
 - 2026-08-12 — the sync is manual by decision; the automation is removed. Deleted `.sync/PLAN.md` (the dispatcher/porter/on-merge design, its phase plan and its cron) and `.sync/RUNBOOK.md` (an incident playbook whose every row diagnosed one of those workflows). Dropped `sync_enabled` from the ledger — a kill-switch for a dispatcher that will not exist reads as "the sync is off" to anyone who finds it, which was already misleading while this file's own procedure ran twelve ports past it — and `stats`, Phase-4 telemetry that was never written to (`noop_ratio: 0` against an actual 47/226). Folded the one runbook row that survives manual work into §6: a cursor SHA that vanishes under an upstream force-push must be moved to the nearest surviving ancestor with a tracking issue, never skipped forward. §6 now spells out the procedure that was previously only implied by the workflows — parent-order reconstruction, verbatim diffs, the gate order with `docs:generate` and `deploy.yml`'s env, ledger reconciliation including the last-entry case, and the `behind` rebase. Also corrected `color-map.json`: `warning` mapped to `air-primary-alert`, the same token as `error`, so the table said the two upstream colors were interchangeable; `air-primary-warning` exists and is used 50 times in `src/theme/`. Last reviewed: 2026-08-12.
 - 2026-08-12 — rebuilt `icon-map.json` and gave it a guard (the content of the closed PR #67, verified rather than imported). The map is now *derived*: for every icon key both sides define — `src/theme/icons.ts` upstream, `src/runtime/dictionary/icons.ts` here — the row is (upstream's lucide name → whatever our dictionary maps that key to), 37 pairs from a 43×39 key intersection at cursor `3dbca02`. Beware the obvious shortcut when re-checking this: the installed `@nuxt/ui@4.8.2` in `node_modules` (pulled in transitively by `nuxtseo-layer-devtools`) is **older than the sync cursor** and is missing keys — three separate reviewers read it and concluded `star` was fabricated and the intersection was 36. Read the raw file at the cursor SHA instead. The derivation turned up three errors in the values #67 proposed, each of which resolves to a real icon and so would have failed no import: `i-lucide-rotate-cw` for what upstream calls `i-lucide-rotate-ccw` (`reload`), `i-lucide-circle-check` for `copyCheck`'s `i-lucide-copy-check`, and `i-lucide-refresh-cw`, which no upstream key uses. It also surfaced seven derivable pairs #67 missed — `drag`, `panelClose`, `panelOpen`, `star`, `stop`, `copyCheck`, `reload` — and, separately, `i-lucide-terminal`, the **only** `i-lucide-*` literal upstream hardcodes under `src/` (`src/theme/prose/code-icon.ts`), which neither the old map nor #67 had even though `prose/CodeIcon.vue` has answered it all along. `error` and `success` gained judgement rows rather than staying unmapped: our `caution` carries a `// this for error` comment, and `copyCheck` already owns the glyph `success` would want. The five entries #67 dropped (`activity`, `arrow-up-to-line`, `house`, `settings`, `user`) are kept — they match no key on either side, which is the hardcoded-literal case the map exists for. **Correcting the record on the five values #67 changed** (`check`, `chevronDown`, `chevronUp`, `minus`, `x`): they are wrong because the map must agree with the dictionary, *not* — as an earlier draft of this entry claimed — because the library never renders them. It does. `Checkbox.vue` renders `main/CheckIcon` and `actions/Minus20Icon`, `Badge.vue` renders `actions/Cross20Icon`, `Button.vue` renders `outline/ChevronDownSIcon`; roughly half of the icon paths under `src/` are hardcoded in components that never read the dictionary, which is #380. That discovery also reshaped the guard: `test/utils/icon-map.spec.ts` allows any icon used anywhere in `src/` rather than only the dictionary's — the narrower rule rejected `terminal`, a correct row — while separately requiring every *derived* row to equal what its semantic key resolves to. That last check is the one with teeth: without it, pointing `i-lucide-check` at another icon the dictionary genuinely uses passed every other assertion. It guards wrong rows, not stale ones; nothing here notices if upstream renames a default. No `.sync/log/` or ledger entry, since this is not a port of an upstream commit — same as #343, #346, #351 and #377. Last reviewed: 2026-08-12.
+- 2026-08-12 — coverage for #363: gave the §2 **`highlight()` takes a fifth `useTokenSearch` argument** invariant a guard. It was recorded during the review of #347 but left untested, and the bullet said so. The parameter and the token-search logic around it are b24ui-only — `c502157b` added `tokens`/`minTokenLength` and `6743f793` the parameter itself, both against the file's old name `src/runtime/utils/fuse.ts`, both shipped in v2.8.0. Immediately before `c502157b` the function took four parameters and computed no `minTokenLength` at all, which is what marks it as locally authored; the `Upstream:` trailer convention is too sparse (16 of ~3200 commits) to carry an inference either way. The later port `557a5178` renamed `fuse.ts` to `search.ts` and carried the divergence across, so a pickaxe on the current path returns only that port — pass `--follow` to see the two commits that introduced it. Until now it had no test at all, so replaying upstream's four-parameter signature would have dropped a shipped feature with nothing going red. Upstream itself has not been re-inspected; treat "upstream has no such parameter" as an inference from b24ui's own history. Last reviewed: 2026-08-12.
diff --git a/test/components/CommandPalette.spec.ts b/test/components/CommandPalette.spec.ts
index 720e9846..c34c2eb7 100644
--- a/test/components/CommandPalette.spec.ts
+++ b/test/components/CommandPalette.spec.ts
@@ -317,4 +317,43 @@ describe('CommandPalette', () => {
       window.HTMLElement.prototype.scrollIntoView = original
     }
   })
+
+  describe('highlighting through fuse', () => {
+    // Everything above sets `labelHtml`/`suffixHtml`/`descriptionHtml` by hand,
+    // which is exactly the path that skips `highlight()`. These mount the real
+    // wiring instead: `processGroupItems` forwards
+    // `fuse.fuseOptions.useTokenSearch` as the fifth argument, and Fuse only
+    // populates `matches` when `includeMatches` is set — which this component
+    // leaves to the caller, so no other case in this file reaches `highlight()`
+    // at all.
+    function paletteWith(label: string, searchTerm: string, useTokenSearch?: boolean, threshold = 0.1) {
+      return mountSuspended(CommandPalette, {
+        props: {
+          groups: [{ id: 'g', items: [{ label }] }],
+          searchTerm,
+          fuse: { fuseOptions: { includeMatches: true, useTokenSearch, threshold, ignoreLocation: true } }
+        } as any
+      })
+    }
+
+    it('marks the match', async () => {
+      const wrapper = await paletteWith('alpha beta', 'alpha')
+
+      expect(wrapper.find('mark').exists()).toBe(true)
+      expect(wrapper.find('mark').text()).toContain('alpha')
+    })
+
+    it('marks each word of a multi-word query only when token search is on', async () => {
+      // Fuse returns the two words as separate regions, six and four characters
+      // long. Off, the threshold is the whole query — ten — and neither reaches
+      // it, so nothing is marked; on, it drops to the shortest word and both do.
+      // An exact-match fixture would not show the difference: Fuse then returns
+      // one region as long as the query, which clears either threshold.
+      const off = await paletteWith('alpha xxxx beta', 'alpha beta', false, 0.4)
+      const on = await paletteWith('alpha xxxx beta', 'alpha beta', true, 0.4)
+
+      expect(off.find('mark').exists()).toBe(false)
+      expect(on.find('mark').exists()).toBe(true)
+    })
+  })
 })
diff --git a/test/utils/search.spec.ts b/test/utils/search.spec.ts
index ffc1503f..02266f14 100644
--- a/test/utils/search.spec.ts
+++ b/test/utils/search.spec.ts
@@ -61,6 +61,52 @@ describe('highlight', () => {
     expect(highlight({ label: 'foo' }, 'foo', 'label')).toBeUndefined()
   })
 
+  describe('useTokenSearch', () => {
+    // b24ui-only fifth argument (see `.sync/PORTING.md` §2). It only moves
+    // `minTokenLength`, the threshold a region must reach to be marked at all:
+    // off, that is the length of the whole search term; on, the length of the
+    // shortest word in it — so multi-word queries highlight each word instead of
+    // nothing.
+    function highlightWith(value: string, indices: [number, number][], searchTerm: string, useTokenSearch?: boolean) {
+      return highlight({ label: value, matches: [{ key: 'label', value, indices }] }, searchTerm, 'label', undefined, useTokenSearch)
+    }
+
+    it('marks a region shorter than the whole term', () => {
+      // The region is 5 long. `alpha beta` is 10, so the full-term threshold
+      // rejects it; the shortest token, `beta`, is 4, so token search accepts it.
+      expect(highlightWith('alpha beta', [[0, 4]], 'alpha beta', true)).toBe('<mark>alpha</mark> beta')
+    })
+
+    it('leaves that same region unmarked when off', () => {
+      expect(highlightWith('alpha beta', [[0, 4]], 'alpha beta', false)).toBe('alpha beta')
+    })
+
+    it('is off by default — the path `CommandPalette` takes unless `fuseOptions` opts in', () => {
+      expect(highlightWith('alpha beta', [[0, 4]], 'alpha beta')).toBe('alpha beta')
+    })
+
+    it('still refuses a region shorter than the shortest token', () => {
+      // `be` is 2; the shortest token `beta` is 4. Token search lowers the bar,
+      // it does not remove it.
+      expect(highlightWith('alpha beta', [[6, 7]], 'alpha beta', true)).toBe('alpha beta')
+    })
+
+    it('falls back to the whole term when the tokenizer matches nothing', () => {
+      // `/[\p{L}\p{M}\p{N}_]+/gu` finds no token in an emoji-only query, so
+      // `tokens` is empty and the threshold stays `searchTerm.length` — 2 code
+      // units for one emoji — rather than collapsing to zero and marking
+      // everything.
+      expect(highlightWith('\u{1F600}\u{1F600}\u{1F600}', [[0, 1]], '\u{1F600}', true))
+        .toBe('<mark>\u{1F600}</mark>\u{1F600}\u{1F600}')
+    })
+
+    it('takes the shortest token, not the first or the longest', () => {
+      // Tokens are `considerable` (12) and `ab` (2); only a 2-long threshold
+      // admits this region.
+      expect(highlightWith('xy considerable', [[0, 1]], 'considerable ab', true)).toBe('<mark>xy</mark> considerable')
+    })
+  })
+
   describe('truncation from the start', () => {
     // Matches a high surrogate not followed by a low one, or a low surrogate not
     // preceded by a high one — i.e. half of an astral character.