chore(rag): PR #167 회차 1 리뷰 반영

- `parse_decompose_response_drops_partial_empty_keeps_valid` 신규
  회귀 핀 — `["", "valid q", "  "]` → `["valid q"]` (trim+filter
  chain 동작 pin).
- `multi_hop_decompose` 의 `stop: Vec::new()` 옆 doc comment 추가 —
  의도 명시 (instruction-following 모델 기대 + prose 추가 시
  MultiHopDecomposeFailed refusal 가 policy). 회차 1 question 의
  답변.
- plan 의 PR-3 implementation order 에 회차 1 carry-over 추가:
  1) ask + ask_multi_hop 의 §4-§9 mirror → 공통 helper 추출,
  2) decompose template 의 substitution corner case → format!
     named arg 으로 교체.

회차 1 의 다른 suggestion (mirror refactor, substitution corner
case, history block helper) 는 PR-3 합리적 timing 으로 plan 에
명시 — 회차 2 reply 에 정리.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

crates/kebab-rag/src/pipeline.rs

@@ -908,6 +908,16 @@ impl RagPipeline {
         let req = GenerateRequest {
             system: MULTI_HOP_DECOMPOSE_SYSTEM_PROMPT.to_string(),
             user,
+            // Empty stop is intentional. Instruction-following models
+            // (gemma3:4b+ / gemma4:e4b / Claude / GPT-4) honor the
+            // "JSON array only" prompt rule, so prose past the
+            // closing `]` is rare. If a downstream LM does append
+            // prose, `parse_decompose_response` returns `None` and
+            // the caller surfaces `MultiHopDecomposeFailed` — that
+            // is the policy. Adding a trailing-`]` stop sequence
+            // risks truncating the array (LM emits the close bracket
+            // and we cut the response one token too early), which
+            // is a worse failure mode than the explicit refusal.
             stop: Vec::new(),
             // JSON array of up to 5 sub-questions is short. 512 is a
             // comfortable cap that fits in any context window without
@@ -1495,6 +1505,16 @@ mod tests {
         assert_eq!(out, vec!["rust async", "tokio runtime"]);
     }
 
+    /// Partial-empty case — drop the empties, keep the rest. Pins
+    /// the trim-then-filter chain in `parse_decompose_response` so a
+    /// future refactor that reorders the steps (e.g. take-then-trim)
+    /// can't accidentally swallow valid sub-queries.
+    #[test]
+    fn parse_decompose_response_drops_partial_empty_keeps_valid() {
+        let out = parse_decompose_response(r#"["", "valid q", "  "]"#).unwrap();
+        assert_eq!(out, vec!["valid q"]);
+    }
+
     #[test]
     fn est_tokens_approx_quarters() {
         assert_eq!(est_tokens(""), 0);

docs/superpowers/plans/2026-05-25-p9-fb-41-multi-hop-rag.md

@@ -112,6 +112,9 @@ XL 작업 — 6 PR 분할 (각 머지 후 누적, 마지막 PR 후 v0.18.0 cut).
    - cap 도달 (max_depth / max_total_sub_queries / max_pool_chunks) 시 forced_stop=true 로 break.
    - synthesize → Answer.hops 에 누적된 HopRecord array 첨부.
 6. decide JSON parse failure → forced_stop synthesize (refusal 아님, 안전한 graceful degrade).
+7. **PR-2 회차 1 carry-over** — 같은 PR 에서 함께 해소:
+   - `ask` + `ask_multi_hop` 의 §4-§9 mirror (~150 줄 중복) → 공통 helper `synthesize_with_packed_context` 추출. history block 처리도 helper 화. drift 위험 차단.
+   - `MULTI_HOP_DECOMPOSE_USER_TEMPLATE` 의 `.replace("{query}", ...).replace("{max_sub_queries}", ...)` corner case → `format!` named arg 또는 strict substitution helper 로 교체. 사용자 query 에 literal `{max_sub_queries}` 포함 시 mis-replace 회피.
 
 **Test**:
 - `multi_hop_decide_stop_triggers_synthesize` — decide 가 `[]` 반환 시 즉시 synthesize.