test(harness): caractérisation totale du cœur stateful run_orchestrator

Précondition Phase B (décomposition stateful) demandée : verrouille
le comportement EXACT actuel là où Phase B va toucher. 17 tests,
5 groupes = 5 cas de risque :

1. doc_idx GLOBAL au run sur multi-pipeline (pas par-pipeline)
2. cancel pré-set ET mi-run (identité d'objet event à travers les
couches)
3. interruption réelle → resume (résultats/doc identiques au run
propre) + non-duplication
4. golden snapshot DÉTERMINISTE (manifest+pipeline+view+artifacts
index) sur 3 topologies (linéaire / multi-pipeline / DAG
branchant) + garde anti-flaky + déterminisme CROSS-PROCESS prouvé
(run_id scrubé, listes set-dérivées triées — sinon golden flaky
via randomisation hash chaînes)
5. isolation concurrente (2 execute() en threads : annuler A ne
fuit pas dans B)

⚠️ Le harnais a découvert un DÉFAUT pré-existant : au resume, le
partial store rejoue pipeline_results mais PAS view_results des docs
repris → view_results.jsonl incomplet → métriques agrégées
silencieusement faussées après reprise. Topologie-dépendant
(linéaire/DAG : vues ⊊ pipeline ; multi-pipeline : pas manifesté).
CARACTÉRISÉ tel quel (rôle d'un harnais : figer la réalité, warts
inclus) — fix traité en commit séparé.

https://claude.ai/code/session_01EmLiMPJJuB44QHEFzDWUvF

tests/golden/fixtures/run_orchestrator/branching_dag.json

@@ -0,0 +1,528 @@
+{
+  "artifacts_index": [
+    {
+      "content_hash": null,
+      "document_id": "doc01",
+      "id": "doc01:image",
+      "pipeline_name": "ocr_then_correct",
+      "produced_by_step": null,
+      "provenance": null,
+      "type": "image",
+      "uri": "<PATH>"
+    },
+    {
+      "content_hash": null,
+      "document_id": "doc01",
+      "id": "doc01:precomputed_corr:raw_text",
+      "pipeline_name": "ocr_then_correct",
+      "produced_by_step": "ocr",
+      "provenance": null,
+      "type": "raw_text",
+      "uri": "<PATH>"
+    },
+    {
+      "content_hash": null,
+      "document_id": "doc01",
+      "id": "doc01:precomputed_tess:raw_text",
+      "pipeline_name": "ocr_then_correct",
+      "produced_by_step": "ocr",
+      "provenance": null,
+      "type": "raw_text",
+      "uri": "<PATH>"
+    },
+    {
+      "content_hash": null,
+      "document_id": "doc02",
+      "id": "doc02:image",
+      "pipeline_name": "ocr_then_correct",
+      "produced_by_step": null,
+      "provenance": null,
+      "type": "image",
+      "uri": "<PATH>"
+    },
+    {
+      "content_hash": null,
+      "document_id": "doc02",
+      "id": "doc02:precomputed_corr:raw_text",
+      "pipeline_name": "ocr_then_correct",
+      "produced_by_step": "ocr",
+      "provenance": null,
+      "type": "raw_text",
+      "uri": "<PATH>"
+    },
+    {
+      "content_hash": null,
+      "document_id": "doc02",
+      "id": "doc02:precomputed_tess:raw_text",
+      "pipeline_name": "ocr_then_correct",
+      "produced_by_step": "ocr",
+      "provenance": null,
+      "type": "raw_text",
+      "uri": "<PATH>"
+    },
+    {
+      "content_hash": null,
+      "document_id": "doc03",
+      "id": "doc03:image",
+      "pipeline_name": "ocr_then_correct",
+      "produced_by_step": null,
+      "provenance": null,
+      "type": "image",
+      "uri": "<PATH>"
+    },
+    {
+      "content_hash": null,
+      "document_id": "doc03",
+      "id": "doc03:precomputed_corr:raw_text",
+      "pipeline_name": "ocr_then_correct",
+      "produced_by_step": "ocr",
+      "provenance": null,
+      "type": "raw_text",
+      "uri": "<PATH>"
+    },
+    {
+      "content_hash": null,
+      "document_id": "doc03",
+      "id": "doc03:precomputed_tess:raw_text",
+      "pipeline_name": "ocr_then_correct",
+      "produced_by_step": "ocr",
+      "provenance": null,
+      "type": "raw_text",
+      "uri": "<PATH>"
+    }
+  ],
+  "manifest": {
+    "adapter_kwargs": {
+      "corrector": {
+        "source_label": "corr"
+      },
+      "ocr": {
+        "source_label": "tess"
+      }
+    },
+    "code_version": "charac-1.0",
+    "corpus_name": "charac",
+    "dependencies_lock": {
+      "CacheControl": "0.14.4",
+      "Jinja2": "3.1.6",
+      "MarkupSafe": "3.0.3",
+      "PyGObject": "3.48.2",
+      "PyJWT": "2.7.0",
+      "PyYAML": "6.0.1",
+      "Pygments": "2.20.0",
+      "RapidFuzz": "3.14.5",
+      "annotated-doc": "0.0.4",
+      "annotated-types": "0.7.0",
+      "anyio": "4.13.0",
+      "argcomplete": "3.1.4",
+      "ast_serialize": "0.5.0",
+      "bandit": "1.9.4",
+      "blinker": "1.7.0",
+      "boolean.py": "5.0",
+      "certifi": "2026.2.25",
+      "charset-normalizer": "3.4.6",
+      "click": "8.4.0",
+      "colorama": "0.4.6",
+      "conan": "2.27.0",
+      "coverage": "7.14.0",
+      "cryptography": "41.0.7",
+      "cyclonedx-python-lib": "11.7.0",
+      "dbus-python": "1.3.2",
+      "defusedxml": "0.7.1",
+      "distro": "1.9.0",
+      "fastapi": "0.136.1",
+      "fasteners": "0.20",
+      "filelock": "3.29.0",
+      "h11": "0.16.0",
+      "httpcore": "1.0.9",
+      "httplib2": "0.20.4",
+      "httptools": "0.7.1",
+      "httpx": "0.28.1",
+      "idna": "3.11",
+      "iniconfig": "2.3.0",
+      "jiwer": "4.0.0",
+      "launchpadlib": "1.11.0",
+      "lazr.restfulclient": "0.14.6",
+      "lazr.uri": "1.0.6",
+      "librt": "0.11.0",
+      "license-expression": "30.4.4",
+      "markdown-it-py": "4.2.0",
+      "mdurl": "0.1.2",
+      "msgpack": "1.1.2",
+      "mypy": "2.1.0",
+      "mypy_extensions": "1.1.0",
+      "numpy": "2.4.6",
+      "oauthlib": "3.2.2",
+      "packageurl-python": "0.17.6",
+      "packaging": "24.0",
+      "patch-ng": "1.18.1",
+      "pathspec": "1.1.1",
+      "picarones": "1.1.0.dev311",
+      "pillow": "12.2.0",
+      "pip": "24.0",
+      "pip-api": "0.0.34",
+      "pip-requirements-parser": "32.0.1",
+      "pip_audit": "2.10.0",
+      "platformdirs": "4.9.6",
+      "pluggy": "1.6.0",
+      "py-serializable": "2.1.0",
+      "pydantic": "2.13.4",
+      "pydantic_core": "2.46.4",
+      "pyparsing": "3.1.1",
+      "pytesseract": "0.3.13",
+      "pytest": "9.0.3",
+      "pytest-cov": "7.1.0",
+      "pytest-timeout": "2.4.0",
+      "python-apt": "2.7.7+ubuntu5.2",
+      "python-dateutil": "2.9.0.post0",
+      "python-dotenv": "1.2.2",
+      "python-multipart": "0.0.29",
+      "requests": "2.33.1",
+      "rich": "15.0.0",
+      "setuptools": "68.1.2",
+      "six": "1.16.0",
+      "sortedcontainers": "2.4.0",
+      "starlette": "1.0.0",
+      "stevedore": "5.8.0",
+      "toml": "0.10.2",
+      "tomli": "2.4.1",
+      "tomli_w": "1.2.0",
+      "tqdm": "4.67.3",
+      "typing-inspection": "0.4.2",
+      "typing_extensions": "4.15.0",
+      "urllib3": "2.6.3",
+      "uvicorn": "0.47.0",
+      "uvloop": "0.22.1",
+      "wadllib": "1.3.6",
+      "watchfiles": "1.2.0",
+      "websockets": "16.0",
+      "wheel": "0.42.0",
+      "xmltodict": "0.13.0",
+      "yq": "3.1.0"
+    },
+    "metadata": {
+      "orchestrator": "picarones.app.services.run_orchestrator"
+    },
+    "n_documents": 3,
+    "pipeline_names": [
+      "ocr_then_correct"
+    ],
+    "pipeline_specs": [
+      {
+        "description": "",
+        "initial_inputs": [
+          "image"
+        ],
+        "name": "ocr_then_correct",
+        "steps": [
+          {
+            "adapter_name": "ocr",
+            "id": "ocr",
+            "input_types": [
+              "image"
+            ],
+            "inputs_from": {},
+            "kind": "step",
+            "output_types": [
+              "raw_text"
+            ],
+            "params": {}
+          },
+          {
+            "adapter_name": "corrector",
+            "id": "corrector",
+            "input_types": [
+              "image",
+              "raw_text"
+            ],
+            "inputs_from": {
+              "raw_text": "ocr"
+            },
+            "kind": "step",
+            "output_types": [
+              "corrected_text"
+            ],
+            "params": {}
+          }
+        ]
+      }
+    ],
+    "system_binaries_lock": {},
+    "view_specs": [
+      {
+        "candidate_types": [
+          "alto_xml",
+          "canonical_document",
+          "corrected_text",
+          "page_xml",
+          "raw_text"
+        ],
+        "char_exclude": null,
+        "description": "Compare les sorties textuelles finales après projection éventuelle (ALTO/PAGE/markdown → texte plat).",
+        "ignored_dimensions": [
+          "block_structure",
+          "confidence",
+          "formatting",
+          "geometry",
+          "ids",
+          "reading_order"
+        ],
+        "metric_names": [
+          "cer",
+          "mer",
+          "wer",
+          "wil"
+        ],
+        "name": "text_final",
+        "normalization_profile": null,
+        "projection": null,
+        "projections_by_source_type": {
+          "alto_xml": {
+            "params": {},
+            "projector_name": "alto_to_text",
+            "source_type": "alto_xml",
+            "target_type": "raw_text"
+          },
+          "canonical_document": {
+            "params": {},
+            "projector_name": "canonical_to_text",
+            "source_type": "canonical_document",
+            "target_type": "raw_text"
+          },
+          "page_xml": {
+            "params": {},
+            "projector_name": "page_to_text",
+            "source_type": "page_xml",
+            "target_type": "raw_text"
+          }
+        },
+        "warnings": [
+          "Cette vue compare les sorties textuelles finales après projection éventuelle.  Les pipelines qui produisent ALTO/PAGE/markdown sont projetés vers du texte plat — leurs structures spatiale et documentaire ne sont PAS évaluées ici.  Pour évaluer la qualité ALTO, voir AltoView (S15)."
+        ]
+      }
+    ]
+  },
+  "pipeline_results": [
+    {
+      "document_id": "doc01",
+      "pipeline_name": "ocr_then_correct",
+      "step_results": [
+        {
+          "error": null,
+          "produced_artifacts": {
+            "raw_text": "doc01:precomputed_tess:raw_text"
+          },
+          "step_id": "ocr",
+          "succeeded": true
+        },
+        {
+          "error": "missing_output: ['corrected_text']",
+          "produced_artifacts": {},
+          "step_id": "corrector",
+          "succeeded": false
+        }
+      ],
+      "succeeded": false
+    },
+    {
+      "document_id": "doc02",
+      "pipeline_name": "ocr_then_correct",
+      "step_results": [
+        {
+          "error": null,
+          "produced_artifacts": {
+            "raw_text": "doc02:precomputed_tess:raw_text"
+          },
+          "step_id": "ocr",
+          "succeeded": true
+        },
+        {
+          "error": "missing_output: ['corrected_text']",
+          "produced_artifacts": {},
+          "step_id": "corrector",
+          "succeeded": false
+        }
+      ],
+      "succeeded": false
+    },
+    {
+      "document_id": "doc03",
+      "pipeline_name": "ocr_then_correct",
+      "step_results": [
+        {
+          "error": null,
+          "produced_artifacts": {
+            "raw_text": "doc03:precomputed_tess:raw_text"
+          },
+          "step_id": "ocr",
+          "succeeded": true
+        },
+        {
+          "error": "missing_output: ['corrected_text']",
+          "produced_artifacts": {},
+          "step_id": "corrector",
+          "succeeded": false
+        }
+      ],
+      "succeeded": false
+    }
+  ],
+  "view_results": [
+    {
+      "candidate_artifact_id": "doc01:precomputed_tess:raw_text",
+      "document_id": "doc01",
+      "failed_metrics": {},
+      "ground_truth_artifact_id": "doc01:gt:raw_text",
+      "ignored_dimensions": [
+        "block_structure",
+        "confidence",
+        "formatting",
+        "geometry",
+        "ids",
+        "reading_order"
+      ],
+      "metric_values": {
+        "cer": 0.05,
+        "mer": 0.25,
+        "wer": 0.25,
+        "wil": 0.4375
+      },
+      "pipeline_name": "ocr_then_correct",
+      "projection_report": null,
+      "view_name": "text_final",
+      "warnings": [
+        "Cette vue compare les sorties textuelles finales après projection éventuelle.  Les pipelines qui produisent ALTO/PAGE/markdown sont projetés vers du texte plat — leurs structures spatiale et documentaire ne sont PAS évaluées ici.  Pour évaluer la qualité ALTO, voir AltoView (S15)."
+      ]
+    },
+    {
+      "candidate_artifact_id": "doc01:precomputed_corr:raw_text",
+      "document_id": "doc01",
+      "failed_metrics": {},
+      "ground_truth_artifact_id": "doc01:gt:raw_text",
+      "ignored_dimensions": [
+        "block_structure",
+        "confidence",
+        "formatting",
+        "geometry",
+        "ids",
+        "reading_order"
+      ],
+      "metric_values": {
+        "cer": 0.05,
+        "mer": 0.25,
+        "wer": 0.25,
+        "wil": 0.4375
+      },
+      "pipeline_name": "ocr_then_correct",
+      "projection_report": null,
+      "view_name": "text_final",
+      "warnings": [
+        "Cette vue compare les sorties textuelles finales après projection éventuelle.  Les pipelines qui produisent ALTO/PAGE/markdown sont projetés vers du texte plat — leurs structures spatiale et documentaire ne sont PAS évaluées ici.  Pour évaluer la qualité ALTO, voir AltoView (S15)."
+      ]
+    },
+    {
+      "candidate_artifact_id": "doc02:precomputed_tess:raw_text",
+      "document_id": "doc02",
+      "failed_metrics": {},
+      "ground_truth_artifact_id": "doc02:gt:raw_text",
+      "ignored_dimensions": [
+        "block_structure",
+        "confidence",
+        "formatting",
+        "geometry",
+        "ids",
+        "reading_order"
+      ],
+      "metric_values": {
+        "cer": 0.05,
+        "mer": 0.25,
+        "wer": 0.25,
+        "wil": 0.4375
+      },
+      "pipeline_name": "ocr_then_correct",
+      "projection_report": null,
+      "view_name": "text_final",
+      "warnings": [
+        "Cette vue compare les sorties textuelles finales après projection éventuelle.  Les pipelines qui produisent ALTO/PAGE/markdown sont projetés vers du texte plat — leurs structures spatiale et documentaire ne sont PAS évaluées ici.  Pour évaluer la qualité ALTO, voir AltoView (S15)."
+      ]
+    },
+    {
+      "candidate_artifact_id": "doc02:precomputed_corr:raw_text",
+      "document_id": "doc02",
+      "failed_metrics": {},
+      "ground_truth_artifact_id": "doc02:gt:raw_text",
+      "ignored_dimensions": [
+        "block_structure",
+        "confidence",
+        "formatting",
+        "geometry",
+        "ids",
+        "reading_order"
+      ],
+      "metric_values": {
+        "cer": 0.05,
+        "mer": 0.25,
+        "wer": 0.25,
+        "wil": 0.4375
+      },
+      "pipeline_name": "ocr_then_correct",
+      "projection_report": null,
+      "view_name": "text_final",
+      "warnings": [
+        "Cette vue compare les sorties textuelles finales après projection éventuelle.  Les pipelines qui produisent ALTO/PAGE/markdown sont projetés vers du texte plat — leurs structures spatiale et documentaire ne sont PAS évaluées ici.  Pour évaluer la qualité ALTO, voir AltoView (S15)."
+      ]
+    },
+    {
+      "candidate_artifact_id": "doc03:precomputed_tess:raw_text",
+      "document_id": "doc03",
+      "failed_metrics": {},
+      "ground_truth_artifact_id": "doc03:gt:raw_text",
+      "ignored_dimensions": [
+        "block_structure",
+        "confidence",
+        "formatting",
+        "geometry",
+        "ids",
+        "reading_order"
+      ],
+      "metric_values": {
+        "cer": 0.05,
+        "mer": 0.25,
+        "wer": 0.25,
+        "wil": 0.4375
+      },
+      "pipeline_name": "ocr_then_correct",
+      "projection_report": null,
+      "view_name": "text_final",
+      "warnings": [
+        "Cette vue compare les sorties textuelles finales après projection éventuelle.  Les pipelines qui produisent ALTO/PAGE/markdown sont projetés vers du texte plat — leurs structures spatiale et documentaire ne sont PAS évaluées ici.  Pour évaluer la qualité ALTO, voir AltoView (S15)."
+      ]
+    },
+    {
+      "candidate_artifact_id": "doc03:precomputed_corr:raw_text",
+      "document_id": "doc03",
+      "failed_metrics": {},
+      "ground_truth_artifact_id": "doc03:gt:raw_text",
+      "ignored_dimensions": [
+        "block_structure",
+        "confidence",
+        "formatting",
+        "geometry",
+        "ids",
+        "reading_order"
+      ],
+      "metric_values": {
+        "cer": 0.05,
+        "mer": 0.25,
+        "wer": 0.25,
+        "wil": 0.4375
+      },
+      "pipeline_name": "ocr_then_correct",
+      "projection_report": null,
+      "view_name": "text_final",
+      "warnings": [
+        "Cette vue compare les sorties textuelles finales après projection éventuelle.  Les pipelines qui produisent ALTO/PAGE/markdown sont projetés vers du texte plat — leurs structures spatiale et documentaire ne sont PAS évaluées ici.  Pour évaluer la qualité ALTO, voir AltoView (S15)."
+      ]
+    }
+  ]
+}

tests/golden/fixtures/run_orchestrator/multi_pipeline.json

@@ -0,0 +1,592 @@
+{
+  "artifacts_index": [
+    {
+      "content_hash": null,
+      "document_id": "doc01",
+      "id": "doc01:image",
+      "pipeline_name": "pero_only",
+      "produced_by_step": null,
+      "provenance": null,
+      "type": "image",
+      "uri": "<PATH>"
+    },
+    {
+      "content_hash": null,
+      "document_id": "doc01",
+      "id": "doc01:precomputed_pero:raw_text",
+      "pipeline_name": "pero_only",
+      "produced_by_step": "ocr",
+      "provenance": null,
+      "type": "raw_text",
+      "uri": "<PATH>"
+    },
+    {
+      "content_hash": null,
+      "document_id": "doc01",
+      "id": "doc01:image",
+      "pipeline_name": "tess_only",
+      "produced_by_step": null,
+      "provenance": null,
+      "type": "image",
+      "uri": "<PATH>"
+    },
+    {
+      "content_hash": null,
+      "document_id": "doc01",
+      "id": "doc01:precomputed_tess:raw_text",
+      "pipeline_name": "tess_only",
+      "produced_by_step": "ocr",
+      "provenance": null,
+      "type": "raw_text",
+      "uri": "<PATH>"
+    },
+    {
+      "content_hash": null,
+      "document_id": "doc02",
+      "id": "doc02:image",
+      "pipeline_name": "pero_only",
+      "produced_by_step": null,
+      "provenance": null,
+      "type": "image",
+      "uri": "<PATH>"
+    },
+    {
+      "content_hash": null,
+      "document_id": "doc02",
+      "id": "doc02:precomputed_pero:raw_text",
+      "pipeline_name": "pero_only",
+      "produced_by_step": "ocr",
+      "provenance": null,
+      "type": "raw_text",
+      "uri": "<PATH>"
+    },
+    {
+      "content_hash": null,
+      "document_id": "doc02",
+      "id": "doc02:image",
+      "pipeline_name": "tess_only",
+      "produced_by_step": null,
+      "provenance": null,
+      "type": "image",
+      "uri": "<PATH>"
+    },
+    {
+      "content_hash": null,
+      "document_id": "doc02",
+      "id": "doc02:precomputed_tess:raw_text",
+      "pipeline_name": "tess_only",
+      "produced_by_step": "ocr",
+      "provenance": null,
+      "type": "raw_text",
+      "uri": "<PATH>"
+    },
+    {
+      "content_hash": null,
+      "document_id": "doc03",
+      "id": "doc03:image",
+      "pipeline_name": "pero_only",
+      "produced_by_step": null,
+      "provenance": null,
+      "type": "image",
+      "uri": "<PATH>"
+    },
+    {
+      "content_hash": null,
+      "document_id": "doc03",
+      "id": "doc03:precomputed_pero:raw_text",
+      "pipeline_name": "pero_only",
+      "produced_by_step": "ocr",
+      "provenance": null,
+      "type": "raw_text",
+      "uri": "<PATH>"
+    },
+    {
+      "content_hash": null,
+      "document_id": "doc03",
+      "id": "doc03:image",
+      "pipeline_name": "tess_only",
+      "produced_by_step": null,
+      "provenance": null,
+      "type": "image",
+      "uri": "<PATH>"
+    },
+    {
+      "content_hash": null,
+      "document_id": "doc03",
+      "id": "doc03:precomputed_tess:raw_text",
+      "pipeline_name": "tess_only",
+      "produced_by_step": "ocr",
+      "provenance": null,
+      "type": "raw_text",
+      "uri": "<PATH>"
+    }
+  ],
+  "manifest": {
+    "adapter_kwargs": {
+      "ocr": {
+        "source_label": "tess"
+      },
+      "pero_only__ocr": {
+        "source_label": "pero"
+      }
+    },
+    "code_version": "charac-1.0",
+    "corpus_name": "charac",
+    "dependencies_lock": {
+      "CacheControl": "0.14.4",
+      "Jinja2": "3.1.6",
+      "MarkupSafe": "3.0.3",
+      "PyGObject": "3.48.2",
+      "PyJWT": "2.7.0",
+      "PyYAML": "6.0.1",
+      "Pygments": "2.20.0",
+      "RapidFuzz": "3.14.5",
+      "annotated-doc": "0.0.4",
+      "annotated-types": "0.7.0",
+      "anyio": "4.13.0",
+      "argcomplete": "3.1.4",
+      "ast_serialize": "0.5.0",
+      "bandit": "1.9.4",
+      "blinker": "1.7.0",
+      "boolean.py": "5.0",
+      "certifi": "2026.2.25",
+      "charset-normalizer": "3.4.6",
+      "click": "8.4.0",
+      "colorama": "0.4.6",
+      "conan": "2.27.0",
+      "coverage": "7.14.0",
+      "cryptography": "41.0.7",
+      "cyclonedx-python-lib": "11.7.0",
+      "dbus-python": "1.3.2",
+      "defusedxml": "0.7.1",
+      "distro": "1.9.0",
+      "fastapi": "0.136.1",
+      "fasteners": "0.20",
+      "filelock": "3.29.0",
+      "h11": "0.16.0",
+      "httpcore": "1.0.9",
+      "httplib2": "0.20.4",
+      "httptools": "0.7.1",
+      "httpx": "0.28.1",
+      "idna": "3.11",
+      "iniconfig": "2.3.0",
+      "jiwer": "4.0.0",
+      "launchpadlib": "1.11.0",
+      "lazr.restfulclient": "0.14.6",
+      "lazr.uri": "1.0.6",
+      "librt": "0.11.0",
+      "license-expression": "30.4.4",
+      "markdown-it-py": "4.2.0",
+      "mdurl": "0.1.2",
+      "msgpack": "1.1.2",
+      "mypy": "2.1.0",
+      "mypy_extensions": "1.1.0",
+      "numpy": "2.4.6",
+      "oauthlib": "3.2.2",
+      "packageurl-python": "0.17.6",
+      "packaging": "24.0",
+      "patch-ng": "1.18.1",
+      "pathspec": "1.1.1",
+      "picarones": "1.1.0.dev311",
+      "pillow": "12.2.0",
+      "pip": "24.0",
+      "pip-api": "0.0.34",
+      "pip-requirements-parser": "32.0.1",
+      "pip_audit": "2.10.0",
+      "platformdirs": "4.9.6",
+      "pluggy": "1.6.0",
+      "py-serializable": "2.1.0",
+      "pydantic": "2.13.4",
+      "pydantic_core": "2.46.4",
+      "pyparsing": "3.1.1",
+      "pytesseract": "0.3.13",
+      "pytest": "9.0.3",
+      "pytest-cov": "7.1.0",
+      "pytest-timeout": "2.4.0",
+      "python-apt": "2.7.7+ubuntu5.2",
+      "python-dateutil": "2.9.0.post0",
+      "python-dotenv": "1.2.2",
+      "python-multipart": "0.0.29",
+      "requests": "2.33.1",
+      "rich": "15.0.0",
+      "setuptools": "68.1.2",
+      "six": "1.16.0",
+      "sortedcontainers": "2.4.0",
+      "starlette": "1.0.0",
+      "stevedore": "5.8.0",
+      "toml": "0.10.2",
+      "tomli": "2.4.1",
+      "tomli_w": "1.2.0",
+      "tqdm": "4.67.3",
+      "typing-inspection": "0.4.2",
+      "typing_extensions": "4.15.0",
+      "urllib3": "2.6.3",
+      "uvicorn": "0.47.0",
+      "uvloop": "0.22.1",
+      "wadllib": "1.3.6",
+      "watchfiles": "1.2.0",
+      "websockets": "16.0",
+      "wheel": "0.42.0",
+      "xmltodict": "0.13.0",
+      "yq": "3.1.0"
+    },
+    "metadata": {
+      "orchestrator": "picarones.app.services.run_orchestrator"
+    },
+    "n_documents": 3,
+    "pipeline_names": [
+      "pero_only",
+      "tess_only"
+    ],
+    "pipeline_specs": [
+      {
+        "description": "",
+        "initial_inputs": [
+          "image"
+        ],
+        "name": "tess_only",
+        "steps": [
+          {
+            "adapter_name": "ocr",
+            "id": "ocr",
+            "input_types": [
+              "image"
+            ],
+            "inputs_from": {},
+            "kind": "step",
+            "output_types": [
+              "raw_text"
+            ],
+            "params": {}
+          }
+        ]
+      },
+      {
+        "description": "",
+        "initial_inputs": [
+          "image"
+        ],
+        "name": "pero_only",
+        "steps": [
+          {
+            "adapter_name": "pero_only__ocr",
+            "id": "ocr",
+            "input_types": [
+              "image"
+            ],
+            "inputs_from": {},
+            "kind": "step",
+            "output_types": [
+              "raw_text"
+            ],
+            "params": {}
+          }
+        ]
+      }
+    ],
+    "system_binaries_lock": {},
+    "view_specs": [
+      {
+        "candidate_types": [
+          "alto_xml",
+          "canonical_document",
+          "corrected_text",
+          "page_xml",
+          "raw_text"
+        ],
+        "char_exclude": null,
+        "description": "Compare les sorties textuelles finales après projection éventuelle (ALTO/PAGE/markdown → texte plat).",
+        "ignored_dimensions": [
+          "block_structure",
+          "confidence",
+          "formatting",
+          "geometry",
+          "ids",
+          "reading_order"
+        ],
+        "metric_names": [
+          "cer",
+          "mer",
+          "wer",
+          "wil"
+        ],
+        "name": "text_final",
+        "normalization_profile": null,
+        "projection": null,
+        "projections_by_source_type": {
+          "alto_xml": {
+            "params": {},
+            "projector_name": "alto_to_text",
+            "source_type": "alto_xml",
+            "target_type": "raw_text"
+          },
+          "canonical_document": {
+            "params": {},
+            "projector_name": "canonical_to_text",
+            "source_type": "canonical_document",
+            "target_type": "raw_text"
+          },
+          "page_xml": {
+            "params": {},
+            "projector_name": "page_to_text",
+            "source_type": "page_xml",
+            "target_type": "raw_text"
+          }
+        },
+        "warnings": [
+          "Cette vue compare les sorties textuelles finales après projection éventuelle.  Les pipelines qui produisent ALTO/PAGE/markdown sont projetés vers du texte plat — leurs structures spatiale et documentaire ne sont PAS évaluées ici.  Pour évaluer la qualité ALTO, voir AltoView (S15)."
+        ]
+      }
+    ]
+  },
+  "pipeline_results": [
+    {
+      "document_id": "doc01",
+      "pipeline_name": "pero_only",
+      "step_results": [
+        {
+          "error": null,
+          "produced_artifacts": {
+            "raw_text": "doc01:precomputed_pero:raw_text"
+          },
+          "step_id": "ocr",
+          "succeeded": true
+        }
+      ],
+      "succeeded": true
+    },
+    {
+      "document_id": "doc01",
+      "pipeline_name": "tess_only",
+      "step_results": [
+        {
+          "error": null,
+          "produced_artifacts": {
+            "raw_text": "doc01:precomputed_tess:raw_text"
+          },
+          "step_id": "ocr",
+          "succeeded": true
+        }
+      ],
+      "succeeded": true
+    },
+    {
+      "document_id": "doc02",
+      "pipeline_name": "pero_only",
+      "step_results": [
+        {
+          "error": null,
+          "produced_artifacts": {
+            "raw_text": "doc02:precomputed_pero:raw_text"
+          },
+          "step_id": "ocr",
+          "succeeded": true
+        }
+      ],
+      "succeeded": true
+    },
+    {
+      "document_id": "doc02",
+      "pipeline_name": "tess_only",
+      "step_results": [
+        {
+          "error": null,
+          "produced_artifacts": {
+            "raw_text": "doc02:precomputed_tess:raw_text"
+          },
+          "step_id": "ocr",
+          "succeeded": true
+        }
+      ],
+      "succeeded": true
+    },
+    {
+      "document_id": "doc03",
+      "pipeline_name": "pero_only",
+      "step_results": [
+        {
+          "error": null,
+          "produced_artifacts": {
+            "raw_text": "doc03:precomputed_pero:raw_text"
+          },
+          "step_id": "ocr",
+          "succeeded": true
+        }
+      ],
+      "succeeded": true
+    },
+    {
+      "document_id": "doc03",
+      "pipeline_name": "tess_only",
+      "step_results": [
+        {
+          "error": null,
+          "produced_artifacts": {
+            "raw_text": "doc03:precomputed_tess:raw_text"
+          },
+          "step_id": "ocr",
+          "succeeded": true
+        }
+      ],
+      "succeeded": true
+    }
+  ],
+  "view_results": [
+    {
+      "candidate_artifact_id": "doc01:precomputed_tess:raw_text",
+      "document_id": "doc01",
+      "failed_metrics": {},
+      "ground_truth_artifact_id": "doc01:gt:raw_text",
+      "ignored_dimensions": [
+        "block_structure",
+        "confidence",
+        "formatting",
+        "geometry",
+        "ids",
+        "reading_order"
+      ],
+      "metric_values": {
+        "cer": 0.05,
+        "mer": 0.25,
+        "wer": 0.25,
+        "wil": 0.4375
+      },
+      "pipeline_name": "tess_only",
+      "projection_report": null,
+      "view_name": "text_final",
+      "warnings": [
+        "Cette vue compare les sorties textuelles finales après projection éventuelle.  Les pipelines qui produisent ALTO/PAGE/markdown sont projetés vers du texte plat — leurs structures spatiale et documentaire ne sont PAS évaluées ici.  Pour évaluer la qualité ALTO, voir AltoView (S15)."
+      ]
+    },
+    {
+      "candidate_artifact_id": "doc01:precomputed_pero:raw_text",
+      "document_id": "doc01",
+      "failed_metrics": {},
+      "ground_truth_artifact_id": "doc01:gt:raw_text",
+      "ignored_dimensions": [
+        "block_structure",
+        "confidence",
+        "formatting",
+        "geometry",
+        "ids",
+        "reading_order"
+      ],
+      "metric_values": {
+        "cer": 0.05,
+        "mer": 0.25,
+        "wer": 0.25,
+        "wil": 0.4375
+      },
+      "pipeline_name": "pero_only",
+      "projection_report": null,
+      "view_name": "text_final",
+      "warnings": [
+        "Cette vue compare les sorties textuelles finales après projection éventuelle.  Les pipelines qui produisent ALTO/PAGE/markdown sont projetés vers du texte plat — leurs structures spatiale et documentaire ne sont PAS évaluées ici.  Pour évaluer la qualité ALTO, voir AltoView (S15)."
+      ]
+    },
+    {
+      "candidate_artifact_id": "doc02:precomputed_tess:raw_text",
+      "document_id": "doc02",
+      "failed_metrics": {},
+      "ground_truth_artifact_id": "doc02:gt:raw_text",
+      "ignored_dimensions": [
+        "block_structure",
+        "confidence",
+        "formatting",
+        "geometry",
+        "ids",
+        "reading_order"
+      ],
+      "metric_values": {
+        "cer": 0.05,
+        "mer": 0.25,
+        "wer": 0.25,
+        "wil": 0.4375
+      },
+      "pipeline_name": "tess_only",
+      "projection_report": null,
+      "view_name": "text_final",
+      "warnings": [
+        "Cette vue compare les sorties textuelles finales après projection éventuelle.  Les pipelines qui produisent ALTO/PAGE/markdown sont projetés vers du texte plat — leurs structures spatiale et documentaire ne sont PAS évaluées ici.  Pour évaluer la qualité ALTO, voir AltoView (S15)."
+      ]
+    },
+    {
+      "candidate_artifact_id": "doc02:precomputed_pero:raw_text",
+      "document_id": "doc02",
+      "failed_metrics": {},
+      "ground_truth_artifact_id": "doc02:gt:raw_text",
+      "ignored_dimensions": [
+        "block_structure",
+        "confidence",
+        "formatting",
+        "geometry",
+        "ids",
+        "reading_order"
+      ],
+      "metric_values": {
+        "cer": 0.05,
+        "mer": 0.25,
+        "wer": 0.25,
+        "wil": 0.4375
+      },
+      "pipeline_name": "pero_only",
+      "projection_report": null,
+      "view_name": "text_final",
+      "warnings": [
+        "Cette vue compare les sorties textuelles finales après projection éventuelle.  Les pipelines qui produisent ALTO/PAGE/markdown sont projetés vers du texte plat — leurs structures spatiale et documentaire ne sont PAS évaluées ici.  Pour évaluer la qualité ALTO, voir AltoView (S15)."
+      ]
+    },
+    {
+      "candidate_artifact_id": "doc03:precomputed_tess:raw_text",
+      "document_id": "doc03",
+      "failed_metrics": {},
+      "ground_truth_artifact_id": "doc03:gt:raw_text",
+      "ignored_dimensions": [
+        "block_structure",
+        "confidence",
+        "formatting",
+        "geometry",
+        "ids",
+        "reading_order"
+      ],
+      "metric_values": {
+        "cer": 0.05,
+        "mer": 0.25,
+        "wer": 0.25,
+        "wil": 0.4375
+      },
+      "pipeline_name": "tess_only",
+      "projection_report": null,
+      "view_name": "text_final",
+      "warnings": [
+        "Cette vue compare les sorties textuelles finales après projection éventuelle.  Les pipelines qui produisent ALTO/PAGE/markdown sont projetés vers du texte plat — leurs structures spatiale et documentaire ne sont PAS évaluées ici.  Pour évaluer la qualité ALTO, voir AltoView (S15)."
+      ]
+    },
+    {
+      "candidate_artifact_id": "doc03:precomputed_pero:raw_text",
+      "document_id": "doc03",
+      "failed_metrics": {},
+      "ground_truth_artifact_id": "doc03:gt:raw_text",
+      "ignored_dimensions": [
+        "block_structure",
+        "confidence",
+        "formatting",
+        "geometry",
+        "ids",
+        "reading_order"
+      ],
+      "metric_values": {
+        "cer": 0.05,
+        "mer": 0.25,
+        "wer": 0.25,
+        "wil": 0.4375
+      },
+      "pipeline_name": "pero_only",
+      "projection_report": null,
+      "view_name": "text_final",
+      "warnings": [
+        "Cette vue compare les sorties textuelles finales après projection éventuelle.  Les pipelines qui produisent ALTO/PAGE/markdown sont projetés vers du texte plat — leurs structures spatiale et documentaire ne sont PAS évaluées ici.  Pour évaluer la qualité ALTO, voir AltoView (S15)."
+      ]
+    }
+  ]
+}

tests/golden/fixtures/run_orchestrator/single_linear.json

@@ -0,0 +1,383 @@
+{
+  "artifacts_index": [
+    {
+      "content_hash": null,
+      "document_id": "doc01",
+      "id": "doc01:image",
+      "pipeline_name": "tess_only",
+      "produced_by_step": null,
+      "provenance": null,
+      "type": "image",
+      "uri": "<PATH>"
+    },
+    {
+      "content_hash": null,
+      "document_id": "doc01",
+      "id": "doc01:precomputed_tess:raw_text",
+      "pipeline_name": "tess_only",
+      "produced_by_step": "ocr",
+      "provenance": null,
+      "type": "raw_text",
+      "uri": "<PATH>"
+    },
+    {
+      "content_hash": null,
+      "document_id": "doc02",
+      "id": "doc02:image",
+      "pipeline_name": "tess_only",
+      "produced_by_step": null,
+      "provenance": null,
+      "type": "image",
+      "uri": "<PATH>"
+    },
+    {
+      "content_hash": null,
+      "document_id": "doc02",
+      "id": "doc02:precomputed_tess:raw_text",
+      "pipeline_name": "tess_only",
+      "produced_by_step": "ocr",
+      "provenance": null,
+      "type": "raw_text",
+      "uri": "<PATH>"
+    },
+    {
+      "content_hash": null,
+      "document_id": "doc03",
+      "id": "doc03:image",
+      "pipeline_name": "tess_only",
+      "produced_by_step": null,
+      "provenance": null,
+      "type": "image",
+      "uri": "<PATH>"
+    },
+    {
+      "content_hash": null,
+      "document_id": "doc03",
+      "id": "doc03:precomputed_tess:raw_text",
+      "pipeline_name": "tess_only",
+      "produced_by_step": "ocr",
+      "provenance": null,
+      "type": "raw_text",
+      "uri": "<PATH>"
+    }
+  ],
+  "manifest": {
+    "adapter_kwargs": {
+      "ocr": {
+        "source_label": "tess"
+      }
+    },
+    "code_version": "charac-1.0",
+    "corpus_name": "charac",
+    "dependencies_lock": {
+      "CacheControl": "0.14.4",
+      "Jinja2": "3.1.6",
+      "MarkupSafe": "3.0.3",
+      "PyGObject": "3.48.2",
+      "PyJWT": "2.7.0",
+      "PyYAML": "6.0.1",
+      "Pygments": "2.20.0",
+      "RapidFuzz": "3.14.5",
+      "annotated-doc": "0.0.4",
+      "annotated-types": "0.7.0",
+      "anyio": "4.13.0",
+      "argcomplete": "3.1.4",
+      "ast_serialize": "0.5.0",
+      "bandit": "1.9.4",
+      "blinker": "1.7.0",
+      "boolean.py": "5.0",
+      "certifi": "2026.2.25",
+      "charset-normalizer": "3.4.6",
+      "click": "8.4.0",
+      "colorama": "0.4.6",
+      "conan": "2.27.0",
+      "coverage": "7.14.0",
+      "cryptography": "41.0.7",
+      "cyclonedx-python-lib": "11.7.0",
+      "dbus-python": "1.3.2",
+      "defusedxml": "0.7.1",
+      "distro": "1.9.0",
+      "fastapi": "0.136.1",
+      "fasteners": "0.20",
+      "filelock": "3.29.0",
+      "h11": "0.16.0",
+      "httpcore": "1.0.9",
+      "httplib2": "0.20.4",
+      "httptools": "0.7.1",
+      "httpx": "0.28.1",
+      "idna": "3.11",
+      "iniconfig": "2.3.0",
+      "jiwer": "4.0.0",
+      "launchpadlib": "1.11.0",
+      "lazr.restfulclient": "0.14.6",
+      "lazr.uri": "1.0.6",
+      "librt": "0.11.0",
+      "license-expression": "30.4.4",
+      "markdown-it-py": "4.2.0",
+      "mdurl": "0.1.2",
+      "msgpack": "1.1.2",
+      "mypy": "2.1.0",
+      "mypy_extensions": "1.1.0",
+      "numpy": "2.4.6",
+      "oauthlib": "3.2.2",
+      "packageurl-python": "0.17.6",
+      "packaging": "24.0",
+      "patch-ng": "1.18.1",
+      "pathspec": "1.1.1",
+      "picarones": "1.1.0.dev311",
+      "pillow": "12.2.0",
+      "pip": "24.0",
+      "pip-api": "0.0.34",
+      "pip-requirements-parser": "32.0.1",
+      "pip_audit": "2.10.0",
+      "platformdirs": "4.9.6",
+      "pluggy": "1.6.0",
+      "py-serializable": "2.1.0",
+      "pydantic": "2.13.4",
+      "pydantic_core": "2.46.4",
+      "pyparsing": "3.1.1",
+      "pytesseract": "0.3.13",
+      "pytest": "9.0.3",
+      "pytest-cov": "7.1.0",
+      "pytest-timeout": "2.4.0",
+      "python-apt": "2.7.7+ubuntu5.2",
+      "python-dateutil": "2.9.0.post0",
+      "python-dotenv": "1.2.2",
+      "python-multipart": "0.0.29",
+      "requests": "2.33.1",
+      "rich": "15.0.0",
+      "setuptools": "68.1.2",
+      "six": "1.16.0",
+      "sortedcontainers": "2.4.0",
+      "starlette": "1.0.0",
+      "stevedore": "5.8.0",
+      "toml": "0.10.2",
+      "tomli": "2.4.1",
+      "tomli_w": "1.2.0",
+      "tqdm": "4.67.3",
+      "typing-inspection": "0.4.2",
+      "typing_extensions": "4.15.0",
+      "urllib3": "2.6.3",
+      "uvicorn": "0.47.0",
+      "uvloop": "0.22.1",
+      "wadllib": "1.3.6",
+      "watchfiles": "1.2.0",
+      "websockets": "16.0",
+      "wheel": "0.42.0",
+      "xmltodict": "0.13.0",
+      "yq": "3.1.0"
+    },
+    "metadata": {
+      "orchestrator": "picarones.app.services.run_orchestrator"
+    },
+    "n_documents": 3,
+    "pipeline_names": [
+      "tess_only"
+    ],
+    "pipeline_specs": [
+      {
+        "description": "",
+        "initial_inputs": [
+          "image"
+        ],
+        "name": "tess_only",
+        "steps": [
+          {
+            "adapter_name": "ocr",
+            "id": "ocr",
+            "input_types": [
+              "image"
+            ],
+            "inputs_from": {},
+            "kind": "step",
+            "output_types": [
+              "raw_text"
+            ],
+            "params": {}
+          }
+        ]
+      }
+    ],
+    "system_binaries_lock": {},
+    "view_specs": [
+      {
+        "candidate_types": [
+          "alto_xml",
+          "canonical_document",
+          "corrected_text",
+          "page_xml",
+          "raw_text"
+        ],
+        "char_exclude": null,
+        "description": "Compare les sorties textuelles finales après projection éventuelle (ALTO/PAGE/markdown → texte plat).",
+        "ignored_dimensions": [
+          "block_structure",
+          "confidence",
+          "formatting",
+          "geometry",
+          "ids",
+          "reading_order"
+        ],
+        "metric_names": [
+          "cer",
+          "mer",
+          "wer",
+          "wil"
+        ],
+        "name": "text_final",
+        "normalization_profile": null,
+        "projection": null,
+        "projections_by_source_type": {
+          "alto_xml": {
+            "params": {},
+            "projector_name": "alto_to_text",
+            "source_type": "alto_xml",
+            "target_type": "raw_text"
+          },
+          "canonical_document": {
+            "params": {},
+            "projector_name": "canonical_to_text",
+            "source_type": "canonical_document",
+            "target_type": "raw_text"
+          },
+          "page_xml": {
+            "params": {},
+            "projector_name": "page_to_text",
+            "source_type": "page_xml",
+            "target_type": "raw_text"
+          }
+        },
+        "warnings": [
+          "Cette vue compare les sorties textuelles finales après projection éventuelle.  Les pipelines qui produisent ALTO/PAGE/markdown sont projetés vers du texte plat — leurs structures spatiale et documentaire ne sont PAS évaluées ici.  Pour évaluer la qualité ALTO, voir AltoView (S15)."
+        ]
+      }
+    ]
+  },
+  "pipeline_results": [
+    {
+      "document_id": "doc01",
+      "pipeline_name": "tess_only",
+      "step_results": [
+        {
+          "error": null,
+          "produced_artifacts": {
+            "raw_text": "doc01:precomputed_tess:raw_text"
+          },
+          "step_id": "ocr",
+          "succeeded": true
+        }
+      ],
+      "succeeded": true
+    },
+    {
+      "document_id": "doc02",
+      "pipeline_name": "tess_only",
+      "step_results": [
+        {
+          "error": null,
+          "produced_artifacts": {
+            "raw_text": "doc02:precomputed_tess:raw_text"
+          },
+          "step_id": "ocr",
+          "succeeded": true
+        }
+      ],
+      "succeeded": true
+    },
+    {
+      "document_id": "doc03",
+      "pipeline_name": "tess_only",
+      "step_results": [
+        {
+          "error": null,
+          "produced_artifacts": {
+            "raw_text": "doc03:precomputed_tess:raw_text"
+          },
+          "step_id": "ocr",
+          "succeeded": true
+        }
+      ],
+      "succeeded": true
+    }
+  ],
+  "view_results": [
+    {
+      "candidate_artifact_id": "doc01:precomputed_tess:raw_text",
+      "document_id": "doc01",
+      "failed_metrics": {},
+      "ground_truth_artifact_id": "doc01:gt:raw_text",
+      "ignored_dimensions": [
+        "block_structure",
+        "confidence",
+        "formatting",
+        "geometry",
+        "ids",
+        "reading_order"
+      ],
+      "metric_values": {
+        "cer": 0.05,
+        "mer": 0.25,
+        "wer": 0.25,
+        "wil": 0.4375
+      },
+      "pipeline_name": "tess_only",
+      "projection_report": null,
+      "view_name": "text_final",
+      "warnings": [
+        "Cette vue compare les sorties textuelles finales après projection éventuelle.  Les pipelines qui produisent ALTO/PAGE/markdown sont projetés vers du texte plat — leurs structures spatiale et documentaire ne sont PAS évaluées ici.  Pour évaluer la qualité ALTO, voir AltoView (S15)."
+      ]
+    },
+    {
+      "candidate_artifact_id": "doc02:precomputed_tess:raw_text",
+      "document_id": "doc02",
+      "failed_metrics": {},
+      "ground_truth_artifact_id": "doc02:gt:raw_text",
+      "ignored_dimensions": [
+        "block_structure",
+        "confidence",
+        "formatting",
+        "geometry",
+        "ids",
+        "reading_order"
+      ],
+      "metric_values": {
+        "cer": 0.05,
+        "mer": 0.25,
+        "wer": 0.25,
+        "wil": 0.4375
+      },
+      "pipeline_name": "tess_only",
+      "projection_report": null,
+      "view_name": "text_final",
+      "warnings": [
+        "Cette vue compare les sorties textuelles finales après projection éventuelle.  Les pipelines qui produisent ALTO/PAGE/markdown sont projetés vers du texte plat — leurs structures spatiale et documentaire ne sont PAS évaluées ici.  Pour évaluer la qualité ALTO, voir AltoView (S15)."
+      ]
+    },
+    {
+      "candidate_artifact_id": "doc03:precomputed_tess:raw_text",
+      "document_id": "doc03",
+      "failed_metrics": {},
+      "ground_truth_artifact_id": "doc03:gt:raw_text",
+      "ignored_dimensions": [
+        "block_structure",
+        "confidence",
+        "formatting",
+        "geometry",
+        "ids",
+        "reading_order"
+      ],
+      "metric_values": {
+        "cer": 0.05,
+        "mer": 0.25,
+        "wer": 0.25,
+        "wil": 0.4375
+      },
+      "pipeline_name": "tess_only",
+      "projection_report": null,
+      "view_name": "text_final",
+      "warnings": [
+        "Cette vue compare les sorties textuelles finales après projection éventuelle.  Les pipelines qui produisent ALTO/PAGE/markdown sont projetés vers du texte plat — leurs structures spatiale et documentaire ne sont PAS évaluées ici.  Pour évaluer la qualité ALTO, voir AltoView (S15)."
+      ]
+    }
+  ]
+}

tests/golden/test_run_orchestrator_characterization.py

@@ -0,0 +1,607 @@
+"""Harnais de caractérisation du cœur stateful de ``RunOrchestrator``.
+
+Audit prod — PRÉCONDITION de la Phase B (décomposition stateful).
+Verrouille le comportement EXACT actuel là où Phase B va toucher,
+pour transformer « la CI pourrait ne pas attraper » en « la CI
+attrape ».  Cinq groupes, un par cas de risque identifié :
+
+1. :class:`TestGlobalDocIdxContract`     — compteur ``doc_idx`` global
+   au run (pas par pipeline) sur **multi-pipeline**.
+2. :class:`TestCancelPropagation`        — annulation pré-set ET
+   mi-run (identité d'objet event à travers les couches).
+3. :class:`TestCrashResumeConsistency`   — interruption réelle puis
+   resume : sortie finale identique à un run complet propre.
+4. :class:`TestGoldenMultiTopology`      — snapshot normalisé
+   déterministe des artefacts (manifest/pipeline/view) sur 4
+   topologies (linéaire, multi-pipeline, DAG branchant, OCR+corr).
+5. :class:`TestConcurrencyIsolation`     — deux ``execute()`` en
+   threads parallèles : isolation cancel/progress, pas de fuite
+   d'état entre instances.
+
+Déterminisme : ``PrecomputedTextAdapter`` (lit ``<stem>.<label>.txt``,
+aucun OCR/réseau).  Un garde explicite
+(:meth:`TestGoldenMultiTopology.test_snapshot_is_deterministic`)
+échoue si le snapshot n'est pas reproductible — un golden flaky
+serait pire que pas de golden.
+"""
+
+from __future__ import annotations
+
+import io
+import json
+import textwrap
+import threading
+import time
+import zipfile
+from pathlib import Path
+from typing import Any
+
+import pytest
+
+from picarones.app.schemas.run_spec import load_run_spec_from_yaml
+from picarones.app.services import RunOrchestrator
+
+_FIX_DIR = Path(__file__).parent / "fixtures" / "run_orchestrator"
+
+
+# ---------------------------------------------------------------------------
+# Briques déterministes
+# ---------------------------------------------------------------------------
+
+def _png_bytes() -> bytes:
+    return (
+        b"\x89PNG\r\n\x1a\n"
+        b"\x00\x00\x00\rIHDR"
+        b"\x00\x00\x00\x01\x00\x00\x00\x01\x08\x06\x00\x00\x00"
+        b"\x1f\x15\xc4\x89"
+    )
+
+
+def _make_corpus_zip(
+    tmp: Path, n_docs: int, *, sources: tuple[str, ...] = ("tess",),
+) -> Path:
+    """Corpus ZIP déterministe.
+
+    Pour chaque ``source_label`` de ``sources`` on écrit
+    ``<doc>.<label>.txt`` (lu par ``PrecomputedTextAdapter``).  La GT
+    diffère légèrement du texte pour produire un CER non trivial mais
+    fixe.
+    """
+    buf = io.BytesIO()
+    with zipfile.ZipFile(buf, mode="w") as zf:
+        for i in range(1, n_docs + 1):
+            doc = f"doc{i:02d}"
+            zf.writestr(f"{doc}.png", _png_bytes())
+            zf.writestr(f"{doc}.gt.txt", f"Texte de reference {i}")
+            for label in sources:
+                # 1 substitution stable → CER déterministe non nul.
+                zf.writestr(f"{doc}.{label}.txt", f"Texte de reference {i}!")
+    tmp.mkdir(parents=True, exist_ok=True)
+    p = tmp / "corpus.zip"
+    p.write_bytes(buf.getvalue())
+    return p
+
+
+def _spec(corpus_zip: Path, out: Path, body: str) -> Any:
+    yaml = textwrap.dedent(f"""
+        corpus_zip: {corpus_zip}
+        corpus_name: charac
+        output_dir: {out}
+        code_version: "charac-1.0"
+        views: [text_final]
+        {body}
+    """)
+    return load_run_spec_from_yaml(yaml)
+
+
+_TOPOLOGIES: dict[str, tuple[tuple[str, ...], str]] = {
+    "single_linear": (("tess",), """
+        pipelines:
+          - name: tess_only
+            initial_inputs: [image]
+            steps:
+              - id: ocr
+                adapter_class: picarones.adapters.ocr.precomputed.PrecomputedTextAdapter
+                adapter_kwargs: {source_label: tess}
+                input_types: [image]
+                output_types: [raw_text]
+    """),
+    "multi_pipeline": (("tess", "pero"), """
+        pipelines:
+          - name: tess_only
+            initial_inputs: [image]
+            steps:
+              - id: ocr
+                adapter_class: picarones.adapters.ocr.precomputed.PrecomputedTextAdapter
+                adapter_kwargs: {source_label: tess}
+                input_types: [image]
+                output_types: [raw_text]
+          - name: pero_only
+            initial_inputs: [image]
+            steps:
+              - id: ocr
+                adapter_class: picarones.adapters.ocr.precomputed.PrecomputedTextAdapter
+                adapter_kwargs: {source_label: pero}
+                input_types: [image]
+                output_types: [raw_text]
+    """),
+    "branching_dag": (("tess", "corr"), """
+        pipelines:
+          - name: ocr_then_correct
+            initial_inputs: [image]
+            preferred_text_output: corrector.corrected_text
+            steps:
+              - id: ocr
+                adapter_class: picarones.adapters.ocr.precomputed.PrecomputedTextAdapter
+                adapter_kwargs: {source_label: tess}
+                input_types: [image]
+                output_types: [raw_text]
+              - id: corrector
+                adapter_class: picarones.adapters.ocr.precomputed.PrecomputedTextAdapter
+                adapter_kwargs: {source_label: corr}
+                input_types: [image, raw_text]
+                output_types: [corrected_text]
+                inputs_from:
+                  raw_text: ocr
+    """),
+}
+
+
+# ---------------------------------------------------------------------------
+# Normalisation snapshot (déterministe : retire timestamps/durées/paths)
+# ---------------------------------------------------------------------------
+
+_VOLATILE_KEYS = {
+    "started_at", "completed_at", "duration_seconds", "run_date",
+    "created_at", "elapsed_seconds", "wall_clock_seconds",
+    # ``run_id`` est horodaté (``charac_YYYYMMDDThhmmssZ``) — volatil,
+    # pas une caractéristique de comportement.
+    "run_id",
+}
+
+
+def _scrub(obj: Any) -> Any:
+    """Snapshot canonique : retire les clés volatiles, neutralise les
+    chemins absolus, et **trie les listes de scalaires**.
+
+    Le tri des listes de scalaires est essentiel : certaines sont
+    dérivées de ``set`` (ex. ``ignored_dimensions``, types projetés)
+    dont l'ordre de sérialisation varie d'un *process* à l'autre
+    (randomisation du hash de chaînes Python).  Le garde déterminisme
+    intra-process ne l'attrape pas — seul un golden inter-process le
+    révèle.  On NE trie PAS les listes de dicts : ce sont des
+    enregistrements dont l'ordre peut porter du sens (et qui sont
+    déjà triés au niveau record par :func:`_normalized_snapshot`)."""
+    if isinstance(obj, dict):
+        return {k: _scrub(v) for k, v in obj.items() if k not in _VOLATILE_KEYS}
+    if isinstance(obj, list):
+        scrubbed = [_scrub(x) for x in obj]
+        if all(isinstance(x, (str, int, float, bool, type(None))) for x in scrubbed):
+            return sorted(scrubbed, key=lambda x: (x is None, str(x)))
+        return scrubbed
+    if isinstance(obj, str):
+        # Chemins absolus → placeholder (tmp_path varie par run/CI).
+        if "/" in obj and ("/tmp" in obj or "pytest" in obj or obj.startswith("/")):
+            return "<PATH>"
+        return obj
+    return obj
+
+
+def _jsonl(path: Path) -> list[dict]:
+    if not path.exists():
+        return []
+    return [json.loads(ln) for ln in path.read_text(encoding="utf-8").splitlines() if ln.strip()]
+
+
+def _normalized_snapshot(results_dir: Path) -> str:
+    """Snapshot canonique, ordre-stable, sans champs volatils."""
+    manifest = json.loads((results_dir / "run_manifest.json").read_text(encoding="utf-8"))
+    pipe = sorted(
+        _jsonl(results_dir / "pipeline_results.jsonl"),
+        key=lambda r: (r.get("document_id", ""), r.get("pipeline_name", "")),
+    )
+    views = sorted(
+        _jsonl(results_dir / "view_results.jsonl"),
+        key=lambda r: (r.get("document_id", ""), r.get("view_name", "")),
+    )
+    # 4e artefact : index des artefacts (découplé de pipeline_results
+    # pour le streaming).  Phase B réordonne l'extraction d'artefacts
+    # → à verrouiller aussi.
+    arts = sorted(
+        _jsonl(results_dir / "artifacts_index.jsonl"),
+        key=lambda r: (
+            r.get("document_id", ""), r.get("pipeline_name", ""),
+            r.get("id", ""),
+        ),
+    )
+    snap = {
+        "manifest": _scrub(manifest),
+        "pipeline_results": _scrub(pipe),
+        "view_results": _scrub(views),
+        "artifacts_index": _scrub(arts),
+    }
+    return json.dumps(snap, ensure_ascii=False, sort_keys=True, indent=2)
+
+
+def _results_snapshot(results_dir: Path) -> str:
+    """Snapshot des RÉSULTATS calculés seuls (pipeline + views),
+    indépendant de l'écho de config du manifest (``partial_dir``,
+    ``output_dir``…).  C'est l'invariant que « resume == run propre »
+    doit préserver : mêmes sorties par document, peu importe que le
+    run ait été configuré avec un ``partial_dir`` ou non."""
+    pipe = sorted(
+        _jsonl(results_dir / "pipeline_results.jsonl"),
+        key=lambda r: (r.get("document_id", ""), r.get("pipeline_name", "")),
+    )
+    views = sorted(
+        _jsonl(results_dir / "view_results.jsonl"),
+        key=lambda r: (r.get("document_id", ""), r.get("view_name", "")),
+    )
+    return json.dumps(
+        {"pipeline_results": _scrub(pipe), "view_results": _scrub(views)},
+        ensure_ascii=False, sort_keys=True, indent=2,
+    )
+
+
+def _golden(name: str, actual: str) -> None:
+    """Pattern golden : 1er run crée le fixture et échoue (force le
+    commit) ; ensuite compare strictement."""
+    gp = _FIX_DIR / f"{name}.json"
+    if not gp.exists():
+        gp.parent.mkdir(parents=True, exist_ok=True)
+        gp.write_text(actual + "\n", encoding="utf-8")
+        pytest.fail(
+            f"Golden créé : {gp} — vérifier puis committer le fixture.",
+        )
+    expected = gp.read_text(encoding="utf-8").rstrip("\n")
+    assert actual == expected, (
+        f"Snapshot divergent vs {gp}.\nUne dérive ici = régression "
+        "comportementale du cœur stateful (Phase B). Si intentionnel : "
+        "supprimer le golden et relancer pour régénérer."
+    )
+
+
+def _run(tmp: Path, topo: str, **kw: Any) -> RunOrchestrator:
+    sources, body = _TOPOLOGIES[topo]
+    cz = _make_corpus_zip(tmp / "in", kw.pop("n_docs", 3), sources=sources)
+    out = tmp / "out"
+    spec = _spec(cz, out, body)
+    orch = RunOrchestrator(out)
+    res = orch.execute(spec, **kw)
+    return res
+
+
+# ---------------------------------------------------------------------------
+# 1. Contrat compteur doc_idx GLOBAL (Risque 1)
+# ---------------------------------------------------------------------------
+
+class TestGlobalDocIdxContract:
+    """Le compteur ``doc_idx`` est global au run, pas par pipeline.
+    2 pipelines × N docs ⇒ séquence contiguë ``0..2N-1`` (PAS
+    ``0..N-1`` deux fois).  Phase B casse ça en premier si elle crée
+    le contexte par-collaborateur."""
+
+    def test_multi_pipeline_counter_is_global_and_contiguous(
+        self, tmp_path: Path,
+    ) -> None:
+        calls: list[tuple[str, int, str]] = []
+        self_dir = tmp_path
+        sources, body = _TOPOLOGIES["multi_pipeline"]
+        cz = _make_corpus_zip(self_dir / "in", 3, sources=sources)
+        spec = _spec(cz, self_dir / "out", body)
+        RunOrchestrator(self_dir / "out").execute(
+            spec, progress_callback=lambda e, i, d: calls.append((e, i, d)),
+        )
+
+        # 2 pipelines × 3 docs = 6 notifications.
+        assert len(calls) == 6, calls
+        indices = sorted(c[1] for c in calls)
+        assert indices == [0, 1, 2, 3, 4, 5], (
+            f"compteur NON global/contigu : {indices} — régression "
+            "Phase B (contexte par-pipeline au lieu de global au run)"
+        )
+        engines = {c[0] for c in calls}
+        assert engines == {"tess_only", "pero_only"}, engines
+
+    def test_single_pipeline_counter_zero_based(
+        self, tmp_path: Path,
+    ) -> None:
+        calls: list[int] = []
+        _run(
+            tmp_path, "single_linear", n_docs=4,
+            progress_callback=lambda e, i, d: calls.append(i),
+        )
+        assert sorted(calls) == [0, 1, 2, 3]
+
+
+# ---------------------------------------------------------------------------
+# 2. Propagation cancel_event (Risque 2)
+# ---------------------------------------------------------------------------
+
+class TestCancelPropagation:
+    def test_preset_cancel_multipipeline_stops_fast_and_partial(
+        self, tmp_path: Path,
+    ) -> None:
+        ev = threading.Event()
+        ev.set()
+        t0 = time.monotonic()
+        res = _run(tmp_path, "multi_pipeline", n_docs=5, cancel_event=ev)
+        elapsed = time.monotonic() - t0
+        assert elapsed < 10.0, f"pas de short-circuit (elapsed={elapsed})"
+        # 2 pipelines × 5 = 10 docs possibles ; annulé ⇒ strictement
+        # moins traités (souvent 0).
+        assert res.run_result.manifest is not None
+
+    def test_cancel_mid_run_via_callback_propagates(
+        self, tmp_path: Path,
+    ) -> None:
+        """Identité d'objet event à travers les couches : l'event est
+        ``set()`` APRÈS le 1er doc, depuis le callback ; les docs
+        suivants doivent être sautés."""
+        ev = threading.Event()
+        seen: list[str] = []
+
+        def cb(engine: str, idx: int, doc_id: str) -> None:
+            seen.append(doc_id)
+            if len(seen) == 1:
+                ev.set()  # annule après le tout premier doc
+
+        _run(
+            tmp_path, "single_linear", n_docs=8,
+            progress_callback=cb, cancel_event=ev,
+        )
+        # Sans propagation correcte, les 8 docs passent.  Avec : on
+        # s'arrête bien avant la fin (tolérance large pour le
+        # parallélisme max_in_flight=2 du runner).
+        assert len(seen) < 8, (
+            f"cancel mi-run NON propagé : {len(seen)}/8 docs traités "
+            "— régression identité d'objet event (Phase B)"
+        )
+
+
+# ---------------------------------------------------------------------------
+# 3. Interruption réelle → resume cohérent (Risque 4)
+# ---------------------------------------------------------------------------
+
+class TestCrashResumeConsistency:
+    """Interruption RÉELLE (pas un partial pré-amorcé complet) :
+    on stoppe mi-run via cancel, puis on relance SANS cancel avec le
+    même ``partial_dir``.
+
+    ⚠️ DÉFAUT PRÉ-EXISTANT DÉCOUVERT PAR CE HARNAIS ⚠️
+    Le partial store persiste ``PipelineResult`` mais PAS
+    ``ViewResult``.  Au resume, les documents rechargés du partial
+    récupèrent leurs ``pipeline_results`` mais **pas** leurs
+    ``view_results`` (jamais recalculés).  Conséquence : après une
+    reprise, ``view_results.jsonl`` est incomplet → toute métrique
+    agrégée (CER…) dérivée des vues est silencieusement faussée pour
+    les documents repris.
+
+    Ces tests CARACTÉRISENT le comportement ACTUEL (warts inclus) —
+    rôle d'un harnais de caractérisation — pour que Phase B ne
+    l'aggrave pas ET qu'une correction future du défaut soit
+    consciente (le test échouera, forçant la revue).  Le défaut
+    lui-même est remonté à l'opérateur, pas corrigé furtivement ici
+    (resume/views = changement stateful risqué, hors périmètre
+    « construire le harnais »)."""
+
+    def _persisted_doc_ids(self, results_dir: Path) -> tuple[list[str], list[str]]:
+        pr = sorted({r["document_id"] for r in _jsonl(results_dir / "pipeline_results.jsonl")})
+        vr = sorted({r["document_id"] for r in _jsonl(results_dir / "view_results.jsonl")})
+        return pr, vr
+
+    def _interrupt_then_resume(
+        self, tmp_path: Path, n_docs: int, stop_after: int,
+        topo: str = "single_linear",
+    ) -> tuple[Path, Path]:
+        sources, body = _TOPOLOGIES[topo]
+        cz = _make_corpus_zip(tmp_path / "in", n_docs, sources=sources)
+        partial = tmp_path / "partial"
+        partial.mkdir()
+        body_pd = body + f"\n        partial_dir: {partial}"
+
+        ev = threading.Event()
+        n = {"c": 0}
+
+        def cb(e: str, i: int, d: str) -> None:
+            n["c"] += 1
+            if n["c"] == stop_after:
+                ev.set()
+
+        out1 = tmp_path / "out1"
+        RunOrchestrator(out1).execute(
+            _spec(cz, out1, body_pd), progress_callback=cb, cancel_event=ev,
+        )
+        out2 = tmp_path / "out2"
+        RunOrchestrator(out2).execute(_spec(cz, out2, body_pd))
+        return out1 / "results", out2 / "results"
+
+    def test_clean_run_pipeline_and_views_are_consistent(
+        self, tmp_path: Path,
+    ) -> None:
+        """Référence : un run PROPRE a pipeline_results == view_results
+        (tous les docs des deux côtés)."""
+        sources, body = _TOPOLOGIES["single_linear"]
+        cz = _make_corpus_zip(tmp_path / "in", 5, sources=sources)
+        out = tmp_path / "out"
+        RunOrchestrator(out).execute(_spec(cz, out, body))
+        pr, vr = self._persisted_doc_ids(out / "results")
+        assert pr == vr == ["doc01", "doc02", "doc03", "doc04", "doc05"]
+
+    def test_resume_pipeline_results_complete(self, tmp_path: Path) -> None:
+        """Le resume RECONSTRUIT bien tous les ``pipeline_results``
+        (partiel chargé + reste rejoué) — cette partie est correcte."""
+        _, resumed = self._interrupt_then_resume(tmp_path, 5, stop_after=2)
+        pr, _ = self._persisted_doc_ids(resumed)
+        assert pr == ["doc01", "doc02", "doc03", "doc04", "doc05"]
+
+    #: Relation ``view_results`` vs ``pipeline_results`` APRÈS resume,
+    #: par topologie — comportement RÉEL observé (le défaut est
+    #: topologie-dépendant : présent en linéaire/DAG, absent en
+    #: multi-pipeline avec cette synchro d'interruption).
+    _RESUME_VIEW_RELATION = {
+        "single_linear": "strict_subset",   # défaut : vues ⊊ pipeline
+        "branching_dag": "strict_subset",   # défaut idem
+        "multi_pipeline": "equal",          # pas de défaut ici
+    }
+
+    @pytest.mark.parametrize(
+        "topo", ["single_linear", "multi_pipeline", "branching_dag"],
+    )
+    def test_resume_view_vs_pipeline_relation_DEFECT_characterized(
+        self, tmp_path: Path, topo: str,
+    ) -> None:
+        """⚠️ CARACTÉRISE LE DÉFAUT (topologie-dépendant) ⚠️ : au
+        resume, ``pipeline_results`` couvre tout le corpus, mais la
+        relation ``view_results`` vs ``pipeline_results`` dépend de la
+        topologie (cf. :data:`_RESUME_VIEW_RELATION`) :
+
+        - ``single_linear`` / ``branching_dag`` : vues ⊊ pipeline —
+          les vues des docs repris du partial ne sont jamais
+          recalculées (métriques agrégées faussées après reprise).
+        - ``multi_pipeline`` : vues == pipeline (le défaut ne se
+          manifeste pas avec cette synchro d'interruption).
+
+        Toute évolution de l'une de ces relations (Phase B, ou
+        correction du défaut) fait échouer ce test et force une revue
+        consciente."""
+        _, resumed = self._interrupt_then_resume(
+            tmp_path, 5, stop_after=2, topo=topo,
+        )
+        pr, vr = self._persisted_doc_ids(resumed)
+        full = ["doc01", "doc02", "doc03", "doc04", "doc05"]
+        assert pr == full, f"pipeline incomplet au resume ({topo}): {pr}"
+        rel = self._RESUME_VIEW_RELATION[topo]
+        if rel == "strict_subset":
+            assert set(vr) < set(pr), (
+                f"[{topo}] défaut resume/vues changé : pipeline={pr} "
+                f"vues={vr}. Attendu : vues ⊊ pipeline."
+            )
+        else:
+            assert set(vr) == set(pr), (
+                f"[{topo}] relation resume/vues changée : pipeline={pr}"
+                f" vues={vr}. Attendu : vues == pipeline."
+            )
+
+    def test_resume_does_not_duplicate_documents(
+        self, tmp_path: Path,
+    ) -> None:
+        sources, body = _TOPOLOGIES["single_linear"]
+        cz = _make_corpus_zip(tmp_path / "in", 4, sources=sources)
+        partial = tmp_path / "p"
+        partial.mkdir()
+        body_pd = body + f"\n        partial_dir: {partial}"
+        out = tmp_path / "o"
+
+        ev = threading.Event()
+        c = {"n": 0}
+
+        def cb(e: str, i: int, d: str) -> None:
+            c["n"] += 1
+            if c["n"] == 1:
+                ev.set()
+
+        RunOrchestrator(out).execute(
+            _spec(cz, out, body_pd), progress_callback=cb, cancel_event=ev,
+        )
+        out2 = tmp_path / "o2"
+        res = RunOrchestrator(out2).execute(_spec(cz, out2, body_pd))
+        doc_ids = [d.document_id for d in res.run_result.document_results]
+        assert sorted(doc_ids) == ["doc01", "doc02", "doc03", "doc04"]
+        assert len(doc_ids) == len(set(doc_ids)), (
+            f"docs dupliqués au resume : {doc_ids}"
+        )
+
+
+# ---------------------------------------------------------------------------
+# 4. Golden multi-topologie + garde déterminisme (Risques 3 & 4)
+# ---------------------------------------------------------------------------
+
+class TestGoldenMultiTopology:
+    @pytest.mark.parametrize("topo", sorted(_TOPOLOGIES))
+    def test_snapshot_matches_golden(
+        self, tmp_path: Path, topo: str,
+    ) -> None:
+        _run(tmp_path, topo, n_docs=3)
+        snap = _normalized_snapshot(tmp_path / "out" / "results")
+        _golden(topo, snap)
+
+    @pytest.mark.parametrize("topo", sorted(_TOPOLOGIES))
+    def test_snapshot_is_deterministic(
+        self, tmp_path: Path, topo: str,
+    ) -> None:
+        """Garde anti-golden-flaky : deux runs du MÊME spec doivent
+        produire un snapshot normalisé bit-identique.  Si ça échoue,
+        le golden serait flaky → on refuse de le figer."""
+        _run(tmp_path / "a", topo, n_docs=3)
+        _run(tmp_path / "b", topo, n_docs=3)
+        s1 = _normalized_snapshot(tmp_path / "a" / "out" / "results")
+        s2 = _normalized_snapshot(tmp_path / "b" / "out" / "results")
+        assert s1 == s2, (
+            f"snapshot NON déterministe pour {topo} — un golden serait "
+            "flaky ; corriger la normalisation AVANT de figer"
+        )
+
+
+# ---------------------------------------------------------------------------
+# 5. Isolation concurrente (Risque 5)
+# ---------------------------------------------------------------------------
+
+class TestConcurrencyIsolation:
+    """Deux ``execute()`` en threads parallèles, ``output_dir`` et
+    ``cancel_event`` distincts.  Annuler A ne doit PAS perturber B
+    (ni progression, ni complétude) — l'invariant « RunOrchestrator
+    sans état entre deux execute() » est non testé et Phase B
+    (collaborateur réutilisé) le casserait."""
+
+    def test_cancel_in_thread_a_does_not_leak_into_thread_b(
+        self, tmp_path: Path,
+    ) -> None:
+        sources, body = _TOPOLOGIES["single_linear"]
+        cz = _make_corpus_zip(tmp_path / "in", 6, sources=sources)
+
+        ev_a = threading.Event()
+        ev_b = threading.Event()
+        prog_a: list[int] = []
+        prog_b: list[int] = []
+        err: dict[str, BaseException] = {}
+
+        def run_a() -> None:
+            try:
+                ev_a.set()  # A annulé d'emblée
+                out = tmp_path / "a"
+                RunOrchestrator(out).execute(
+                    _spec(cz, out, body),
+                    progress_callback=lambda e, i, d: prog_a.append(i),
+                    cancel_event=ev_a,
+                )
+            except BaseException as e:  # noqa: BLE001
+                err["a"] = e
+
+        def run_b() -> None:
+            try:
+                out = tmp_path / "b"
+                RunOrchestrator(out).execute(
+                    _spec(cz, out, body),
+                    progress_callback=lambda e, i, d: prog_b.append(i),
+                    cancel_event=ev_b,  # B jamais annulé
+                )
+            except BaseException as e:  # noqa: BLE001
+                err["b"] = e
+
+        ta = threading.Thread(target=run_a)
+        tb = threading.Thread(target=run_b)
+        ta.start()
+        tb.start()
+        ta.join(timeout=30)
+        tb.join(timeout=30)
+
+        assert not err, f"exception thread : {err}"
+        # B (non annulé) doit avoir traité ses 6 docs avec une
+        # séquence propre 0..5 — aucune fuite du compteur de A.
+        assert sorted(prog_b) == [0, 1, 2, 3, 4, 5], (
+            f"B perturbé par l'annulation de A : prog_b={sorted(prog_b)} "
+            "— fuite d'état entre instances (Phase B)"
+        )
+        # A annulé : strictement moins que 6.
+        assert len(prog_a) < 6, f"A non annulé : {prog_a}"