fix(syb): 空格分隔的颜色尺码按逗号同规则拆分 (#284)

#274 删除 ambiguousPattern 以接受描述性颜色（黑色+白色 簡約親膚），
意图正确，但副作用是「黑色 XL」也被整条当成颜色、尺码丢失且判为
success。success 的行不进 AI 解析队列（ai_parse_batch.go 只选
uncertain/failed），因此没有任何机制会纠正它们。2026-09-15 线上有
152 条 parse_status='success' 且 target_size='' 的明细。

不恢复被删的旧闸——那会退回 #274 之前重新拒绝描述性颜色。改为把
空格当作与逗号同等的分隔符，复用逗号分支已有的规则：恰好一个空格
分隔 token 带明确尺码特征时才拆分，不构成猜测。

- 多于一个 token 命中尺码特征时不拆，对齐逗号分支「两侧均像尺码则
  无法安全识别颜色」。
- 没有尺码 token 时原样保留为颜色，#274 的意图不受影响。
- 一并修正 parse.go 中停留在旧 uncertain 描述的文档注释。

新增 parse_whitespace_split_test.go 锁住完整对照表，含 #274 描述性
颜色不被退回的回归。

`[必须]` app/goauto/sybimport 存在 12 个先于本工单的失败用例，来自
#272/#274 改变契约后未更新的旧断言，已记录为 #285。本次提交前后失败
集合逐条比对完全一致，未新增。

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NTDbDcwbDw1TSAcE6wfh2F

server/app/goauto/sybimport/parse.go

@@ -20,11 +20,46 @@ var bracketPattern = regexp.MustCompile(`【[^】]*】`)
 // deliberately narrow pattern.
 var explicitSizePattern = regexp.MustCompile(`(?i)^(?:均(?:码|碼|号|號)|one\s*size|free\s*size|x{0,4}[sml]|[2-9]xl|(?:加大|大|中|小)(?:码|碼|号|號)|\d+(?:\.\d+)?(?:cm|mm|m|码|碼|号|號|公分)|\d+(?:\.\d+)?(?:[-~～至到]\d+(?:\.\d+)?)?(?:斤|公斤|千克|kg))$`)
 
-// ambiguousPattern flags leftover separators or multi-token noise after
-// bracket stripping — the signal that a "clean" split still isn't reliable.
-// Observed in the real SYB sample: "黑色+白色【純棉兩件裝】 簡約親膚" strips
-// its bracket to "黑色+白色 簡約親膚", which still carries a '+' and internal
-// whitespace, so it must not be reported as a confident match.
+// splitOnWhitespace applies the comma rule to a spec that has no comma.
+//
+// `[必须]` #274 removed the old ambiguity guard so that descriptive colors
+// like "黑色+白色 簡約親膚" are accepted instead of rejected. That is right, but
+// it also let "黑色 XL" through as one confident **color**, dropping the size and
+// marking the row success — and success rows never reach the AI parse queue
+// (ai_parse_batch.go selects parse_status IN ('uncertain','failed')), so nothing
+// ever corrects them. 152 such rows existed in production on 2026-09-15.
+//
+// Whitespace is therefore treated exactly like the comma: split only when
+// **exactly one** whitespace-delimited token carries an explicit size signal.
+// That is the same non-guessing rule the comma branch already uses, so
+// "黑色 XL" recovers as color+size while "黑色+白色 簡約親膚" (no size token)
+// keeps #274's descriptive-color behaviour untouched.
+func splitOnWhitespace(value string) (color string, size string, ok bool) {
+	fields := strings.Fields(value)
+	if len(fields) < 2 {
+		return "", "", false
+	}
+	sizeIndex := -1
+	for i, field := range fields {
+		if explicitSizePattern.MatchString(field) {
+			if sizeIndex >= 0 {
+				// More than one size-looking token: no safe colour decision,
+				// same as the comma branch's both-sides-are-size case.
+				return "", "", false
+			}
+			sizeIndex = i
+		}
+	}
+	if sizeIndex < 0 {
+		return "", "", false
+	}
+	remainder := append(append([]string{}, fields[:sizeIndex]...), fields[sizeIndex+1:]...)
+	color = strings.Join(remainder, " ")
+	if color == "" {
+		return "", "", false
+	}
+	return color, fields[sizeIndex], true
+}
 
 // ParseResult is the color/size candidate extracted from one productSpec
 // string, plus how much the caller should trust it.
@@ -43,10 +78,11 @@ type ParseResult struct {
 // Rules, derived from real SYB samples (demo/shunyunbaoerp_stock_list.har)
 // plus the boundary cases already confirmed in the #41 prototype:
 //   - empty/whitespace-only input -> failed, nothing to extract.
-//   - no comma present (e.g. "均碼") -> uncertain: the whole string, with
-//     brackets stripped, becomes the size candidate; color stays empty. A
-//     single token with no separator cannot be split into two dimensions
-//     without guessing which one it is.
+//   - no comma present -> a single token becomes size when it carries an
+//     explicit size signal ("均碼") and colour otherwise (#274's descriptive
+//     colours). Multiple whitespace-delimited tokens are split by the same
+//     one-explicit-size rule the comma branch uses, so "黑色 XL" yields both
+//     dimensions instead of one over-long colour (see splitOnWhitespace).
 //   - comma present and exactly one side has an explicit size signal -> that
 //     side is size and the other side is color. This supports both observed
 //     SYB orders without allowing AI or a fuzzy color dictionary to swap roles.
@@ -75,6 +111,9 @@ func Parse(raw string) ParseResult {
 		if explicitSizePattern.MatchString(size) {
 			return ParseResult{Size: size, Status: models.SYBParseStatusSuccess, Note: "仅识别到尺码"}
 		}
+		if color, sizeToken, ok := splitOnWhitespace(size); ok {
+			return ParseResult{Color: color, Size: sizeToken, Status: models.SYBParseStatusSuccess}
+		}
 		return ParseResult{Color: size, Status: models.SYBParseStatusSuccess, Note: "仅识别到颜色"}
 	}
 
@@ -90,6 +129,9 @@ func Parse(raw string) ParseResult {
 		if explicitSizePattern.MatchString(only) {
 			return ParseResult{Size: only, Status: models.SYBParseStatusSuccess, Note: "仅识别到尺码"}
 		}
+		if color, sizeToken, ok := splitOnWhitespace(only); ok {
+			return ParseResult{Color: color, Size: sizeToken, Status: models.SYBParseStatusSuccess}
+		}
 		if only != "" {
 			return ParseResult{Color: only, Status: models.SYBParseStatusSuccess, Note: "仅识别到颜色"}
 		}

server/app/goauto/sybimport/parse_whitespace_split_test.go

@@ -0,0 +1,80 @@
+package sybimport_test
+
+import (
+	"testing"
+
+	"go-admin/app/goauto/models"
+	"go-admin/app/goauto/sybimport"
+)
+
+// #284: 空格分隔的「颜色 尺码」曾被整条当成颜色，尺码丢失且状态判为
+// success。success 的行不进 AI 解析队列（ai_parse_batch.go 只选 uncertain/failed），
+// 因此没有任何机制会纠正它们。本组用例锁住修复后的完整对照表，包括
+// #274 “接受描述性颜色”的意图必须保持不变。
+func TestParseSplitsWhitespaceWhenExactlyOneTokenIsASize(t *testing.T) {
+	cases := []struct {
+		name  string
+		raw   string
+		color string
+		size  string
+	}{
+		{"颜色在前", "黑色 XL", "黑色", "XL"},
+		{"单字母尺码", "紅色 M", "紅色", "M"},
+		// 与逗号分支一致：角色反转只看哪一侧带明确尺码特征，不看位置。
+		{"尺码在前", "XL 黑色", "黑色", "XL"},
+		{"颜色含多个词", "黑色 XL 加厚", "黑色 加厚", "XL"},
+	}
+	for _, item := range cases {
+		t.Run(item.name, func(t *testing.T) {
+			result := sybimport.Parse(item.raw)
+			if result.Status != models.SYBParseStatusSuccess {
+				t.Fatalf("status = %q, want success", result.Status)
+			}
+			if result.Color != item.color || result.Size != item.size {
+				t.Fatalf("got color=%q size=%q, want color=%q size=%q",
+					result.Color, result.Size, item.color, item.size)
+			}
+		})
+	}
+}
+
+// `[必须]` #274 的意图是让描述性颜色被接受而不是被拒绝。#284 只能在
+// “恰好一个 token 带明确尺码特征”时介入；没有尺码 token 的字串必须原样保留为
+// 颜色，否则就把 #274 退回去了。
+func TestParseKeepsDescriptiveColorsWithoutASizeToken(t *testing.T) {
+	// 来自真实 SYB 样本：“黑色+白色【純棉兩件裝】 簡約親膚”剥离备注后的形态。
+	result := sybimport.Parse("黑色+白色 簡約親膚")
+	if result.Status != models.SYBParseStatusSuccess {
+		t.Fatalf("status = %q, want success", result.Status)
+	}
+	if result.Color != "黑色+白色 簡約親膚" || result.Size != "" {
+		t.Fatalf("descriptive colour was altered: color=%q size=%q", result.Color, result.Size)
+	}
+}
+
+// 多个 token 都像尺码时无法安全地挑出颜色，与逗号分支「两侧均具尺码特征」
+// 同一道理，保守处理：不拆。实际数据中该形态的占比尚无证据（见 #284 风险节）。
+func TestParseDoesNotSplitWhenSeveralTokensLookLikeSizes(t *testing.T) {
+	result := sybimport.Parse("S M L")
+	if result.Size != "" {
+		t.Fatalf("ambiguous multi-size spec must not be split, got size=%q", result.Size)
+	}
+}
+
+// 单 token 的行为是 #274 定下的，#284 不得触碰。
+func TestParseSingleTokenBehaviourIsUnchanged(t *testing.T) {
+	if result := sybimport.Parse("均碼"); result.Size != "均碼" || result.Color != "" {
+		t.Fatalf("size-only token changed: color=%q size=%q", result.Color, result.Size)
+	}
+	if result := sybimport.Parse("黑色"); result.Color != "黑色" || result.Size != "" {
+		t.Fatalf("colour-only token changed: color=%q size=%q", result.Color, result.Size)
+	}
+}
+
+// 逗号仍然优先，空格规则不得干扰已有的逗号拆分。
+func TestParseCommaStillTakesPrecedenceOverWhitespace(t *testing.T) {
+	result := sybimport.Parse("黑色,XL")
+	if result.Color != "黑色" || result.Size != "XL" {
+		t.Fatalf("comma split changed: color=%q size=%q", result.Color, result.Size)
+	}
+}