fix(agent): recall_lessons tokenizer keeps single CJK runes

cloudAgentLessonSearchTokens dropped any token shorter than 2 runes.
That filter was copied from English stop-word logic (a / I), but a single
Han rune such as the ones meaning hook, gag or twist is already a complete
semantic unit in Chinese. When the whole keyword consisted of such runes,
tokens came back empty and cloudAgentSearchLessons silently fell back to
"list the first N lessons" - the user thinks a search happened but gets an
arbitrary list.

Fix: keep a token when it is exactly one Han rune (unicode.Han); single
ASCII/digit runes and empty strings are still dropped as before.

Test: TestLessonSearchTokensKeepsSingleCJKRune fails on the old logic
(got [反转 钩子], want [反转 钩子 梗 钩]) and passes after the fix.
The isomorphic copy in skill_search (cloud_agent_tools.go) is fixed
separately on the #580 branch and deliberately not included here.

Gates: go build ./... and go vet EXIT=0; gofmt clean on LF content; the 3
go-sqlite3 cgo-stub failures in internal/app are pre-existing and unrelated.

backend/internal/app/cloud_agent_lessons.go

@@ -545,7 +545,12 @@ func cloudAgentLessonSearchTokens(keyword string) []string {
 	seen := map[string]bool{}
 	for _, raw := range strings.FieldsFunc(keyword, cloudAgentLessonTokenSeparator) {
 		token := strings.ToLower(strings.TrimSpace(raw))
-		if utf8.RuneCountInString(token) < 2 || seen[token] {
+		// 单字停用词过滤照搬的是英文逻辑（a / I）；汉字单字「梗」「钩」「戏」本身是完整
+		// 语义的最小单位。一刀切丢掉后 tokens 为空，检索会静默回落为「列前 N 条」。
+		// 这里只放行单个汉字的 token，英文/数字单字与空串仍按停用词丢掉。
+		first, size := utf8.DecodeRuneInString(token)
+		singleHan := size == len(token) && unicode.Is(unicode.Han, first)
+		if (!singleHan && utf8.RuneCountInString(token) < 2) || seen[token] {
 			continue
 		}
 		seen[token] = true