Enhance Path QA and Stage Matching Logic

- Introduced multistage path quality assurance (QA) functionality to improve exercise relevance and feedback through structured tiers and optimization hints.
- Updated stage specifications to include `start_state` and `target_state` for better contextualization in roadmap matching.
- Enhanced semantic brief construction with technique sibling exclusions to refine exercise selection based on primary topics.
- Improved path retrieval logic to incorporate new parameters for nuanced matching against learning goals.
- Incremented application version to reflect these updates.

backend/migrations/090_ai_prompt_stage_transition_states.sql

@@ -0,0 +1,43 @@
+-- Migration 090: Stufenspecs — start_state / target_state pro Major Step (Soll-Verkettung)
+
+UPDATE ai_prompts SET
+    description = 'Phase C: Stufenspezifikation inkl. Soll-Start und Stufen-Ziel je Major Step.',
+    template = $t$Du bist Assistent für Kampfsport-Trainer und spezifizierst didaktische Stufen eines Progressionsgraphen.
+
+Anfrage: {{goal_query}}
+Zielanalyse: {{goal_analysis_json}}
+Major Steps: {{major_steps_json}}
+Planungs-Intent (Pfadweite Regeln): {{intent_context_json}}
+Semantic Brief: {{semantic_brief_json}}
+
+Jede Stufe ist ein Übergang im Gesamtpfad:
+- start_state: Soll-Zustand zu Beginn (= Ziel der vorherigen Stufe; Stufe 0 = Pfad-Start)
+- target_state: Zielzustand nach dieser Stufe (= Soll für die nächste Stufe)
+- learning_goal: messbares Lernziel der Übungssuche (was die Übung bringen soll)
+
+Felder je Major Step:
+- load_profile, exercise_type, success_criteria, anti_patterns (wie bisher)
+
+Regeln:
+1. start_state/target_state aus Zielanalyse und Major Steps ableiten — konsistente Kette.
+2. explicit_exclusions aus intent_context in anti_patterns jeder Stufe.
+3. success_criteria: prüfbar an Kurzbeschreibung + Übungsziel.
+4. Keine erfundenen Ausschlüsse.
+
+Antworte NUR mit JSON:
+{
+  "stage_specs": [
+    {
+      "major_step_index": 0,
+      "start_state": "…",
+      "target_state": "…",
+      "learning_goal": "…",
+      "load_profile": ["koordination"],
+      "exercise_type": "kihon_einzel",
+      "success_criteria": ["…"],
+      "anti_patterns": ["…"]
+    }
+  ]
+}$t$,
+    default_template = template
+WHERE slug = 'planning_progression_stage_spec';

backend/planning_exercise_path_builder.py

@@ -13,6 +13,8 @@ from pydantic import BaseModel, Field
 
 from tenant_context import TenantContext, library_content_visibility_sql
 from planning_exercise_profiles import PlanningTargetProfile
+from planning_path_qa_pipeline import run_multistage_path_qa
+from planning_stage_context import build_contextualized_stage_goal, resolve_path_start_target
 from planning_exercise_path_qa import (
     apply_llm_path_reorder,
     build_path_qa_summary,
@@ -191,6 +193,8 @@ def _pick_best_path_hit(
     stage_anti_patterns: Optional[List[str]] = None,
     roadmap_stage_match: bool = False,
     stage_match_brief: Optional[PlanningSemanticBrief] = None,
+    path_primary_topic: Optional[str] = None,
+    path_technique_excludes: Optional[List[str]] = None,
 ) -> Optional[Dict[str, Any]]:
     return pick_best_path_hit(
         hits,
@@ -200,6 +204,8 @@ def _pick_best_path_hit(
         stage_anti_patterns=stage_anti_patterns,
         roadmap_stage_match=roadmap_stage_match,
         stage_match_brief=stage_match_brief,
+        path_primary_topic=path_primary_topic,
+        path_technique_excludes=path_technique_excludes,
     )
 
 
@@ -304,6 +310,8 @@ def _run_path_step_retrieval(
     stage_success_criteria: Optional[List[str]] = None,
     stage_load_profile: Optional[List[str]] = None,
     path_context_note: Optional[str] = None,
+    path_primary_topic: Optional[str] = None,
+    path_technique_excludes: Optional[List[str]] = None,
 ) -> Tuple[List[Dict[str, Any]], PlanningTargetProfile, Dict[str, Any], str]:
     step_query = step_query_override or step_retrieval_query(
         semantic_brief, goal_query, step_index, max_steps
@@ -346,6 +354,8 @@ def _run_path_step_retrieval(
         "stage_success_criteria": list(stage_success_criteria or []),
         "stage_load_profile": list(stage_load_profile or []),
         "path_context_note": (path_context_note or "").strip() or None,
+        "path_primary_topic": (path_primary_topic or "").strip() or None,
+        "path_technique_excludes": list(path_technique_excludes or []),
     }
     pack = apply_progression_context_to_pack(
         cur,
@@ -514,6 +524,10 @@ def _annotate_roadmap_step(
     anti = list(anti_patterns_override or stage_spec.anti_patterns or [])
     if anti:
         step["roadmap_anti_patterns"] = anti
+    if (stage_spec.start_state or "").strip():
+        step["roadmap_start_state"] = stage_spec.start_state.strip()
+    if (stage_spec.target_state or "").strip():
+        step["roadmap_target_state"] = stage_spec.target_state.strip()
     step["roadmap_match_source"] = "stage_spec"
     if skill_expectations:
         step["skill_expectations"] = skill_expectations
@@ -562,6 +576,11 @@ def _build_steps_roadmap_first(
         if roadmap_ctx.resolved_structured
         else None
     )
+    path_start, path_target = resolve_path_start_target(
+        structured=roadmap_ctx.resolved_structured,
+        goal_analysis=roadmap_ctx.goal_analysis,
+    )
+    stage_count = len(stage_specs)
     brief_summary = (
         roadmap_ctx.semantic_brief
         if roadmap_ctx.semantic_brief
@@ -593,6 +612,17 @@ def _build_steps_roadmap_first(
         )
         step_kind = resolve_step_exercise_kind_filter(stage_spec, body.exercise_kind_any)
         stage_goal = (stage_spec.learning_goal or "").strip()
+        stage_start = (stage_spec.start_state or "").strip()
+        stage_target = (stage_spec.target_state or "").strip()
+        contextual_goal = build_contextualized_stage_goal(
+            learning_goal=stage_goal,
+            start_state=stage_start,
+            target_state=stage_target,
+            path_target_state=path_target,
+            path_start_state=path_start,
+            stage_index=step_index,
+            stage_count=stage_count,
+        )
         path_context_note = None
         if rs_dump:
             ctx_parts = [
@@ -607,6 +637,14 @@ def _build_steps_roadmap_first(
             extra_context=path_context_note,
         )
         stage_anti = list(dict.fromkeys([*(stage_spec.anti_patterns or []), *path_anti]))
+        path_primary = (semantic_brief.primary_topic or "").strip()
+        path_tech_excludes = list(semantic_brief.exclude_phrases or [])
+        if semantic_brief.topic_type == "technique" and path_primary:
+            from planning_exercise_semantics import technique_sibling_excludes
+
+            for item in technique_sibling_excludes(path_primary):
+                if item not in path_tech_excludes:
+                    path_tech_excludes.append(item)
         stage_match_brief = build_stage_match_brief(
             learning_goal=stage_goal,
             anti_patterns=stage_anti,
@@ -615,6 +653,12 @@ def _build_steps_roadmap_first(
             phase=major.phase if major else None,
             path_context_note=path_context_note,
             path_anti_patterns=path_anti,
+            path_primary_topic=path_primary or None,
+            path_technique_excludes=path_tech_excludes or None,
+            stage_start_state=stage_start or None,
+            stage_target_state=stage_target or None,
+            path_target_state=path_target or None,
+            contextualized_learning_goal=contextual_goal or None,
         )
 
         hits, _, _, _ = _run_path_step_retrieval(
@@ -641,6 +685,8 @@ def _build_steps_roadmap_first(
             stage_success_criteria=list(stage_spec.success_criteria or []),
             stage_load_profile=list(stage_spec.load_profile or []),
             path_context_note=path_context_note,
+            path_primary_topic=path_primary or None,
+            path_technique_excludes=path_tech_excludes or None,
         )
 
         hit = _pick_best_path_hit(
@@ -651,6 +697,8 @@ def _build_steps_roadmap_first(
             stage_anti_patterns=stage_anti or None,
             roadmap_stage_match=True,
             stage_match_brief=stage_match_brief,
+            path_primary_topic=path_primary or None,
+            path_technique_excludes=path_tech_excludes or None,
         )
 
         if not hit:
@@ -878,6 +926,13 @@ def _run_evaluate_only_path_qa(
                 roadmap_snapshot=path_roadmap_snapshot,
             )
 
+    multistage_qa = run_multistage_path_qa(
+        off_topic_steps=off_topic_steps,
+        stripped_off_topic=stripped_off_topic,
+        gaps=gaps,
+        llm_qa=llm_qa,
+        llm_applied=llm_qa_applied,
+    )
     path_qa = build_path_qa_summary(
         gaps=gaps,
         bridge_inserts=bridge_inserts,
@@ -890,6 +945,7 @@ def _run_evaluate_only_path_qa(
         reorder_applied=False,
         reorder_notes=[],
         roadmap_qa_mode=roadmap_qa_mode,
+        multistage_qa=multistage_qa,
     )
     return {
         "path_qa": path_qa,
@@ -1318,6 +1374,14 @@ def suggest_progression_path(
             if offer.get("offer_id") not in seen_offer_ids:
                 gap_fill_offers.append(offer)
 
+    multistage_qa = run_multistage_path_qa(
+        off_topic_steps=off_topic_steps,
+        stripped_off_topic=stripped_off_topic,
+        gaps=gaps,
+        llm_qa=llm_qa,
+        llm_applied=llm_qa_applied,
+        roadmap_unfilled=roadmap_unfilled if roadmap_first else None,
+    )
     path_qa = build_path_qa_summary(
         gaps=gaps,
         bridge_inserts=bridge_inserts,
@@ -1330,6 +1394,7 @@ def suggest_progression_path(
         reorder_applied=reorder_applied,
         reorder_notes=reorder_notes,
         roadmap_qa_mode=roadmap_qa_mode,
+        multistage_qa=multistage_qa,
     )
 
     target_profile_summary = path_target_profile.to_summary_dict(cur)

backend/planning_exercise_path_qa.py

@@ -23,10 +23,12 @@ from planning_exercise_semantics import (
     brief_to_summary_dict,
     exercise_passes_path_semantic_gate,
     exercise_passes_stage_learning_goal_gate,
+    exercise_passes_technique_path_scope,
     resolve_path_anti_patterns,
     score_exercise_semantic_relevance,
     semantic_brief_for_stage,
     step_phase_for_index,
+    technique_sibling_excludes,
 )
 
 _logger = logging.getLogger("shinkan.planning_exercise_path_qa")
@@ -434,6 +436,31 @@ def detect_off_topic_steps(
                 }
             )
             continue
+        primary = (brief.primary_topic or "").strip()
+        if brief.topic_type == "technique" and primary:
+            siblings = technique_sibling_excludes(primary)
+            stage_goal_pre = (step.get("roadmap_learning_goal") or "").strip()
+            if not exercise_passes_technique_path_scope(
+                primary_topic=primary,
+                title=bundle["title"],
+                summary=bundle["summary"],
+                goal=bundle["goal"],
+                learning_goal=stage_goal_pre,
+                sibling_excludes=siblings,
+                relaxed=False,
+            ):
+                off_topic.append(
+                    {
+                        "step_index": idx,
+                        "exercise_id": int(step["exercise_id"]),
+                        "title": step.get("title") or bundle["title"],
+                        "semantic_score": 0.0,
+                        "expected_phase": (step.get("roadmap_phase") or "").strip().lower() or None,
+                        "issue": "technique_scope",
+                        "reasons": [f"Passt nicht zur Haupttechnik „{primary}“"],
+                    }
+                )
+                continue
         stage_goal = (step.get("roadmap_learning_goal") or "").strip()
         phase = (step.get("roadmap_phase") or "").strip().lower() or step_phase_for_index(
             brief, idx, total
@@ -599,6 +626,7 @@ def build_path_qa_summary(
     reorder_applied: bool = False,
     reorder_notes: Optional[Sequence[str]] = None,
     roadmap_qa_mode: Optional[str] = None,
+    multistage_qa: Optional[Mapping[str, Any]] = None,
 ) -> Dict[str, Any]:
     offers = list(gap_fill_offers or [])
     off_topic = list(off_topic_steps or [])
@@ -619,6 +647,10 @@ def build_path_qa_summary(
         "reorder_notes": list(reorder_notes or []),
         "roadmap_qa_mode": roadmap_qa_mode,
     }
+    if multistage_qa:
+        summary["qa_tiers"] = list(multistage_qa.get("qa_tiers") or [])
+        summary["optimization_hints"] = list(multistage_qa.get("optimization_hints") or [])
+        summary["optimization_hint_count"] = int(multistage_qa.get("optimization_hint_count") or 0)
     if llm_qa:
         summary["overall_ok"] = bool(llm_qa.get("overall_ok", True))
         summary["quality_score"] = llm_qa.get("quality_score")

backend/planning_exercise_retrieval.py

@@ -351,6 +351,8 @@ def rank_visible_library_hits(
                 stage_semantic_score=stage_semantic_score,
                 anti_patterns=pack.get("stage_anti_patterns"),
                 step_phase=step_phase,
+                path_primary_topic=pack.get("path_primary_topic"),
+                path_technique_excludes=pack.get("path_technique_excludes"),
             ):
                 score_penalty = max(0.0, score_penalty - 0.10)
                 stage_match_reason = "Passt zum Stufen-Lernziel"

backend/planning_exercise_semantics.py

@@ -180,6 +180,51 @@ def _find_technique_in_text(q_lower: str) -> Optional[Tuple[str, Tuple[str, ...]
     return None
 
 
+def technique_sibling_excludes(primary_topic: str) -> List[str]:
+    """Andere Techniken derselben Familie (z. B. Mae/Yoko bei Mawashi) — aus Katalog."""
+    topic = _normalize_phrase(primary_topic)
+    if not topic:
+        return []
+    hit = _find_technique_in_text(topic)
+    if not hit:
+        return []
+    out: List[str] = []
+    for raw in hit[1]:
+        for expanded in _expand_stage_exclude_phrase(raw):
+            if expanded and expanded not in out:
+                out.append(expanded)
+    return out[:16]
+
+
+def exercise_passes_technique_path_scope(
+    *,
+    primary_topic: str,
+    title: str,
+    summary: str = "",
+    goal: str = "",
+    learning_goal: str = "",
+    sibling_excludes: Optional[Sequence[str]] = None,
+    relaxed: bool = False,
+) -> bool:
+    """
+    Technik-Pfad: keine Geschwister-Technik; Haupttechnik im Übungstext oder Stufen-Lernziel.
+    """
+    primary = _normalize_phrase(primary_topic)
+    if not primary:
+        return True
+
+    blob = _blob_from_fields(title, summary, goal, [])
+    excludes = list(sibling_excludes or technique_sibling_excludes(primary))
+    if excludes and _blob_matches_stage_excludes(blob, excludes):
+        return False
+
+    in_exercise = _phrase_in_blob(primary, blob)
+    in_stage = _phrase_in_blob(primary, _blob_from_fields("", "", learning_goal, []))
+    if in_exercise or in_stage:
+        return True
+    return relaxed
+
+
 def _detect_development_arc(q_lower: str) -> List[str]:
     found: List[str] = []
     for phase, markers in _ARC_PHASES:
@@ -850,13 +895,19 @@ def build_stage_match_brief(
     phase: Optional[str] = None,
     path_context_note: Optional[str] = None,
     path_anti_patterns: Optional[Sequence[str]] = None,
+    path_primary_topic: Optional[str] = None,
+    path_technique_excludes: Optional[Sequence[str]] = None,
+    stage_start_state: Optional[str] = None,
+    stage_target_state: Optional[str] = None,
+    path_target_state: Optional[str] = None,
+    contextualized_learning_goal: Optional[str] = None,
 ) -> PlanningSemanticBrief:
     """
     Stufen-zentrierter Semantik-Brief — unabhängig vom Gesamt-Pfad-Thema.
 
     Primär für Roadmap-Match: Bewertung gegen Titel + Kurzbeschreibung + Übungsziel.
     """
-    lg = (learning_goal or "").strip()
+    lg = (contextualized_learning_goal or learning_goal or "").strip()
     if len(lg) < 3:
         return PlanningSemanticBrief(semantic_strength=0.0)
 
@@ -865,9 +916,20 @@ def build_stage_match_brief(
         s = str(raw or "").strip()
         if s and s not in merged_anti:
             merged_anti.append(s)
+    primary_path = _normalize_phrase(path_primary_topic or "")
+    if primary_path:
+        for item in technique_sibling_excludes(primary_path):
+            if item not in merged_anti:
+                merged_anti.append(item)
+    for raw in path_technique_excludes or []:
+        for expanded in _expand_stage_exclude_phrase(str(raw or "")):
+            if expanded and expanded not in merged_anti:
+                merged_anti.append(expanded)
     constraints = parse_stage_goal_constraints(lg, merged_anti)
     must: List[str] = []
     norm_lg = _normalize_phrase(lg)
+    if primary_path and primary_path not in must:
+        must.insert(0, primary_path[:120])
     for token in constraints.positive_tokens:
         if token not in must:
             must.append(token)
@@ -883,6 +945,10 @@ def build_stage_match_brief(
             must.append(s[:60])
 
     retrieval_parts = [norm_lg]
+    for raw in (stage_start_state, stage_target_state, path_target_state):
+        s = _normalize_phrase(str(raw or ""))[:200]
+        if s and s not in retrieval_parts:
+            retrieval_parts.append(s)
     if path_context_note:
         note = _normalize_phrase(path_context_note)[:200]
         if note:
@@ -950,12 +1016,14 @@ def exercise_passes_stage_fit(
     stage_semantic_score: Optional[float] = None,
     anti_patterns: Optional[Sequence[str]] = None,
     step_phase: Optional[str] = None,
+    path_primary_topic: Optional[str] = None,
+    path_technique_excludes: Optional[Sequence[str]] = None,
     min_stage_semantic: float = _MIN_STAGE_FIT_SEMANTIC,
     relaxed: bool = False,
 ) -> bool:
     """Allgemeines Stufen-Fit-Gate: voller Übungstext vs. Stufen-Brief."""
     lg = (learning_goal or "").strip()
-    if len(lg) < 3:
+    if len(lg) < 3 and not (path_primary_topic or "").strip():
         return True
 
     blob = _blob_from_fields(title, summary, goal, [])
@@ -963,6 +1031,18 @@ def exercise_passes_stage_fit(
     if constraints.exclude_phrases and _blob_matches_stage_excludes(blob, constraints.exclude_phrases):
         return False
 
+    primary_path = (path_primary_topic or "").strip()
+    if primary_path and not exercise_passes_technique_path_scope(
+        primary_topic=primary_path,
+        title=title,
+        summary=summary,
+        goal=goal,
+        learning_goal=lg,
+        sibling_excludes=path_technique_excludes,
+        relaxed=relaxed,
+    ):
+        return False
+
     brief = stage_brief or build_stage_match_brief(
         learning_goal=lg,
         anti_patterns=anti_patterns,
@@ -1102,6 +1182,8 @@ def pick_best_path_hit(
     stage_anti_patterns: Optional[Sequence[str]] = None,
     roadmap_stage_match: bool = False,
     stage_match_brief: Optional[PlanningSemanticBrief] = None,
+    path_primary_topic: Optional[str] = None,
+    path_technique_excludes: Optional[Sequence[str]] = None,
 ) -> Optional[Dict[str, Any]]:
     """Gestufte Auswahl: strikt → relaxed → optional Notfall-Fallback."""
     if not hits:
@@ -1138,6 +1220,8 @@ def pick_best_path_hit(
                     stage_brief=stage_brief,
                     stage_semantic_score=stage_sem,
                     anti_patterns=stage_anti_patterns,
+                    path_primary_topic=path_primary_topic,
+                    path_technique_excludes=path_technique_excludes,
                     relaxed=not strict,
                 ):
                     continue
@@ -1209,8 +1293,10 @@ __all__ = [
     "merge_semantic_brief_llm",
     "parse_stage_goal_constraints",
     "pick_best_path_hit",
+    "exercise_passes_technique_path_scope",
     "score_exercise_stage_fit",
     "semantic_brief_for_stage",
+    "technique_sibling_excludes",
     "resolve_semantic_skill_weights",
     "score_exercise_semantic_relevance",
     "semantic_core_phrases",

backend/planning_intent_context.py

@@ -14,6 +14,7 @@ from typing import Any, Dict, List, Mapping, Optional, Sequence
 from planning_exercise_semantics import (
     PlanningSemanticBrief,
     resolve_path_anti_patterns,
+    technique_sibling_excludes,
 )
 
 _NEGATION_CLAUSE_RE = re.compile(
@@ -46,14 +47,18 @@ class PlanningIntentContext:
     path_success_criteria: List[str] = field(default_factory=list)
     explicit_exclusions: List[str] = field(default_factory=list)
     context_notes: str = ""
+    topic_type: str = "general"
+    technique_sibling_excludes: List[str] = field(default_factory=list)
 
     def to_api_dict(self) -> Dict[str, Any]:
         return {
             "source_query": self.source_query,
             "primary_topic": self.primary_topic,
+            "topic_type": self.topic_type,
             "path_anti_patterns": self.path_anti_patterns[:16],
             "path_success_criteria": self.path_success_criteria[:10],
             "explicit_exclusions": self.explicit_exclusions[:10],
+            "technique_sibling_excludes": self.technique_sibling_excludes[:16],
             "context_notes": self.context_notes[:1200] or None,
         }
 
@@ -101,15 +106,32 @@ def build_planning_intent_context(
             path_success.append(line)
 
     topic = (primary_topic or ga.get("primary_topic") or "").strip()
-    if semantic_brief and not topic:
-        topic = (semantic_brief.primary_topic or "").strip()
+    topic_type = "general"
+    siblings: List[str] = []
+    if semantic_brief:
+        if not topic:
+            topic = (semantic_brief.primary_topic or "").strip()
+        topic_type = (semantic_brief.topic_type or "general").strip().lower()
+        if topic_type == "technique" and topic:
+            siblings = technique_sibling_excludes(topic)
+            for raw in semantic_brief.exclude_phrases or []:
+                s = str(raw or "").strip()
+                if s and s.lower() not in {x.lower() for x in siblings}:
+                    siblings.append(s[:120])
+
+    if topic_type == "technique" and topic:
+        line = f"Haupttechnik {topic} in Kurzbeschreibung oder Übungsziel erkennbar"
+        if line not in path_success:
+            path_success.insert(0, line)
 
     return PlanningIntentContext(
         source_query=(goal_query or "").strip(),
         primary_topic=topic,
+        topic_type=topic_type,
         path_anti_patterns=path_anti,
         path_success_criteria=path_success,
         explicit_exclusions=explicit,
+        technique_sibling_excludes=siblings[:16],
         context_notes=combined_notes[:1200],
     )
 
@@ -148,13 +170,23 @@ def finalize_stage_spec_artifact(
             *(spec.anti_patterns or []),
             *intent.explicit_exclusions,
             *intent.path_anti_patterns,
+            *intent.technique_sibling_excludes,
+            (
+                f"andere Technik als {intent.primary_topic}"
+                if intent.topic_type == "technique" and intent.primary_topic
+                else ""
+            ),
         ],
         limit=14,
     )
+    stage_start = (spec.start_state or "").strip()
+    stage_target = (spec.target_state or "").strip()
     success = _dedupe_preserve(
         [
             *(spec.success_criteria or []),
             *intent.path_success_criteria,
+            (f"Soll-Start der Stufe erreichbar: {stage_start[:180]}" if stage_start else ""),
+            (f"Stufen-Ziel erreichbar: {stage_target[:180]}" if stage_target else ""),
             (
                 f"Übung liefert messbar: {learning_goal[:160]}"
                 if learning_goal

backend/planning_path_qa_pipeline.py

@@ -0,0 +1,176 @@
+"""
+Mehrstufige Pfad-QS — Findings pro Stufe, daraus Optimierungspotenziale ableiten.
+
+Stufen (allgemein, domänenneutral):
+  1. deterministische Gates (Technik-Scope, Ausschlüsse, Stufen-Fit)
+  2. Übergangs-Kohärenz (Lücken zwischen Schritten)
+  3. LLM-Ganzpfad-Bewertung (Empfehlungen, keine Auto-Patches)
+"""
+from __future__ import annotations
+
+from typing import Any, Dict, List, Mapping, Optional, Sequence
+
+
+_ACTION_BY_ISSUE: Dict[str, str] = {
+    "technique_scope": "rematch_slot",
+    "path_exclude": "rematch_slot",
+    "stage_mismatch": "refine_stage_spec",
+    "off_topic": "rematch_slot",
+    "gap": "bridge_or_gap_fill",
+    "large_gap": "bridge_or_gap_fill",
+    "roadmap_unfilled": "rematch_slot",
+}
+
+
+def _action_for_finding(finding: Mapping[str, Any]) -> str:
+    issue = str(finding.get("issue") or finding.get("type") or "").strip().lower()
+    if finding.get("is_large_gap"):
+        return "bridge_or_gap_fill"
+    return _ACTION_BY_ISSUE.get(issue, "review")
+
+
+def _hint_from_finding(finding: Mapping[str, Any], *, tier: str) -> Dict[str, Any]:
+    step_index = finding.get("step_index")
+    if step_index is None:
+        step_index = finding.get("major_step_index")
+    issue = str(finding.get("issue") or finding.get("type") or tier)
+    action = _action_for_finding(finding)
+    title = str(finding.get("title") or finding.get("removed_title") or "").strip()
+    reasons = finding.get("reasons") or []
+    reason = reasons[0] if reasons else str(finding.get("rationale") or finding.get("detail") or "")
+
+    hint: Dict[str, Any] = {
+        "tier": tier,
+        "action": action,
+        "issue": issue,
+        "step_index": step_index,
+        "title": title or None,
+        "reason": (reason or "")[:400] or None,
+    }
+    if finding.get("roadmap_learning_goal"):
+        hint["roadmap_learning_goal"] = finding.get("roadmap_learning_goal")
+    if finding.get("roadmap_major_step_index") is not None:
+        hint["roadmap_major_step_index"] = finding.get("roadmap_major_step_index")
+    return {k: v for k, v in hint.items() if v is not None and v != ""}
+
+
+def derive_optimization_hints(
+    tiers: Sequence[Mapping[str, Any]],
+) -> List[Dict[str, Any]]:
+    """Aus QS-Stufen konkrete Optimierungsaktionen (ohne anfrage-spezifische Heuristiken)."""
+    hints: List[Dict[str, Any]] = []
+    seen: set[tuple] = set()
+    for tier in tiers:
+        tier_id = str(tier.get("id") or "")
+        for finding in tier.get("findings") or []:
+            if not isinstance(finding, dict):
+                continue
+            hint = _hint_from_finding(finding, tier=tier_id)
+            key = (
+                hint.get("tier"),
+                hint.get("action"),
+                hint.get("step_index"),
+                hint.get("issue"),
+            )
+            if key in seen:
+                continue
+            seen.add(key)
+            hints.append(hint)
+    return hints[:24]
+
+
+def run_multistage_path_qa(
+    *,
+    off_topic_steps: Sequence[Mapping[str, Any]],
+    stripped_off_topic: Sequence[Mapping[str, Any]],
+    gaps: Sequence[Mapping[str, Any]],
+    llm_qa: Optional[Mapping[str, Any]] = None,
+    llm_applied: bool = False,
+    roadmap_unfilled: Optional[Sequence[Mapping[str, Any]]] = None,
+) -> Dict[str, Any]:
+    """Orchestriert QS-Stufen und leitet Optimierungspotenziale ab."""
+    tier1_findings: List[Dict[str, Any]] = []
+    for item in stripped_off_topic or off_topic_steps or []:
+        if isinstance(item, dict):
+            tier1_findings.append(dict(item))
+
+    tier2_findings: List[Dict[str, Any]] = [dict(g) for g in gaps if isinstance(g, dict)]
+
+    tier3_findings: List[Dict[str, Any]] = []
+    llm_recommendations: List[str] = []
+    if llm_applied and llm_qa:
+        q_score = llm_qa.get("quality_score")
+        tier3_findings.append(
+            {
+                "issue": "llm_assessment",
+                "quality_score": q_score,
+                "overall_ok": llm_qa.get("overall_ok"),
+                "detail": llm_qa.get("summary") or llm_qa.get("assessment"),
+            }
+        )
+        for raw in llm_qa.get("recommendations") or llm_qa.get("suggestions") or []:
+            s = str(raw or "").strip()
+            if s:
+                llm_recommendations.append(s[:500])
+
+    unfilled = list(roadmap_unfilled or [])
+    if unfilled:
+        for item in unfilled:
+            if isinstance(item, (list, tuple)) and len(item) >= 2:
+                idx, spec = item[0], item[1]
+                tier1_findings.append(
+                    {
+                        "issue": "roadmap_unfilled",
+                        "step_index": int(idx),
+                        "roadmap_major_step_index": getattr(spec, "major_step_index", idx),
+                        "roadmap_learning_goal": getattr(spec, "learning_goal", None),
+                        "reasons": ["Keine passende Übung für Roadmap-Stufe"],
+                    }
+                )
+            elif isinstance(item, dict):
+                tier1_findings.append({**item, "issue": item.get("issue") or "roadmap_unfilled"})
+
+    tiers: List[Dict[str, Any]] = [
+        {
+            "id": "tier1_deterministic",
+            "label": "Deterministische Gates",
+            "finding_count": len(tier1_findings),
+            "findings": tier1_findings[:16],
+        },
+        {
+            "id": "tier2_transitions",
+            "label": "Übergangs-Kohärenz",
+            "finding_count": len(tier2_findings),
+            "findings": tier2_findings[:12],
+        },
+        {
+            "id": "tier3_llm_holistic",
+            "label": "LLM-Ganzpfad",
+            "finding_count": len(tier3_findings),
+            "findings": tier3_findings,
+            "recommendations": llm_recommendations[:8],
+            "applied": bool(llm_applied),
+        },
+    ]
+    optimization_hints = derive_optimization_hints(tiers)
+    for rec in llm_recommendations[:5]:
+        optimization_hints.append(
+            {
+                "tier": "tier3_llm_holistic",
+                "action": "review_roadmap",
+                "issue": "llm_recommendation",
+                "reason": rec,
+            }
+        )
+
+    return {
+        "qa_tiers": tiers,
+        "optimization_hints": optimization_hints[:28],
+        "optimization_hint_count": len(optimization_hints),
+    }
+
+
+__all__ = [
+    "derive_optimization_hints",
+    "run_multistage_path_qa",
+]

backend/planning_progression_roadmap.py

@@ -104,6 +104,10 @@ class RoadmapArtifact(BaseModel):
 class StageSpecArtifact(BaseModel):
     major_step_index: int = Field(ge=0)
     learning_goal: str = ""
+    """Soll-Start dieser Stufe (= Zielzustand der vorherigen Stufe / Pfad-Start)."""
+    start_state: str = ""
+    """Zielzustand dieser Stufe (= Soll für den nächsten Schritt)."""
+    target_state: str = ""
     load_profile: List[str] = Field(default_factory=list)
     exercise_type: str = ""
     success_criteria: List[str] = Field(default_factory=list)
@@ -941,6 +945,8 @@ def roadmap_context_from_override(
                 StageSpecArtifact(
                     major_step_index=i,
                     learning_goal=(spec.learning_goal or majors[i].learning_goal).strip(),
+                    start_state=(spec.start_state or "").strip(),
+                    target_state=(spec.target_state or "").strip(),
                     load_profile=list(spec.load_profile or []),
                     exercise_type=(spec.exercise_type or "").strip(),
                     success_criteria=list(spec.success_criteria or []),
@@ -1004,6 +1010,19 @@ def roadmap_context_from_override(
             semantic_brief=enriched_brief,
         ),
     )
+    from planning_stage_context import derive_stage_specs_transition_states, resolve_path_start_target
+
+    path_start, path_target = resolve_path_start_target(
+        structured=structured,
+        goal_analysis=goal_analysis,
+    )
+    stage_specs = derive_stage_specs_transition_states(
+        stage_specs,
+        majors,
+        path_start=path_start,
+        path_target=path_target,
+        goal_analysis=goal_analysis,
+    )
 
     return ProgressionRoadmapContext(
         goal_query=goal_query.strip(),
@@ -1225,6 +1244,19 @@ def run_progression_roadmap_pipeline(
         intent=intent,
         fallback_specs=heuristic_specs,
     )
+    from planning_stage_context import derive_stage_specs_transition_states, resolve_path_start_target
+
+    path_start, path_target = resolve_path_start_target(
+        structured=resolved,
+        goal_analysis=goal_analysis,
+    )
+    ctx.stage_specs = derive_stage_specs_transition_states(
+        ctx.stage_specs,
+        roadmap.major_steps,
+        path_start=path_start,
+        path_target=path_target,
+        goal_analysis=goal_analysis,
+    )
 
     if ctx.llm_goal_analysis_applied or ctx.llm_roadmap_applied or ctx.llm_stage_spec_applied:
         ctx.pipeline_phase = "roadmap_v1_llm"

backend/planning_stage_context.py

@@ -0,0 +1,140 @@
+"""
+Stufen-Kontext im Gesamtziel — Start/Ziel pro Roadmap-Stufe für Matching und QS.
+
+Übertragbar auf Trainingsplanung: Abschnitt-Soll (= Ende Vorabschnitt), Abschnitt-Ziel.
+"""
+from __future__ import annotations
+
+from typing import List, Optional, Sequence
+
+from planning_progression_roadmap import (
+    GoalAnalysisArtifact,
+    MajorStep,
+    RoadmapStructuredInput,
+    StageSpecArtifact,
+)
+
+
+def build_contextualized_stage_goal(
+    *,
+    learning_goal: str,
+    start_state: str = "",
+    target_state: str = "",
+    path_target_state: str = "",
+    path_start_state: str = "",
+    stage_index: int = 0,
+    stage_count: int = 1,
+) -> str:
+    """Stufen-Lernziel eingebettet in Übergang und Gesamtziel (für Brief/Retrieval)."""
+    lg = (learning_goal or "").strip()
+    if not lg:
+        return ""
+
+    parts: List[str] = []
+    start = (start_state or "").strip()
+    target = (target_state or "").strip()
+    path_end = (path_target_state or "").strip()
+    path_begin = (path_start_state or "").strip()
+
+    if start:
+        parts.append(f"Soll-Start: {start[:220]}")
+    elif path_begin and stage_index == 0:
+        parts.append(f"Pfad-Start: {path_begin[:220]}")
+    if target:
+        parts.append(f"Stufen-Ziel: {target[:220]}")
+    parts.append(f"Lernziel: {lg[:280]}")
+    if path_end:
+        if stage_index >= max(0, stage_count - 1):
+            parts.append(f"Gesamtziel: {path_end[:220]}")
+        else:
+            parts.append(f"Gesamtziel (Kontext): {path_end[:180]}")
+
+    return " | ".join(parts)[:900]
+
+
+def derive_stage_specs_transition_states(
+    stage_specs: Sequence[StageSpecArtifact],
+    major_steps: Sequence[MajorStep],
+    *,
+    path_start: str = "",
+    path_target: str = "",
+    goal_analysis: Optional[GoalAnalysisArtifact] = None,
+) -> List[StageSpecArtifact]:
+    """
+    Verkettete Soll-/Zielzustände je Stufe.
+
+    - Stufe 0 start = Pfad-Start
+    - Stufe n start = Zielzustand Stufe n-1 (Ziel des vorherigen Schritts)
+    - Letzte Stufe target = Pfad-Gesamtziel (falls gesetzt)
+    """
+    start_path = (path_start or "").strip()
+    end_path = (path_target or "").strip()
+    if goal_analysis:
+        if not start_path:
+            start_path = (goal_analysis.start_assumption or "").strip()
+        if not end_path:
+            end_path = (goal_analysis.target_state or "").strip()
+
+    by_idx = {int(s.major_step_index): s for s in stage_specs}
+    majors = sorted(major_steps, key=lambda m: m.index)
+    if not majors:
+        return list(stage_specs)
+
+    out: List[StageSpecArtifact] = []
+    prev_target = start_path
+    last_idx = majors[-1].index
+
+    for major in majors:
+        spec = by_idx.get(major.index)
+        if spec is None:
+            spec = StageSpecArtifact(
+                major_step_index=major.index,
+                learning_goal=major.learning_goal,
+            )
+
+        explicit_start = (spec.start_state or "").strip()
+        explicit_target = (spec.target_state or "").strip()
+        stage_start = explicit_start or prev_target or start_path
+        if explicit_target:
+            stage_target = explicit_target
+        elif major.index == last_idx and end_path:
+            stage_target = end_path
+        else:
+            stage_target = (major.learning_goal or spec.learning_goal or "").strip()
+
+        prev_target = stage_target
+        out.append(
+            spec.model_copy(
+                update={
+                    "start_state": (stage_start or "")[:500],
+                    "target_state": (stage_target or "")[:500],
+                }
+            )
+        )
+    return out
+
+
+def resolve_path_start_target(
+    *,
+    structured: Optional[RoadmapStructuredInput] = None,
+    goal_analysis: Optional[GoalAnalysisArtifact] = None,
+) -> tuple[str, str]:
+    """Pfadweiter Start- und Zielzustand für Stufen-Verkettung."""
+    start = ""
+    target = ""
+    if structured:
+        start = (structured.start_situation or "").strip()
+        target = (structured.target_state or "").strip()
+    if goal_analysis:
+        if not start:
+            start = (goal_analysis.start_assumption or "").strip()
+        if not target:
+            target = (goal_analysis.target_state or "").strip()
+    return start, target
+
+
+__all__ = [
+    "build_contextualized_stage_goal",
+    "derive_stage_specs_transition_states",
+    "resolve_path_start_target",
+]

backend/tests/test_planning_roadmap_stage_match.py

@@ -5,10 +5,12 @@ from planning_exercise_semantics import (
     enrich_brief_with_path_constraints,
     exercise_passes_stage_learning_goal_gate,
     exercise_passes_stage_fit,
+    exercise_passes_technique_path_scope,
     pick_best_path_hit,
     resolve_path_anti_patterns,
     score_exercise_stage_fit,
     semantic_brief_for_stage,
+    technique_sibling_excludes,
 )
 from planning_exercise_path_qa import strip_off_topic_steps_from_path
 
@@ -247,6 +249,47 @@ def test_pick_best_skips_kumite_for_mawashi_athletic_path():
     assert int(chosen["id"]) == 2
 
 
+def test_technique_scope_rejects_sibling_geri_for_mawashi_path():
+    siblings = technique_sibling_excludes("mawashi geri")
+    assert any("mae" in s for s in siblings)
+    assert not exercise_passes_technique_path_scope(
+        primary_topic="mawashi geri",
+        title="Mae Geri Grundtechnik",
+        summary="Front kick",
+        goal="Präzision Mae Geri",
+        learning_goal="Sprungvorbereitung für Mawashi Geri",
+        sibling_excludes=siblings,
+    )
+    assert exercise_passes_technique_path_scope(
+        primary_topic="mawashi geri",
+        title="Sprungkraft Plyometrie",
+        summary="Absprung",
+        goal="Vorbereitung gesprungener Mawashi Geri",
+        learning_goal="Sprungvorbereitung für Mawashi Geri",
+        sibling_excludes=siblings,
+    )
+
+
+def test_stage_fit_rejects_yoko_geri_on_mawashi_roadmap_stage():
+    brief = build_semantic_brief("gesprungener Mawashi Geri Sprungphase")
+    primary = brief.primary_topic or "mawashi geri"
+    stage_goal = "Koordination Sprungphase Mawashi Geri"
+    stage_brief = build_stage_match_brief(
+        learning_goal=stage_goal,
+        path_primary_topic=primary,
+        path_technique_excludes=technique_sibling_excludes(primary),
+    )
+    assert not exercise_passes_stage_fit(
+        learning_goal=stage_goal,
+        title="Yoko Geri seitlicher Tritt",
+        summary="Seitwärtskick",
+        goal="Yoko Geri Technik",
+        stage_brief=stage_brief,
+        path_primary_topic=primary,
+        path_technique_excludes=technique_sibling_excludes(primary),
+    )
+
+
 def test_strip_off_topic_removes_partial_when_most_steps_bad():
     steps = [{"exercise_id": i, "title": f"E{i}"} for i in range(1, 8)]
     off_topic = [{"step_index": i, "issue": "path_exclude"} for i in range(5)]

backend/tests/test_planning_stage_context.py

@@ -0,0 +1,65 @@
+"""Tests Stufen-Kontext (Start/Ziel-Verkettung) und mehrstufige QS."""
+from planning_path_qa_pipeline import derive_optimization_hints, run_multistage_path_qa
+from planning_progression_roadmap import MajorStep, StageSpecArtifact
+from planning_stage_context import (
+    build_contextualized_stage_goal,
+    derive_stage_specs_transition_states,
+)
+
+
+def test_derive_stage_transition_chain():
+    majors = [
+        MajorStep(index=0, phase="grundlage", learning_goal="Stand-Mawashi", consolidates=["m1"]),
+        MajorStep(index=1, phase="vertiefung", learning_goal="Sprungvorbereitung", consolidates=["m2"]),
+        MajorStep(index=2, phase="perfektion", learning_goal="Gesprungener Mawashi", consolidates=["m3"]),
+    ]
+    specs = [
+        StageSpecArtifact(major_step_index=0, learning_goal=majors[0].learning_goal),
+        StageSpecArtifact(major_step_index=1, learning_goal=majors[1].learning_goal),
+        StageSpecArtifact(major_step_index=2, learning_goal=majors[2].learning_goal),
+    ]
+    out = derive_stage_specs_transition_states(
+        specs,
+        majors,
+        path_start="Anfänger mit Grundstellung",
+        path_target="Sauberer gesprungener Mawashi Geri",
+    )
+    assert out[0].start_state == "Anfänger mit Grundstellung"
+    assert out[1].start_state == out[0].target_state
+    assert out[2].target_state == "Sauberer gesprungener Mawashi Geri"
+
+
+def test_contextualized_stage_goal_includes_path_target():
+    text = build_contextualized_stage_goal(
+        learning_goal="Sprungkoordination",
+        start_state="Stand-Mawashi sicher",
+        target_state="Explosiver Absprung",
+        path_target_state="Gesprungener Mawashi Geri",
+        stage_index=1,
+        stage_count=3,
+    )
+    assert "Sprungkoordination" in text
+    assert "Gesamtziel" in text
+    assert "Soll-Start" in text
+
+
+def test_multistage_qa_emits_optimization_hints():
+    result = run_multistage_path_qa(
+        off_topic_steps=[],
+        stripped_off_topic=[
+            {
+                "step_index": 2,
+                "issue": "technique_scope",
+                "title": "Yoko Geri",
+                "reasons": ["Passt nicht zur Haupttechnik"],
+            }
+        ],
+        gaps=[{"from_title": "A", "to_title": "B", "gap_score": 0.6, "is_large_gap": True}],
+        llm_qa={"quality_score": 0.25, "recommendations": ["Athletisches Training ergänzen"]},
+        llm_applied=True,
+    )
+    assert len(result["qa_tiers"]) == 3
+    hints = result["optimization_hints"]
+    assert any(h.get("action") == "rematch_slot" for h in hints)
+    assert any(h.get("action") == "bridge_or_gap_fill" for h in hints)
+    assert derive_optimization_hints(result["qa_tiers"])

backend/version.py

@@ -1,8 +1,8 @@
 # Shinkan Jinkendo Version Information
 
-APP_VERSION = "0.8.223"
+APP_VERSION = "0.8.225"
 BUILD_DATE = "2026-06-07"
-DB_SCHEMA_VERSION = "20260607089"
+DB_SCHEMA_VERSION = "20260607090"
 
 MODULE_VERSIONS = {
     "legal_documents": "1.4.0",  # Admin: Live-Vorschau pro Abschnitt + modale Vollvorschau (Editor + Dokumentenliste)
@@ -53,6 +53,22 @@ MODULE_VERSIONS = {
 }
 
 CHANGELOG = [
+    {
+        "version": "0.8.225",
+        "date": "2026-06-07",
+        "changes": [
+            "Stufen start_state/target_state (Soll-Verkettung) + kontextualisiertes Matching.",
+            "Mehrstufige Pfad-QS (tier1–3) mit optimization_hints; Migration 090 Prompt.",
+        ],
+    },
+    {
+        "version": "0.8.224",
+        "date": "2026-06-07",
+        "changes": [
+            "Technik-Pfad-Scope: Geschwister-Techniken (Mae/Yoko bei Mawashi) als hartes Gate in Match/QS.",
+            "path_primary_topic in build_stage_match_brief; Intent technique_sibling_excludes.",
+        ],
+    },
     {
         "version": "0.8.223",
         "date": "2026-06-07",

docs/architecture/PLANNING_PROGRESSION_GRAPH_KI.md

@@ -241,6 +241,28 @@ LLM: Migration **089** — Prompts `planning_progression_goal_analysis` + `plann
 
 Matching: `anti_patterns` + `success_criteria` → `build_stage_match_brief` → Retrieval-Gate (Titel + Summary + Goal).
 
+### Stufen im Gesamtziel (`planning_stage_context.py`)
+
+| Feld | Bedeutung |
+|------|-----------|
+| `start_state` | Soll-Start der Stufe (= `target_state` der Vorstufe / Pfad-Start) |
+| `target_state` | Ziel nach dieser Stufe (= Soll für den nächsten Schritt) |
+| `build_contextualized_stage_goal()` | Lernziel + Start + Stufen-Ziel + Gesamtziel → Brief/Retrieval |
+
+Deterministisch: `derive_stage_specs_transition_states()` nach Roadmap-Pipeline; LLM kann Felder überschreiben (Prompt **090**).
+
+### Mehrstufige Pfad-QS (`planning_path_qa_pipeline.py`)
+
+| Stufe | Inhalt | Ableitung |
+|-------|--------|-----------|
+| **tier1** | Deterministische Gates (Technik-Scope, Ausschlüsse, unfilled) | `optimization_hints` → `rematch_slot`, `refine_stage_spec` |
+| **tier2** | Übergangs-Lücken zwischen Schritten | `bridge_or_gap_fill` |
+| **tier3** | LLM-Ganzpfad + Empfehlungen | `review_roadmap` |
+
+API: `path_qa.qa_tiers`, `path_qa.optimization_hints` — **kein** anfrage-spezifischer Patch, sondern strukturierte Rückkopplung. Auto-Rematch-Schleife: Backlog (QS → Aktion → erneutes Match).
+
+Phase G: gleiche Tiers für Abschnitts-QS nach Einheits-Match.
+
 ## 9. Fähigkeiten-Scoring-Anbindung
 
 Modul: `planning_skill_expectations.py`