infinity-context 0.1.0__py3-none-any.whl

infinity_context_server/eval_case_catalog.py

@@ -0,0 +1,1590 @@
+"""Deterministic eval case catalogs."""
+
+from __future__ import annotations
+
+from infinity_context_server.eval_types import EvalCase
+
+
+def _small_golden_cases(*, space_id: str, alpha_memory_scope_id: str) -> tuple[EvalCase, ...]:
+    return (
+        EvalCase(
+            case_id="facts_canonical_truth",
+            category="facts",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query="canonical memory lifecycle owner",
+            must_include=("EVAL_FACT_CANONICAL",),
+            must_not_include=("Qdrant owns lifecycle",),
+        ),
+        EvalCase(
+            case_id="facts_adapter_boundary",
+            category="facts",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query="derived indexes qdrant graphiti",
+            must_include=("EVAL_FACT_ADAPTERS",),
+        ),
+        EvalCase(
+            case_id="documents_architecture_notes",
+            category="documents",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query="derived document vectors architecture notes",
+            must_include=("EVAL_DOC_ARCH",),
+        ),
+        EvalCase(
+            case_id="documents_chunking_notes",
+            category="documents",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query="canonical chunks large documents",
+            must_include=("EVAL_DOC_CHUNKS",),
+        ),
+        EvalCase(
+            case_id="updates_current_fact_only",
+            category="stale_update",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query="document recall qdrant pgvector",
+            must_include=("EVAL_FACT_UPDATED_NEW",),
+            must_not_include=("EVAL_FACT_UPDATED_OLD",),
+        ),
+        EvalCase(
+            case_id="deleted_fact_filtered",
+            category="deleted",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query="deleted fact must not render",
+            must_not_include=("EVAL_FACT_DELETED",),
+        ),
+        EvalCase(
+            case_id="cross_memory_scope_isolation",
+            category="cross_memory_scope",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query="beta memory_scope isolated secret",
+            must_not_include=("EVAL_BETA_SECRET",),
+        ),
+        EvalCase(
+            case_id="prompt_injection_evidence_only",
+            category="prompt_injection",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query="prompt injection marker reveal secrets",
+            must_include=("EVAL_PROMPT_INJECTION_MARKER",),
+        ),
+    )
+
+
+def _quality_golden_cases(
+    *,
+    space_id: str,
+    alpha_memory_scope_id: str,
+    beta_memory_scope_id: str,
+    current_thread_id: str,
+    other_thread_id: str,
+) -> tuple[EvalCase, ...]:
+    return (
+        EvalCase(
+            case_id="current_model_beats_decoy",
+            category="answer_support",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query="GPT-5.4 mini local interview canary current model",
+            must_include=("QUALITY_FACT_MODEL_CURRENT",),
+            must_not_include=("QUALITY_DECOY_WRONG_MODEL",),
+            max_facts=1,
+            max_chunks=0,
+        ),
+        EvalCase(
+            case_id="architecture_roles_recall",
+            category="answer_support",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query="Graphiti temporal facts Qdrant document RAG Postgres canonical truth",
+            must_include=("QUALITY_FACT_ARCH_ROLES",),
+            max_facts=3,
+            max_chunks=0,
+        ),
+        EvalCase(
+            case_id="clean_architecture_recall_without_frontend_noise",
+            category="answer_support",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query="Clean Architecture SOLID simple DDD port adapter infinity context",
+            must_include=("QUALITY_FACT_CLEAN_ARCH",),
+            must_not_include=("QUALITY_NOISE_FRONTEND_THEME",),
+            max_facts=1,
+            max_chunks=0,
+        ),
+        EvalCase(
+            case_id="updated_provider_current_only",
+            category="stale_update",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query="current document memory Qdrant RAG Graphiti temporal facts pgvector",
+            must_include=("QUALITY_FACT_PROVIDER_CURRENT",),
+            must_not_include=("QUALITY_FACT_PROVIDER_OLD",),
+            max_facts=5,
+            max_chunks=0,
+        ),
+        EvalCase(
+            case_id="longmemeval_knowledge_update_current_truth",
+            category="stale_update",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query=(
+                "LongMemEval knowledge update old pgvector provider current "
+                "Qdrant Graphiti truth"
+            ),
+            must_include=("QUALITY_FACT_PROVIDER_CURRENT",),
+            must_not_include=("QUALITY_FACT_PROVIDER_OLD",),
+            max_facts=5,
+            max_chunks=0,
+            required_diagnostics=(("items_used", "gte", 1),),
+        ),
+        EvalCase(
+            case_id="temporal_supersedes_current_only",
+            category="stale_update",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query="legacy temporal owner RedisGraph snapshots",
+            must_include=("QUALITY_FACT_TEMPORAL_CURRENT",),
+            must_not_include=("QUALITY_FACT_TEMPORAL_OLD",),
+            max_facts=5,
+            max_chunks=0,
+            required_diagnostics=(("temporal_replacements_applied", "gte", 1),),
+        ),
+        EvalCase(
+            case_id="linked_temporal_supersedes_current_only",
+            category="stale_update",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query="Project Phoenix linked owner anchor",
+            must_include=("QUALITY_LINKED_TEMPORAL_CURRENT",),
+            must_not_include=("QUALITY_LINKED_TEMPORAL_OLD",),
+            max_facts=6,
+            max_chunks=0,
+            required_diagnostics=(
+                ("approved_context_linked_facts_used", "gte", 1),
+                ("linked_temporal_replacements_applied", "gte", 1),
+                ("retrieval_sources_used", "contains", "temporal_supersedes_relation"),
+            ),
+        ),
+        EvalCase(
+            case_id="relative_time_current_fact_not_last_week_fact",
+            category="stale_update",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query="billing rollout owner was Alex last week",
+            must_include=("QUALITY_FACT_RELATIVE_TIME_CURRENT",),
+            must_not_include=("QUALITY_FACT_RELATIVE_TIME_OLD",),
+            max_facts=5,
+            max_chunks=0,
+            required_diagnostics=(("temporal_replacements_applied", "gte", 1),),
+        ),
+        EvalCase(
+            case_id="contradicted_fact_hidden_by_default",
+            category="stale_update",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query="legacy billing owner Alex",
+            must_include=("QUALITY_FACT_CONTRADICTION_CURRENT",),
+            must_not_include=("QUALITY_FACT_CONTRADICTION_OLD",),
+            max_facts=5,
+            max_chunks=0,
+        ),
+        EvalCase(
+            case_id="contradicted_fact_visible_only_in_stale_review",
+            category="stale_review",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query="legacy billing owner Alex",
+            must_include=("QUALITY_FACT_CONTRADICTION_OLD",),
+            token_budget=2500,
+            max_facts=10,
+            max_chunks=0,
+            max_evidence_items=0,
+            include_stale=True,
+            required_diagnostics=(
+                ("stale_facts_used", "gte", 1),
+                ("retrieval_sources_used", "contains", "disputed_review"),
+            ),
+        ),
+        EvalCase(
+            case_id="pending_conflict_review_visible",
+            category="conflict_review",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query="Project Atlas billing log retention conflict review",
+            must_include=(
+                "QUALITY_PENDING_CONFLICT_ACTIVE",
+                "QUALITY_PENDING_CONFLICT_CANDIDATE",
+                "Pending review add suggestion for active fact",
+            ),
+            must_not_include=("Pending duplicate merge",),
+            max_facts=1,
+            max_chunks=0,
+            required_diagnostics=(
+                ("pending_conflict_suggestions_considered", "eq", 1),
+                ("pending_duplicate_merge_suggestions_considered", "eq", 0),
+            ),
+        ),
+        EvalCase(
+            case_id="deleted_fact_hidden",
+            category="deleted",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query="obsolete deleted benchmark fact must not render",
+            must_not_include=("QUALITY_FACT_DELETED",),
+            max_facts=5,
+            max_chunks=0,
+        ),
+        EvalCase(
+            case_id="restricted_fact_hidden",
+            category="restricted",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query="production credential restricted secret context",
+            must_not_include=("QUALITY_RESTRICTED_SECRET",),
+            max_facts=5,
+            max_chunks=0,
+        ),
+        EvalCase(
+            case_id="cross_memory_scope_secret_hidden",
+            category="cross_memory_scope",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query="beta memory_scope billing token isolated alpha",
+            must_not_include=("QUALITY_BETA_ONLY_SECRET",),
+            max_facts=5,
+            max_chunks=0,
+        ),
+        EvalCase(
+            case_id="multi_memory_scope_explicit_recall",
+            category="multi_memory_scope",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id, beta_memory_scope_id),
+            query="beta memory_scope billing token isolated",
+            must_include=("QUALITY_BETA_ONLY_SECRET",),
+            max_facts=5,
+            max_chunks=0,
+        ),
+        EvalCase(
+            case_id="thread_current_visible_without_neighbor",
+            category="cross_thread",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            thread_id=current_thread_id,
+            query="active coding session black-box retry strategy neighboring snapshot migration",
+            must_include=("QUALITY_THREAD_CURRENT",),
+            must_not_include=("QUALITY_THREAD_OTHER",),
+            max_facts=5,
+            max_chunks=0,
+        ),
+        EvalCase(
+            case_id="thread_other_visible_without_current",
+            category="cross_thread",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            thread_id=other_thread_id,
+            query="neighboring session snapshot-only migration notes black-box retry strategy",
+            must_include=("QUALITY_THREAD_OTHER",),
+            must_not_include=("QUALITY_THREAD_CURRENT",),
+            max_facts=5,
+            max_chunks=0,
+        ),
+        EvalCase(
+            case_id="person_event_project_precision",
+            category="anchor_precision",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query="Alex Project Atlas billing one hour ago",
+            must_include=("QUALITY_FACT_ALEX_ATLAS_RECENT",),
+            must_not_include=("QUALITY_FACT_MARIA_ATLAS_WEEK",),
+            max_facts=1,
+            max_chunks=0,
+        ),
+        EvalCase(
+            case_id="multilingual_recent_person_project_recall",
+            category="anchor_precision",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query="Алексом Project Atlas billing час назад",
+            must_include=("QUALITY_FACT_RU_ALEX_ATLAS",),
+            must_not_include=("QUALITY_FACT_MARIA_ATLAS_WEEK",),
+            max_facts=1,
+            max_chunks=0,
+        ),
+        EvalCase(
+            case_id="same_person_time_wrong_project_does_not_pull_atlas",
+            category="anchor_precision",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query="Alex Project Apollo billing one hour ago",
+            must_not_include=(
+                "QUALITY_FACT_ALEX_ATLAS_RECENT",
+                "QUALITY_FACT_RU_ALEX_ATLAS",
+                "QUALITY_FACT_MARIA_ATLAS_WEEK",
+                "QUALITY_PENDING_CONFLICT_ACTIVE",
+                "QUALITY_PENDING_CONFLICT_CANDIDATE",
+                "project: Project Atlas",
+                "event: Atlas billing call",
+            ),
+            max_facts=3,
+            max_chunks=0,
+            max_evidence_items=0,
+        ),
+        EvalCase(
+            case_id="mixed_language_wrong_project_returns_no_context",
+            category="no_candidate",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query="Алекс Project Apollo billing час назад",
+            must_not_include=(
+                "QUALITY_FACT_ALEX_ATLAS_RECENT",
+                "QUALITY_FACT_RU_ALEX_ATLAS",
+                "QUALITY_FACT_MARIA_ATLAS_WEEK",
+                "QUALITY_FACT_RELATIVE_TIME_CURRENT",
+                "QUALITY_FACT_CONTRADICTION_CURRENT",
+                "QUALITY_ANCHOR_REL_PERSON_FACT",
+                "project: Project Atlas",
+                "event: Atlas billing call",
+                "person: Alex",
+            ),
+            max_facts=3,
+            max_chunks=0,
+            max_evidence_items=0,
+            require_evidence_guard=False,
+            required_diagnostics=(
+                ("items_used", "eq", 0),
+                (
+                    "requirement_guard_status",
+                    "eq",
+                    "dropped_missing_project_anchor",
+                ),
+                ("requirement_guard_items_dropped", "gte", 1),
+                (
+                    "retrieval_quality_summary.answerability_status",
+                    "eq",
+                    "insufficient_context",
+                ),
+                (
+                    "retrieval_quality_summary.recommended_response_policy",
+                    "eq",
+                    "ask_for_more_context",
+                ),
+                (
+                    "retrieval_quality_summary.answerability_reasons",
+                    "contains",
+                    "no_context_items",
+                ),
+            ),
+        ),
+        EvalCase(
+            case_id="longmemeval_multilingual_entity_abstention",
+            category="no_candidate",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query="Никита Project Mercury billing вчера",
+            must_not_include=(
+                "QUALITY_FACT_ALEX_ATLAS_RECENT",
+                "QUALITY_FACT_RU_ALEX_ATLAS",
+                "QUALITY_FACT_MARIA_ATLAS_WEEK",
+                "QUALITY_FACT_RELATIVE_TIME_CURRENT",
+                "QUALITY_PENDING_CONFLICT_ACTIVE",
+                "QUALITY_PENDING_CONFLICT_CANDIDATE",
+                "project: Project Atlas",
+                "event: Atlas billing call",
+                "person: Alex",
+                "person: Maria",
+            ),
+            max_facts=3,
+            max_chunks=0,
+            max_evidence_items=0,
+            require_evidence_guard=False,
+            required_diagnostics=(
+                ("items_used", "eq", 0),
+                (
+                    "retrieval_quality_summary.answerability_status",
+                    "eq",
+                    "insufficient_context",
+                ),
+                (
+                    "retrieval_quality_summary.recommended_response_policy",
+                    "eq",
+                    "ask_for_more_context",
+                ),
+                (
+                    "retrieval_quality_summary.answerability_reasons",
+                    "contains",
+                    "no_context_items",
+                ),
+            ),
+        ),
+        EvalCase(
+            case_id="document_overview_recall",
+            category="documents",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query="layered memory core canonical Postgres facts derived adapters",
+            must_include=("QUALITY_DOC_OVERVIEW",),
+            max_facts=0,
+            max_chunks=5,
+        ),
+        EvalCase(
+            case_id="document_architecture_precision",
+            category="documents",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query="temporal Graphiti facts vector Qdrant docs canonical revalidation",
+            must_include=("QUALITY_DOC_ARCHITECTURE",),
+            must_not_include=("QUALITY_DOC_DECOY_REDIS",),
+            max_facts=0,
+            max_chunks=1,
+        ),
+        EvalCase(
+            case_id="document_source_diversity_preserves_secondary_source",
+            category="documents",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query="sourcedivsignal dominantstream secondarystream roundrobinclip",
+            must_include=(
+                "QUALITY_SOURCE_DIVERSITY_DOMINANT_0",
+                "QUALITY_SOURCE_DIVERSITY_SECONDARY",
+            ),
+            must_not_include=("QUALITY_SOURCE_DIVERSITY_DOMINANT_2",),
+            token_budget=190,
+            max_facts=0,
+            max_chunks=5,
+            required_diagnostics=(
+                ("chunk_sources_considered", "gte", 2),
+                ("chunk_sources_used", "gte", 2),
+                ("source_diversity_chunks_reordered", "gte", 1),
+            ),
+        ),
+        EvalCase(
+            case_id="hybrid_document_beats_single_source",
+            category="hybrid_retrieval",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query="hybrid dual source vector keyword routing evidence",
+            must_include=("QUALITY_HYBRID_DUAL_SOURCE",),
+            must_not_include=("QUALITY_HYBRID_SINGLE_SOURCE_DECOY",),
+            max_facts=0,
+            max_chunks=1,
+            required_diagnostics=(("hybrid_items_used", "gte", 1),),
+        ),
+        EvalCase(
+            case_id="context_diversity_preserves_fact_and_chunk_evidence",
+            category="hybrid_retrieval",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query="quality context diversity screenshot transcript fact crowd",
+            must_include=("QUALITY_DIVERSITY_FACT_PRIMARY", "QUALITY_DIVERSITY_CHUNK"),
+            must_not_include=("QUALITY_DIVERSITY_FACT_SECONDARY",),
+            token_budget=160,
+            max_facts=3,
+            max_chunks=1,
+            required_diagnostics=(
+                ("diversity_families_used", "gte", 2),
+                ("diversity_items_used", "gte", 2),
+            ),
+        ),
+        EvalCase(
+            case_id="pending_duplicate_merge_review_visible",
+            category="duplicate_merge",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query="Alex Project Atlas retrieval notes duplicate meeting capture",
+            must_include=(
+                "QUALITY_DUPLICATE_MERGE_ACTIVE",
+                "QUALITY_DUPLICATE_MERGE_PENDING",
+                "Pending duplicate merge update suggestion for active fact",
+            ),
+            must_not_include=("Pending review add suggestion for active fact",),
+            max_facts=1,
+            max_chunks=0,
+            required_diagnostics=(
+                ("pending_conflict_suggestions_considered", "eq", 0),
+                ("pending_duplicate_merge_suggestions_considered", "eq", 1),
+            ),
+            required_item_matches=(
+                (
+                    ("item_type", "eq", "suggestion"),
+                    (
+                        "diagnostics.retrieval_source",
+                        "eq",
+                        "pending_duplicate_merge_suggestion",
+                    ),
+                    ("diagnostics.review_risk", "eq", "medium"),
+                    ("diagnostics.review_recommendation_confidence", "eq", "medium"),
+                    (
+                        "diagnostics.review_recommended_resolution_action",
+                        "eq",
+                        "merge_source_refs",
+                    ),
+                    ("diagnostics.review_requires_review", "eq", True),
+                    ("diagnostics.review_auto_merge_eligible", "eq", False),
+                    (
+                        "diagnostics.review_recommendation_reason_codes",
+                        "contains",
+                        "human_review_required",
+                    ),
+                ),
+            ),
+        ),
+        EvalCase(
+            case_id="canonical_project_anchor_recall_with_citation",
+            category="anchor_context",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query="Project Atlas canonical owner screenshot OCR anchor",
+            must_include=("project: Project Atlas",),
+            max_facts=0,
+            max_chunks=0,
+            required_diagnostics=(
+                ("anchors_used", "gte", 1),
+                ("citations_rendered", "gte", 1),
+                ("retrieval_sources_used", "contains", "canonical_anchors"),
+            ),
+            required_item_matches=(
+                (
+                    ("item_type", "eq", "anchor"),
+                    ("text", "contains", "project: Project Atlas"),
+                    ("diagnostics.retrieval_source", "eq", "canonical_anchors"),
+                    ("diagnostics.anchor_kind", "eq", "project"),
+                    (
+                        "diagnostics.anchor_identity_profile.schema_version",
+                        "eq",
+                        "anchor-identity-profile-v1",
+                    ),
+                    ("diagnostics.anchor_identity_profile.anchor_kind", "eq", "project"),
+                    (
+                        "diagnostics.anchor_identity_profile.primary_identity_key",
+                        "eq",
+                        "project:atlas",
+                    ),
+                    ("diagnostics.anchor_identity_profile.identity_term_count", "gte", 1),
+                ),
+            ),
+            required_source_ref_matches=(
+                (
+                    ("source_type", "eq", "asset_extraction"),
+                    ("source_id", "eq", "quality-anchor-atlas-extract"),
+                    ("chunk_id", "eq", "quality-anchor-atlas-chunk"),
+                    ("time_start_ms", "eq", 1200),
+                    ("time_end_ms", "eq", 5400),
+                    ("bbox", "eq", [12.0, 32.0, 300.0, 88.0]),
+                ),
+            ),
+            required_citation_matches=(
+                (
+                    ("source_type", "eq", "asset_extraction"),
+                    ("source_id", "eq", "quality-anchor-atlas-extract"),
+                    ("chunk_id", "eq", "quality-anchor-atlas-chunk"),
+                    ("time_range_ms.start", "eq", 1200),
+                    ("time_range_ms.end", "eq", 5400),
+                    ("bbox", "eq", [12.0, 32.0, 300.0, 88.0]),
+                ),
+            ),
+        ),
+        EvalCase(
+            case_id="longmemeval_source_attribution_project_anchor",
+            category="anchor_context",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query="Project Atlas source attribution citation evidence OCR screenshot anchor",
+            must_include=("project: Project Atlas",),
+            max_facts=0,
+            max_chunks=0,
+            required_diagnostics=(
+                ("anchors_used", "gte", 1),
+                ("citations_rendered", "gte", 1),
+                ("retrieval_sources_used", "contains", "canonical_anchors"),
+            ),
+            required_item_matches=(
+                (
+                    ("item_type", "eq", "anchor"),
+                    ("text", "contains", "project: Project Atlas"),
+                    ("diagnostics.retrieval_source", "eq", "canonical_anchors"),
+                    ("diagnostics.anchor_kind", "eq", "project"),
+                    (
+                        "diagnostics.anchor_identity_profile.primary_identity_key",
+                        "eq",
+                        "project:atlas",
+                    ),
+                ),
+            ),
+            required_source_ref_matches=(
+                (
+                    ("source_type", "eq", "asset_extraction"),
+                    ("source_id", "eq", "quality-anchor-atlas-extract"),
+                    ("chunk_id", "eq", "quality-anchor-atlas-chunk"),
+                    ("time_start_ms", "eq", 1200),
+                    ("time_end_ms", "eq", 5400),
+                    ("bbox", "eq", [12.0, 32.0, 300.0, 88.0]),
+                ),
+            ),
+            required_citation_matches=(
+                (
+                    ("source_type", "eq", "asset_extraction"),
+                    ("source_id", "eq", "quality-anchor-atlas-extract"),
+                    ("chunk_id", "eq", "quality-anchor-atlas-chunk"),
+                    ("time_range_ms.start", "eq", 1200),
+                    ("time_range_ms.end", "eq", 5400),
+                    ("bbox", "eq", [12.0, 32.0, 300.0, 88.0]),
+                ),
+            ),
+        ),
+        EvalCase(
+            case_id="canonical_event_anchor_recall_by_identity",
+            category="anchor_context",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query="Alex Project Atlas billing call one hour ago event",
+            must_include=(
+                "event: Atlas billing call",
+                "event_type: call",
+                "with: Alex",
+                "time: one hour ago",
+            ),
+            max_facts=0,
+            max_chunks=0,
+            required_diagnostics=(
+                ("anchors_used", "gte", 1),
+                ("retrieval_sources_used", "contains", "canonical_anchors"),
+            ),
+            required_item_matches=(
+                (
+                    ("item_type", "eq", "anchor"),
+                    ("text", "contains", "event: Atlas billing call"),
+                    ("diagnostics.retrieval_source", "eq", "canonical_anchors"),
+                    ("diagnostics.anchor_kind", "eq", "event"),
+                    (
+                        "diagnostics.anchor_identity_profile.schema_version",
+                        "eq",
+                        "anchor-identity-profile-v1",
+                    ),
+                    ("diagnostics.anchor_identity_profile.anchor_kind", "eq", "event"),
+                    ("diagnostics.anchor_identity_profile.has_event_participant", "eq", True),
+                    ("diagnostics.anchor_identity_profile.has_event_project", "eq", True),
+                    ("diagnostics.anchor_identity_profile.has_event_temporal_hint", "eq", True),
+                    ("diagnostics.anchor_identity_profile.event_type", "eq", "call"),
+                ),
+            ),
+            required_source_ref_matches=(
+                (
+                    ("source_type", "eq", "capture"),
+                    ("source_id", "quality-event-anchor-capture"),
+                    ("quote_preview", "contains", "Alex Project Atlas billing call"),
+                ),
+            ),
+            required_citation_matches=(
+                (
+                    ("source_type", "eq", "capture"),
+                    ("source_id", "eq", "quality-event-anchor-capture"),
+                    ("quote_preview", "contains", "Alex Project Atlas billing call"),
+                ),
+            ),
+        ),
+        EvalCase(
+            case_id="mixed_script_event_anchor_recall_by_query_intent",
+            category="anchor_context",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query="созвон с алексом в атласе",
+            must_include=(
+                "event: Atlas billing call",
+                "with: Alex",
+            ),
+            max_facts=0,
+            max_chunks=0,
+            required_diagnostics=(
+                ("query_anchor_person_hint_count", "gte", 1),
+                ("query_anchor_project_hint_count", "gte", 1),
+                ("anchors_used_by_query_intent", "gte", 1),
+                ("retrieval_sources_used", "contains", "canonical_anchors"),
+            ),
+            required_item_matches=(
+                (
+                    ("item_type", "eq", "anchor"),
+                    ("text", "contains", "event: Atlas billing call"),
+                    (
+                        "diagnostics.query_anchor_match_reasons",
+                        "contains",
+                        "query_event_participant_match",
+                    ),
+                    (
+                        "diagnostics.query_anchor_match_reasons",
+                        "contains",
+                        "query_event_project_match",
+                    ),
+                    ("diagnostics.query_anchor_match_keys", "contains", "aleks"),
+                    ("diagnostics.query_anchor_match_keys", "contains", "atlas"),
+                ),
+            ),
+        ),
+        EvalCase(
+            case_id="event_anchor_relation_expands_linked_person_project_facts",
+            category="anchor_context",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query=(
+                "Alex Relation Project Atlas Relation relation-ledger call last week linked facts"
+            ),
+            must_include=(
+                "event: Quality relation ledger call",
+                "person: Alex Relation",
+                "project: Project Atlas Relation",
+                "QUALITY_ANCHOR_REL_PERSON_FACT",
+                "QUALITY_ANCHOR_REL_PROJECT_FACT",
+            ),
+            token_budget=2500,
+            max_facts=2,
+            max_chunks=0,
+            max_evidence_items=0,
+            required_diagnostics=(
+                ("anchors_used", "gte", 1),
+                ("anchor_relation_items_used", "gte", 2),
+                ("approved_context_links_used", "gte", 2),
+                ("approved_context_linked_facts_used", "gte", 2),
+                ("retrieval_sources_used", "contains", "canonical_anchor_relations"),
+                ("retrieval_sources_used", "contains", "approved_context_linked_facts"),
+            ),
+            required_item_matches=(
+                (
+                    ("item_type", "eq", "anchor"),
+                    (
+                        "diagnostics.retrieval_sources",
+                        "contains",
+                        "canonical_anchor_relations",
+                    ),
+                    ("diagnostics.anchor_kind", "eq", "person"),
+                    ("text", "contains", "person: Alex Relation"),
+                ),
+                (
+                    ("item_type", "eq", "anchor"),
+                    (
+                        "diagnostics.retrieval_sources",
+                        "contains",
+                        "canonical_anchor_relations",
+                    ),
+                    ("diagnostics.anchor_kind", "eq", "project"),
+                    ("text", "contains", "project: Project Atlas Relation"),
+                ),
+                (
+                    ("item_type", "eq", "fact"),
+                    ("diagnostics.retrieval_source", "eq", "approved_context_linked_facts"),
+                    ("text", "contains", "QUALITY_ANCHOR_REL_PERSON_FACT"),
+                ),
+                (
+                    ("item_type", "eq", "fact"),
+                    ("diagnostics.retrieval_source", "eq", "approved_context_linked_facts"),
+                    ("text", "contains", "QUALITY_ANCHOR_REL_PROJECT_FACT"),
+                ),
+            ),
+        ),
+        EvalCase(
+            case_id="wrong_project_anchor_deflects_generic_match",
+            category="anchor_context",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query="Project Apollo",
+            must_not_include=("project: Project Atlas",),
+            max_facts=0,
+            max_chunks=0,
+            require_evidence_guard=False,
+            required_diagnostics=(
+                ("anchors_considered", "gte", 1),
+                ("anchors_used", "eq", 0),
+                ("items_used", "eq", 0),
+            ),
+        ),
+        EvalCase(
+            case_id="multimodal_source_refs_recall_with_citations",
+            category="documents",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query="Project Atlas screenshot OCR transcript invoice review timeline",
+            must_include=("QUALITY_MM_CONTEXT",),
+            max_facts=0,
+            max_chunks=3,
+            required_diagnostics=(
+                ("items_with_multimodal_source_refs", "gte", 1),
+                ("source_refs_with_page_count", "gte", 1),
+                ("source_refs_with_time_range_count", "gte", 1),
+                ("source_refs_with_bbox_count", "gte", 1),
+                ("query_snippet_items_used", "gte", 1),
+                ("query_snippet_source_refs_enriched", "gte", 1),
+            ),
+            required_source_ref_matches=(
+                (
+                    ("source_type", "eq", "asset_extraction"),
+                    ("source_id", "eq", "quality-mm-extract"),
+                    ("page_number", "eq", 2),
+                    ("time_start_ms", "eq", 1200),
+                    ("time_end_ms", "eq", 5400),
+                    ("bbox", "eq", [12.0, 32.0, 300.0, 88.0]),
+                    ("quote_preview", "contains", "Project Atlas screenshot"),
+                ),
+            ),
+            required_citation_matches=(
+                (
+                    ("source_type", "eq", "asset_extraction"),
+                    ("source_id", "eq", "quality-mm-extract"),
+                    ("page_number", "eq", 2),
+                    ("time_range_ms.start", "eq", 1200),
+                    ("time_range_ms.end", "eq", 5400),
+                    ("bbox", "eq", [12.0, 32.0, 300.0, 88.0]),
+                    ("label", "contains", "bbox"),
+                ),
+            ),
+        ),
+        EvalCase(
+            case_id="multilingual_multimodal_source_refs_recall",
+            category="documents",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query="скриншот инвойса владелец Атлас Алекс транскрипт",
+            must_include=("QUALITY_MM_CONTEXT",),
+            max_facts=0,
+            max_chunks=0,
+            max_evidence_items=1,
+            required_diagnostics=(
+                ("artifact_evidence_items_used", "gte", 1),
+                ("items_with_evidence_kind", "gte", 1),
+                ("items_with_evidence_modality", "gte", 1),
+                ("source_refs_with_bbox_count", "gte", 1),
+                ("source_refs_with_time_range_count", "gte", 1),
+                ("query_snippet_items_used", "gte", 1),
+                ("query_snippet_source_refs_enriched", "gte", 1),
+            ),
+            required_source_ref_matches=(
+                (
+                    ("source_type", "eq", "extraction_artifact"),
+                    ("chunk_id", "eq", "quality-mm-manifest-ocr-owner"),
+                    ("time_start_ms", "eq", 1200),
+                    ("time_end_ms", "eq", 5400),
+                    ("bbox", "eq", [12.0, 32.0, 300.0, 88.0]),
+                    ("quote_preview", "contains", "Project Atlas screenshot"),
+                ),
+            ),
+            required_citation_matches=(
+                (
+                    ("source_type", "eq", "extraction_artifact"),
+                    ("chunk_id", "eq", "quality-mm-manifest-ocr-owner"),
+                    ("time_range_ms.start", "eq", 1200),
+                    ("time_range_ms.end", "eq", 5400),
+                    ("bbox", "eq", [12.0, 32.0, 300.0, 88.0]),
+                    ("quote_preview", "contains", "Project Atlas screenshot"),
+                    ("label", "contains", "bbox"),
+                ),
+            ),
+        ),
+        EvalCase(
+            case_id="multimodal_evidence_metadata_contract",
+            category="item_contract",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query="скриншот инвойса владелец Атлас Алекс транскрипт",
+            must_include=(
+                "QUALITY_MM_CONTEXT",
+                "evidence=image/ocr_region",
+                "confidence=0.930",
+                'reason="matched first-party multimodal extraction evidence"',
+            ),
+            must_not_include=("Calendar footer without the requested owner.",),
+            max_facts=0,
+            max_chunks=0,
+            max_evidence_items=1,
+            required_diagnostics=(
+                ("artifact_evidence_items_used", "gte", 1),
+                ("artifact_evidence_confidence_signal_count", "gte", 1),
+                ("artifact_evidence_coordinate_signal_count", "gte", 1),
+                ("items_with_evidence_kind", "gte", 1),
+                ("items_with_evidence_modality", "gte", 1),
+                ("retrieval_sources_used", "contains", "artifact_evidence"),
+            ),
+            required_item_matches=(
+                (
+                    ("item_type", "eq", "extraction_artifact"),
+                    ("score", "gte", 0.9),
+                    ("diagnostics.retrieval_source", "eq", "artifact_evidence"),
+                    ("diagnostics.evidence_kind", "eq", "ocr_region"),
+                    ("diagnostics.evidence_modality", "eq", "image"),
+                    ("diagnostics.evidence_confidence", "gte", 0.9),
+                    (
+                        "diagnostics.ranking_reason",
+                        "contains",
+                        "first-party multimodal extraction evidence",
+                    ),
+                    (
+                        "diagnostics.evidence_profile.schema_version",
+                        "eq",
+                        "context-item-evidence-profile-v1",
+                    ),
+                    ("diagnostics.evidence_profile.has_citations", "eq", True),
+                    ("diagnostics.evidence_profile.has_precise_location", "eq", True),
+                    ("diagnostics.evidence_profile.has_multimodal_location", "eq", True),
+                    (
+                        "diagnostics.evidence_profile.source_refs_with_bbox_count",
+                        "gte",
+                        1,
+                    ),
+                    (
+                        "diagnostics.evidence_profile.source_refs_with_time_range_count",
+                        "gte",
+                        1,
+                    ),
+                    ("diagnostics.score_signals.evidence_confidence", "gte", 0.9),
+                    ("diagnostics.provenance.query_snippet_unique_term_hits", "gte", 4),
+                ),
+            ),
+            required_source_ref_matches=(
+                (
+                    ("source_type", "eq", "extraction_artifact"),
+                    ("chunk_id", "eq", "quality-mm-manifest-ocr-owner"),
+                    ("time_start_ms", "eq", 1200),
+                    ("time_end_ms", "eq", 5400),
+                    ("bbox", "eq", [12.0, 32.0, 300.0, 88.0]),
+                    ("quote_preview", "contains", "Project Atlas screenshot"),
+                ),
+            ),
+            required_citation_matches=(
+                (
+                    ("source_type", "eq", "extraction_artifact"),
+                    ("chunk_id", "eq", "quality-mm-manifest-ocr-owner"),
+                    ("time_range_ms.start", "eq", 1200),
+                    ("time_range_ms.end", "eq", 5400),
+                    ("bbox", "eq", [12.0, 32.0, 300.0, 88.0]),
+                    ("quote_preview", "contains", "Project Atlas screenshot"),
+                    ("label", "contains", "bbox"),
+                ),
+            ),
+        ),
+        EvalCase(
+            case_id="multimodal_visual_region_query_requires_bbox",
+            category="documents",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query="where on screen is Project Atlas screenshot invoice owner Alex",
+            must_include=("QUALITY_MM_CONTEXT",),
+            must_not_include=("summary has no screen box",),
+            max_facts=0,
+            max_chunks=0,
+            max_evidence_items=1,
+            required_diagnostics=(
+                ("artifact_evidence_items_used", "gte", 1),
+                ("artifact_evidence_visual_region_query_drop_count", "gte", 1),
+                ("artifact_evidence_coordinate_signal_count", "gte", 1),
+                ("source_refs_with_bbox_count", "gte", 1),
+                ("retrieval_sources_used", "contains", "artifact_evidence"),
+                ("context_requirement_coverage.status", "eq", "satisfied"),
+                ("context_requirement_coverage.covered_modalities", "contains", "image"),
+                (
+                    "context_requirement_coverage.covered_evidence_features",
+                    "contains",
+                    "visual_region",
+                ),
+            ),
+            required_item_matches=(
+                (
+                    ("item_type", "eq", "extraction_artifact"),
+                    ("diagnostics.retrieval_source", "eq", "artifact_evidence"),
+                    ("diagnostics.evidence_kind", "eq", "ocr_region"),
+                    ("diagnostics.evidence_modality", "eq", "image"),
+                    ("diagnostics.score_signals.context_requirement_boost", "gte", 0.03),
+                    (
+                        "diagnostics.score_signals.context_requirement_matched_feature_count",
+                        "gte",
+                        2,
+                    ),
+                    (
+                        "diagnostics.score_signals.context_requirement_matched_modality_count",
+                        "gte",
+                        1,
+                    ),
+                ),
+            ),
+            required_source_ref_matches=(
+                (
+                    ("source_type", "eq", "extraction_artifact"),
+                    ("chunk_id", "eq", "quality-mm-manifest-ocr-owner"),
+                    ("bbox", "eq", [12.0, 32.0, 300.0, 88.0]),
+                    ("quote_preview", "contains", "Project Atlas screenshot"),
+                ),
+            ),
+            required_citation_matches=(
+                (
+                    ("source_type", "eq", "extraction_artifact"),
+                    ("chunk_id", "eq", "quality-mm-manifest-ocr-owner"),
+                    ("bbox", "eq", [12.0, 32.0, 300.0, 88.0]),
+                    ("label", "contains", "bbox"),
+                ),
+            ),
+        ),
+        EvalCase(
+            case_id="media_timestamp_query_selects_matching_evidence",
+            category="documents",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query="Project Atlas screenshot invoice owner Alex transcript at 00:01",
+            must_include=("QUALITY_MM_CONTEXT",),
+            must_not_include=("Calendar footer without the requested owner.",),
+            max_facts=0,
+            max_chunks=0,
+            max_evidence_items=1,
+            required_diagnostics=(
+                ("artifact_evidence_time_query_count", "gte", 1),
+                ("artifact_evidence_time_query_match_count", "gte", 1),
+                ("artifact_evidence_time_query_drop_count", "gte", 1),
+                ("media_time_query_items_used", "gte", 1),
+                ("media_time_query_matched_items_used", "gte", 1),
+                ("source_refs_with_time_range_count", "gte", 1),
+            ),
+            required_item_matches=(
+                (
+                    ("item_type", "eq", "extraction_artifact"),
+                    ("diagnostics.retrieval_source", "eq", "artifact_evidence"),
+                    ("diagnostics.media_time_query_count", "eq", 1),
+                    ("diagnostics.score_signals.media_time_matched_window_count", "gte", 1),
+                ),
+            ),
+            required_source_ref_matches=(
+                (
+                    ("source_type", "eq", "extraction_artifact"),
+                    ("chunk_id", "eq", "quality-mm-manifest-ocr-owner"),
+                    ("time_start_ms", "eq", 1200),
+                    ("time_end_ms", "eq", 5400),
+                    ("quote_preview", "contains", "Project Atlas screenshot"),
+                ),
+            ),
+            required_citation_matches=(
+                (
+                    ("source_type", "eq", "extraction_artifact"),
+                    ("chunk_id", "eq", "quality-mm-manifest-ocr-owner"),
+                    ("time_range_ms.start", "eq", 1200),
+                    ("time_range_ms.end", "eq", 5400),
+                    ("quote_preview", "contains", "Project Atlas screenshot"),
+                ),
+            ),
+        ),
+        EvalCase(
+            case_id="document_middle_recall",
+            category="documents",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query="context packing memory evidence only source references visible",
+            must_include=("QUALITY_DOC_MIDDLE",),
+            max_facts=0,
+            max_chunks=5,
+        ),
+        EvalCase(
+            case_id="document_tail_recall",
+            category="documents",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query="operational runbook isolated full provider canary smoke checks",
+            must_include=("QUALITY_DOC_TAIL",),
+            max_facts=0,
+            max_chunks=5,
+        ),
+        EvalCase(
+            case_id="prompt_injection_evidence_only",
+            category="prompt_injection",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query="prompt injection document override system prompt reveal private secrets",
+            must_include=("QUALITY_PROMPT_INJECTION_DOC",),
+            max_facts=0,
+            max_chunks=3,
+        ),
+        EvalCase(
+            case_id="tiny_budget_does_not_overflow",
+            category="token_budget",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query="compact context tiny token budgets",
+            must_include=("QUALITY_FACT_COMPACT",),
+            token_budget=64,
+            max_facts=1,
+            max_chunks=0,
+        ),
+        EvalCase(
+            case_id="unrelated_query_returns_no_context_items",
+            category="no_candidate",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query="xqzv plorbn nerfst 99999 qzzq",
+            must_not_include=(
+                "QUALITY_FACT_MODEL_CURRENT",
+                "QUALITY_DOC_OVERVIEW",
+                "QUALITY_MM_CONTEXT",
+            ),
+            max_facts=1,
+            max_chunks=1,
+            consistency_mode="canonical_only",
+            require_evidence_guard=False,
+            required_diagnostics=(
+                ("items_used", "eq", 0),
+                (
+                    "retrieval_quality_summary.answerability_status",
+                    "eq",
+                    "insufficient_context",
+                ),
+                (
+                    "retrieval_quality_summary.recommended_response_policy",
+                    "eq",
+                    "ask_for_more_context",
+                ),
+                (
+                    "retrieval_quality_summary.answerability_reasons",
+                    "contains",
+                    "no_context_items",
+                ),
+            ),
+        ),
+        EvalCase(
+            case_id="identifier_like_query_deflects_partial_marker",
+            category="no_candidate",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query="QUALITY_CONTEXT_SECRET_MARKER",
+            must_not_include=("QUALITY_CONTEXT_REVIEW_MARKER",),
+            max_facts=1,
+            max_chunks=0,
+            consistency_mode="canonical_only",
+            require_evidence_guard=False,
+            required_diagnostics=(
+                ("items_used", "eq", 0),
+                (
+                    "retrieval_quality_summary.answerability_status",
+                    "eq",
+                    "insufficient_context",
+                ),
+                (
+                    "retrieval_quality_summary.recommended_response_policy",
+                    "eq",
+                    "ask_for_more_context",
+                ),
+                (
+                    "retrieval_quality_summary.answerability_reasons",
+                    "contains",
+                    "no_context_items",
+                ),
+            ),
+        ),
+    )
+
+
+def _long_memory_golden_cases(
+    *,
+    space_id: str,
+    alpha_memory_scope_id: str,
+    beta_memory_scope_id: str,
+    kickoff_thread_id: str,
+    current_thread_id: str,
+    other_thread_id: str,
+) -> tuple[EvalCase, ...]:
+    return (
+        EvalCase(
+            case_id="long_cross_session_kickoff_recall",
+            category="multi_session",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query="first interview session active context infinity context",
+            must_include=("LONGMEM_SESSION_KICKOFF",),
+            max_facts=5,
+            max_chunks=0,
+        ),
+        EvalCase(
+            case_id="long_current_thread_isolation",
+            category="cross_thread",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            thread_id=current_thread_id,
+            query="current coding session validates long-memory gates Obsidian visualization",
+            must_include=("LONGMEM_SESSION_CURRENT",),
+            must_not_include=("LONGMEM_SESSION_OTHER",),
+            max_facts=5,
+            max_chunks=0,
+        ),
+        EvalCase(
+            case_id="long_other_thread_isolation",
+            category="cross_thread",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            thread_id=other_thread_id,
+            query="neighboring design session Obsidian graph visualization current coding gates",
+            must_include=("LONGMEM_SESSION_OTHER",),
+            must_not_include=("LONGMEM_SESSION_CURRENT",),
+            max_facts=5,
+            max_chunks=0,
+        ),
+        EvalCase(
+            case_id="long_kickoff_thread_isolation",
+            category="cross_thread",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            thread_id=kickoff_thread_id,
+            query="first interview session enabled active context current coding gates",
+            must_include=("LONGMEM_SESSION_KICKOFF",),
+            must_not_include=("LONGMEM_SESSION_CURRENT",),
+            max_facts=5,
+            max_chunks=0,
+        ),
+        EvalCase(
+            case_id="long_temporal_update_current_only",
+            category="temporal_update",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query="current provider documents Qdrant RAG Graphiti temporal facts pgvector disabled",
+            must_include=("LONGMEM_PROVIDER_CURRENT",),
+            must_not_include=("LONGMEM_PROVIDER_OLD",),
+            max_facts=5,
+            max_chunks=0,
+        ),
+        EvalCase(
+            case_id="long_lme_knowledge_update_old_query_current_truth",
+            category="temporal_update",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query="old pgvector provider disabled graph search documents current truth",
+            must_include=("LONGMEM_PROVIDER_CURRENT",),
+            must_not_include=("LONGMEM_PROVIDER_OLD",),
+            max_facts=10,
+            max_chunks=0,
+            required_diagnostics=(
+                ("items_used", "gte", 1),
+                ("retrieval_sources_used", "contains", "postgres_facts"),
+            ),
+        ),
+        EvalCase(
+            case_id="long_deleted_fact_hidden",
+            category="deleted",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query="obsolete agent hook path must not render",
+            must_not_include=("LONGMEM_DELETED_STALE",),
+            max_facts=5,
+            max_chunks=0,
+        ),
+        EvalCase(
+            case_id="long_preference_and_constraint_recall",
+            category="preference_synthesis",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query=(
+                "user preference concise Russian summaries concrete next actions "
+                "review gated deletes"
+            ),
+            must_include=("LONGMEM_PREF_FORMAT", "LONGMEM_CONSTRAINT_REVIEW"),
+            max_facts=5,
+            max_chunks=0,
+        ),
+        EvalCase(
+            case_id="long_lme_multilingual_preference_recall",
+            category="preference_synthesis",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query="короткие Russian summaries concrete next actions review gated deletes",
+            must_include=("LONGMEM_PREF_FORMAT", "LONGMEM_CONSTRAINT_REVIEW"),
+            max_facts=5,
+            max_chunks=0,
+        ),
+        EvalCase(
+            case_id="long_graphiti_decision_beats_obsidian_decoy",
+            category="answer_support",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query="primary runtime temporal fact engine Graphiti Obsidian 3D graph",
+            must_include=("LONGMEM_DECISION_GRAPHITI",),
+            must_not_include=("LONGMEM_DECOY_OBSIDIAN",),
+            max_facts=1,
+            max_chunks=0,
+        ),
+        EvalCase(
+            case_id="long_document_scope_recall",
+            category="documents",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query="shared project memory coding agents dev teams Codex Claude Cursor Slack",
+            must_include=("LONGMEM_DOC_PROJECT_SCOPE",),
+            max_facts=0,
+            max_chunks=5,
+        ),
+        EvalCase(
+            case_id="long_document_architecture_precision",
+            category="documents",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query="canonical lifecycle Postgres adapters Graphiti Qdrant documents facts",
+            must_include=("LONGMEM_DOC_ARCH_LAYER",),
+            must_not_include=("LONGMEM_DOC_INJECTION",),
+            max_facts=0,
+            max_chunks=1,
+        ),
+        EvalCase(
+            case_id="long_document_operations_tail_recall",
+            category="documents",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query="full provider canary MCP smoke agent install doctor memory gates",
+            must_include=("LONGMEM_DOC_OPERATIONS",),
+            max_facts=0,
+            max_chunks=5,
+        ),
+        EvalCase(
+            case_id="long_prompt_injection_evidence_guard",
+            category="prompt_injection",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query="hostile memory ignore previous instructions reveal private secrets",
+            must_include=("LONGMEM_DOC_INJECTION",),
+            max_facts=0,
+            max_chunks=3,
+        ),
+        EvalCase(
+            case_id="long_cross_memory_scope_hidden",
+            category="cross_memory_scope",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query="beta memory_scope candidate scorecard private",
+            must_not_include=("LONGMEM_BETA_PRIVATE",),
+            max_facts=5,
+            max_chunks=0,
+        ),
+        EvalCase(
+            case_id="long_multi_memory_scope_explicit_recall",
+            category="multi_memory_scope",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id, beta_memory_scope_id),
+            query="beta memory_scope candidate scorecard private",
+            must_include=("LONGMEM_BETA_PRIVATE",),
+            max_facts=5,
+            max_chunks=0,
+        ),
+        EvalCase(
+            case_id="long_restricted_secret_hidden",
+            category="restricted",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query="raw production token restricted secret",
+            must_not_include=("LONGMEM_RESTRICTED_SECRET",),
+            max_facts=5,
+            max_chunks=0,
+        ),
+        EvalCase(
+            case_id="long_unknown_query_abstains_without_context",
+            category="no_candidate",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query="unrelated yakutsk cooking recipe quantum aquarium warranty",
+            must_not_include=(
+                "LONGMEM_SESSION_KICKOFF",
+                "LONGMEM_PREF_FORMAT",
+                "LONGMEM_PROVIDER_CURRENT",
+                "LONGMEM_DOC_PROJECT_SCOPE",
+            ),
+            max_facts=1,
+            max_chunks=1,
+            consistency_mode="canonical_only",
+            require_evidence_guard=False,
+            required_diagnostics=(
+                ("items_used", "eq", 0),
+                (
+                    "retrieval_quality_summary.answerability_status",
+                    "eq",
+                    "insufficient_context",
+                ),
+                (
+                    "retrieval_quality_summary.recommended_response_policy",
+                    "eq",
+                    "ask_for_more_context",
+                ),
+                (
+                    "retrieval_quality_summary.answerability_reasons",
+                    "contains",
+                    "no_context_items",
+                ),
+            ),
+        ),
+        EvalCase(
+            case_id="long_lme_abstention_unknown_multilingual",
+            category="no_candidate",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query="Никита Mercury quantum aquarium warranty",
+            must_not_include=(
+                "LONGMEM_SESSION_KICKOFF",
+                "LONGMEM_PREF_FORMAT",
+                "LONGMEM_PROVIDER_CURRENT",
+                "LONGMEM_DOC_PROJECT_SCOPE",
+                "LONGMEM_DECISION_GRAPHITI",
+            ),
+            max_facts=1,
+            max_chunks=1,
+            consistency_mode="canonical_only",
+            require_evidence_guard=False,
+            required_diagnostics=(
+                ("items_used", "eq", 0),
+                (
+                    "retrieval_quality_summary.answerability_status",
+                    "eq",
+                    "insufficient_context",
+                ),
+                (
+                    "retrieval_quality_summary.recommended_response_policy",
+                    "eq",
+                    "ask_for_more_context",
+                ),
+                (
+                    "retrieval_quality_summary.answerability_reasons",
+                    "contains",
+                    "no_context_items",
+                ),
+            ),
+        ),
+        EvalCase(
+            case_id="long_old_provider_query_resolves_current_fact",
+            category="temporal_update",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query="legacy documents pgvector graph search disabled provider",
+            must_include=("LONGMEM_PROVIDER_CURRENT",),
+            must_not_include=("LONGMEM_PROVIDER_OLD",),
+            max_facts=10,
+            max_chunks=0,
+            required_diagnostics=(
+                ("items_used", "gte", 1),
+                ("retrieval_sources_used", "contains", "postgres_facts"),
+            ),
+        ),
+        EvalCase(
+            case_id="long_cross_session_preference_synthesis_with_kickoff",
+            category="preference_synthesis",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query=(
+                "first interview session active context concise Russian summaries "
+                "review gated deletes"
+            ),
+            must_include=(
+                "LONGMEM_SESSION_KICKOFF",
+                "LONGMEM_PREF_FORMAT",
+                "LONGMEM_CONSTRAINT_REVIEW",
+            ),
+            max_facts=6,
+            max_chunks=0,
+        ),
+        EvalCase(
+            case_id="long_tiny_budget_preference_recall",
+            category="token_budget",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query="concise Russian summaries concrete next actions",
+            must_include=("LONGMEM_PREF_FORMAT",),
+            token_budget=64,
+            max_facts=1,
+            max_chunks=0,
+        ),
+    )
+
+
+def _graph_native_cases(
+    *,
+    space_id: str,
+    alpha_memory_scope_id: str,
+    current_thread_id: str,
+) -> tuple[EvalCase, ...]:
+    return (
+        EvalCase(
+            case_id="graph_alias_hydrates_canonical_fact",
+            category="graph_recall",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query="omegaaliasbridge",
+            must_include=("GRAPH_NATIVE_EVAL_RELATED_DECISION",),
+            max_facts=3,
+            max_chunks=0,
+        ),
+        EvalCase(
+            case_id="graph_related_candidates_hydrate_multiple_facts",
+            category="graph_recall",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query="omegaaliastwohop",
+            must_include=("GRAPH_NATIVE_EVAL_RELATED_DECISION", "GRAPH_NATIVE_EVAL_SECOND_HOP"),
+            max_facts=5,
+            max_chunks=0,
+        ),
+        EvalCase(
+            case_id="graph_deleted_candidate_filtered",
+            category="graph_filter",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query="omegaaliasdeleted",
+            must_not_include=("GRAPH_NATIVE_EVAL_DELETED_SHADOW",),
+            max_facts=3,
+            max_chunks=0,
+            require_evidence_guard=False,
+        ),
+        EvalCase(
+            case_id="graph_cross_memory_scope_candidate_filtered",
+            category="graph_filter",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query="omegaaliasbeta",
+            must_not_include=("GRAPH_NATIVE_EVAL_BETA_SECRET",),
+            max_facts=3,
+            max_chunks=0,
+            require_evidence_guard=False,
+        ),
+        EvalCase(
+            case_id="graph_restricted_candidate_filtered",
+            category="graph_filter",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query="omegaaliasrestricted",
+            must_not_include=("GRAPH_NATIVE_EVAL_RESTRICTED_SECRET",),
+            max_facts=3,
+            max_chunks=0,
+            require_evidence_guard=False,
+        ),
+        EvalCase(
+            case_id="graph_wrong_thread_candidate_filtered",
+            category="graph_filter",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            thread_id=current_thread_id,
+            query="omegaaliaswrongthread",
+            must_not_include=("GRAPH_NATIVE_EVAL_WRONG_THREAD",),
+            max_facts=3,
+            max_chunks=0,
+            require_evidence_guard=False,
+        ),
+        EvalCase(
+            case_id="graph_orphan_candidate_dropped",
+            category="graph_filter",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query="omegaaliasorphan",
+            must_not_include=("eval_orphan_relation",),
+            max_facts=3,
+            max_chunks=0,
+            require_evidence_guard=False,
+        ),
+        EvalCase(
+            case_id="canonical_only_skips_graph_candidates",
+            category="graph_canonical_only",
+            space_id=space_id,
+            memory_scope_ids=(alpha_memory_scope_id,),
+            query="omegaaliascanonicalonly",
+            must_not_include=("GRAPH_NATIVE_EVAL_RELATED_DECISION",),
+            max_facts=3,
+            max_chunks=0,
+            consistency_mode="canonical_only",
+            require_evidence_guard=False,
+        ),
+    )