fix(distill): preserve tables in technical spec docs via table-density detection

internal/distill/distill.go

@@ -464,23 +464,30 @@ func doHTTP(req *http.Request) ([]byte, error) {
 // preserve all precise identifiers for FTS search.
 func isTechnicalCatalog(content string) bool {
 	lines := strings.Split(content, "\n")
-	if len(lines) < 150 {
+	if len(lines) < 50 {
 		return false
 	}
 	// Count numbered entity patterns: M01, M02... or similar coded lists
 	numberedRe := regexp.MustCompile(`\bM\d{2}\b`)
 	tableRe := regexp.MustCompile(`(?i)(iceberg_|dim_|dwd_|hive_|matrixdb|\.im_edge\.)`)
 	numberedCount := 0
 	tableCount := 0
+	tableLineCount := 0
 	for _, line := range lines {
 		if numberedRe.MatchString(line) {
 			numberedCount++
 		}
 		if tableRe.MatchString(line) {
 			tableCount++
 		}
+		if strings.HasPrefix(strings.TrimSpace(line), "|") {
+			tableLineCount++
+		}
 	}
-	return numberedCount >= 5 || tableCount >= 10
+	// High table density (>30% of lines are table rows) indicates a data dictionary
+	// or field catalog that should be preserved verbatim rather than summarized.
+	tableRatio := float64(tableLineCount) / float64(len(lines))
+	return numberedCount >= 5 || tableCount >= 10 || tableRatio >= 0.30
 }
 
 // lightweightNote generates a source-note without LLM distillation.

internal/distill/prompt.go

@@ -69,6 +69,12 @@ preserve EVERY item completely — including its ID/code, name, source system, a
 storage table names (Hive/Iceberg/MatrixDB paths, database.schema.table identifiers).
 Do NOT summarize, merge, or omit any entry. Partial preservation is a critical failure.
 
+TABLE PRESERVATION RULE (MANDATORY for doc_type=技术规范 only): If the document is a
+technical specification (data dictionary, field catalog, API spec, system design doc),
+reproduce ALL Markdown tables verbatim in the Key Facts section. Do NOT convert table
+rows into prose. Every cell value must appear exactly as in the source. Omitting or
+summarizing a table in a 技术规范 document is a critical failure.
+
 ## Quotes
 Notable direct quotes from the document (if any). If none, write "None."
 

internal/kbinit/schema/templates/source-note.md

@@ -75,6 +75,12 @@ supports: []    # PATHS ONLY: same rule as related_to.
   If the source contains numbered/coded items (e.g. M01-M43, API list, field catalog),
   ALL items MUST be preserved — do NOT summarize, merge, or omit any entry.
   Each item must retain: code/ID, name, source system, storage table names, and any specific technical identifiers.
+
+  TABLE PRESERVATION RULE (MANDATORY for doc_type=技术规范 only):
+  If this is a technical specification (data dictionary, field catalog, API spec),
+  reproduce ALL Markdown tables VERBATIM in Key Facts. Do NOT convert table rows
+  into prose. Every cell value must appear exactly as in the source.
+  Omitting or summarizing a table in a 技术规范 document is a critical failure.
   Partial preservation is a critical quality failure.
 -->